'2'

app.py

@@ -0,0 +1,63 @@
+import os
+import sys
+
+import ffmpeg
+import torch
+import warnings
+
+#验证SDK token
+from modelscope.hub.api import HubApi
+api = HubApi()
+api.login(os.getenv('token'))
+
+#模型下载
+from modelscope import snapshot_download
+model_dir = snapshot_download('yuzhouaoyou/uvr')
+
+warnings.filterwarnings("ignore")
+os.system('pip install librosa==0.9.1')
+os.system('pip install numpy==1.23.5')
+#os.system('pip install flask')
+
+sys.path.append(os.getcwd())
+from infer.modules.uvr5.mdxnet import MDXNetDereverb
+from infer.modules.uvr5.vr import AudioPre
+
+def run_uvr(file):
+    out_path = 'opt'
+    func = AudioPre(agg=10,model_path=f'{model_dir}/HP5.pth',device='cuda:0' if torch.cuda.is_available() else 'cpu',is_half=False)
+    need_reformat = True
+    try:
+        info = ffmpeg.probe(file, cmd="ffprobe")
+        if info["streams"][0]["channels"] == 2 and info["streams"][0]["sample_rate"] == "44100":
+            need_reformat = False
+    except:
+        need_reformat = True
+
+    if need_reformat:
+        tmp_path = "%s/%s.reformatted.wav" % (
+            os.path.join(os.environ["TEMP"]),
+            os.path.basename(file),
+        )
+        os.system(
+            "ffmpeg -i %s -vn -acodec pcm_s16le -ac 2 -ar 44100 %s -y"
+            % (file, tmp_path)
+        )
+        file = tmp_path
+    func._path_audio_(file, out_path, out_path, 'mp3')
+    return '%s/instrument_%s_10.mp3' % (out_path,os.path.basename(file)), '%s/vocal_%s_10.mp3' % (out_path,os.path.basename(file))
+    
+import gradio as gr
+
+def f(file):
+    i,v = run_uvr(file)
+    return [i,v]
+
+demo = gr.Interface(
+    fn=f,
+    inputs=[gr.File()],
+    outputs=[gr.Files()],
+    title="UVR 人声分离",
+)
+
+demo.launch()

infer/lib/audio.py

@@ -0,0 +1,73 @@
+import os
+import traceback
+
+import librosa
+import numpy as np
+import av
+from io import BytesIO
+
+
+def wav2(i, o, format):
+    inp = av.open(i, "rb")
+    if format == "m4a":
+        format = "mp4"
+    out = av.open(o, "wb", format=format)
+    if format == "ogg":
+        format = "libvorbis"
+    if format == "mp4":
+        format = "aac"
+
+    ostream = out.add_stream(format)
+
+    for frame in inp.decode(audio=0):
+        for p in ostream.encode(frame):
+            out.mux(p)
+
+    for p in ostream.encode(None):
+        out.mux(p)
+
+    out.close()
+    inp.close()
+
+
+def audio2(i, o, format, sr):
+    inp = av.open(i, "rb")
+    out = av.open(o, "wb", format=format)
+    if format == "ogg":
+        format = "libvorbis"
+    if format == "f32le":
+        format = "pcm_f32le"
+
+    ostream = out.add_stream(format, channels=1)
+    ostream.sample_rate = sr
+
+    for frame in inp.decode(audio=0):
+        for p in ostream.encode(frame):
+            out.mux(p)
+
+    out.close()
+    inp.close()
+
+
+def load_audio(file, sr):
+    file = (
+        file.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+    )  # 防止小白拷路径头尾带了空格和"和回车
+    if os.path.exists(file) == False:
+        raise RuntimeError(
+            "You input a wrong audio path that does not exists, please fix it!"
+        )
+    try:
+        with open(file, "rb") as f:
+            with BytesIO() as out:
+                audio2(f, out, "f32le", sr)
+                return np.frombuffer(out.getvalue(), np.float32).flatten()
+
+    except AttributeError:
+        audio = file[1] / 32768.0
+        if len(audio.shape) == 2:
+            audio = np.mean(audio, -1)
+        return librosa.resample(audio, orig_sr=file[0], target_sr=16000)
+
+    except:
+        raise RuntimeError(traceback.format_exc())

infer/lib/infer_pack/attentions.py

@@ -0,0 +1,459 @@
+import copy
+import math
+from typing import Optional
+
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from infer.lib.infer_pack import commons, modules
+from infer.lib.infer_pack.modules import LayerNorm
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size=1,
+        p_dropout=0.0,
+        window_size=10,
+        **kwargs
+    ):
+        super(Encoder, self).__init__()
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = int(n_layers)
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+
+        self.drop = nn.Dropout(p_dropout)
+        self.attn_layers = nn.ModuleList()
+        self.norm_layers_1 = nn.ModuleList()
+        self.ffn_layers = nn.ModuleList()
+        self.norm_layers_2 = nn.ModuleList()
+        for i in range(self.n_layers):
+            self.attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels,
+                    hidden_channels,
+                    n_heads,
+                    p_dropout=p_dropout,
+                    window_size=window_size,
+                )
+            )
+            self.norm_layers_1.append(LayerNorm(hidden_channels))
+            self.ffn_layers.append(
+                FFN(
+                    hidden_channels,
+                    hidden_channels,
+                    filter_channels,
+                    kernel_size,
+                    p_dropout=p_dropout,
+                )
+            )
+            self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+    def forward(self, x, x_mask):
+        attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+        x = x * x_mask
+        zippep = zip(
+            self.attn_layers, self.norm_layers_1, self.ffn_layers, self.norm_layers_2
+        )
+        for attn_layers, norm_layers_1, ffn_layers, norm_layers_2 in zippep:
+            y = attn_layers(x, x, attn_mask)
+            y = self.drop(y)
+            x = norm_layers_1(x + y)
+
+            y = ffn_layers(x, x_mask)
+            y = self.drop(y)
+            x = norm_layers_2(x + y)
+        x = x * x_mask
+        return x
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size=1,
+        p_dropout=0.0,
+        proximal_bias=False,
+        proximal_init=True,
+        **kwargs
+    ):
+        super(Decoder, self).__init__()
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.proximal_bias = proximal_bias
+        self.proximal_init = proximal_init
+
+        self.drop = nn.Dropout(p_dropout)
+        self.self_attn_layers = nn.ModuleList()
+        self.norm_layers_0 = nn.ModuleList()
+        self.encdec_attn_layers = nn.ModuleList()
+        self.norm_layers_1 = nn.ModuleList()
+        self.ffn_layers = nn.ModuleList()
+        self.norm_layers_2 = nn.ModuleList()
+        for i in range(self.n_layers):
+            self.self_attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels,
+                    hidden_channels,
+                    n_heads,
+                    p_dropout=p_dropout,
+                    proximal_bias=proximal_bias,
+                    proximal_init=proximal_init,
+                )
+            )
+            self.norm_layers_0.append(LayerNorm(hidden_channels))
+            self.encdec_attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout
+                )
+            )
+            self.norm_layers_1.append(LayerNorm(hidden_channels))
+            self.ffn_layers.append(
+                FFN(
+                    hidden_channels,
+                    hidden_channels,
+                    filter_channels,
+                    kernel_size,
+                    p_dropout=p_dropout,
+                    causal=True,
+                )
+            )
+            self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+    def forward(self, x, x_mask, h, h_mask):
+        """
+        x: decoder input
+        h: encoder output
+        """
+        self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(
+            device=x.device, dtype=x.dtype
+        )
+        encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+        x = x * x_mask
+        for i in range(self.n_layers):
+            y = self.self_attn_layers[i](x, x, self_attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_0[i](x + y)
+
+            y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_1[i](x + y)
+
+            y = self.ffn_layers[i](x, x_mask)
+            y = self.drop(y)
+            x = self.norm_layers_2[i](x + y)
+        x = x * x_mask
+        return x
+
+
+class MultiHeadAttention(nn.Module):
+    def __init__(
+        self,
+        channels,
+        out_channels,
+        n_heads,
+        p_dropout=0.0,
+        window_size=None,
+        heads_share=True,
+        block_length=None,
+        proximal_bias=False,
+        proximal_init=False,
+    ):
+        super(MultiHeadAttention, self).__init__()
+        assert channels % n_heads == 0
+
+        self.channels = channels
+        self.out_channels = out_channels
+        self.n_heads = n_heads
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+        self.heads_share = heads_share
+        self.block_length = block_length
+        self.proximal_bias = proximal_bias
+        self.proximal_init = proximal_init
+        self.attn = None
+
+        self.k_channels = channels // n_heads
+        self.conv_q = nn.Conv1d(channels, channels, 1)
+        self.conv_k = nn.Conv1d(channels, channels, 1)
+        self.conv_v = nn.Conv1d(channels, channels, 1)
+        self.conv_o = nn.Conv1d(channels, out_channels, 1)
+        self.drop = nn.Dropout(p_dropout)
+
+        if window_size is not None:
+            n_heads_rel = 1 if heads_share else n_heads
+            rel_stddev = self.k_channels**-0.5
+            self.emb_rel_k = nn.Parameter(
+                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+                * rel_stddev
+            )
+            self.emb_rel_v = nn.Parameter(
+                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+                * rel_stddev
+            )
+
+        nn.init.xavier_uniform_(self.conv_q.weight)
+        nn.init.xavier_uniform_(self.conv_k.weight)
+        nn.init.xavier_uniform_(self.conv_v.weight)
+        if proximal_init:
+            with torch.no_grad():
+                self.conv_k.weight.copy_(self.conv_q.weight)
+                self.conv_k.bias.copy_(self.conv_q.bias)
+
+    def forward(
+        self, x: torch.Tensor, c: torch.Tensor, attn_mask: Optional[torch.Tensor] = None
+    ):
+        q = self.conv_q(x)
+        k = self.conv_k(c)
+        v = self.conv_v(c)
+
+        x, _ = self.attention(q, k, v, mask=attn_mask)
+
+        x = self.conv_o(x)
+        return x
+
+    def attention(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+    ):
+        # reshape [b, d, t] -> [b, n_h, t, d_k]
+        b, d, t_s = key.size()
+        t_t = query.size(2)
+        query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
+        key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+        value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+
+        scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
+        if self.window_size is not None:
+            assert (
+                t_s == t_t
+            ), "Relative attention is only available for self-attention."
+            key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
+            rel_logits = self._matmul_with_relative_keys(
+                query / math.sqrt(self.k_channels), key_relative_embeddings
+            )
+            scores_local = self._relative_position_to_absolute_position(rel_logits)
+            scores = scores + scores_local
+        if self.proximal_bias:
+            assert t_s == t_t, "Proximal bias is only available for self-attention."
+            scores = scores + self._attention_bias_proximal(t_s).to(
+                device=scores.device, dtype=scores.dtype
+            )
+        if mask is not None:
+            scores = scores.masked_fill(mask == 0, -1e4)
+            if self.block_length is not None:
+                assert (
+                    t_s == t_t
+                ), "Local attention is only available for self-attention."
+                block_mask = (
+                    torch.ones_like(scores)
+                    .triu(-self.block_length)
+                    .tril(self.block_length)
+                )
+                scores = scores.masked_fill(block_mask == 0, -1e4)
+        p_attn = F.softmax(scores, dim=-1)  # [b, n_h, t_t, t_s]
+        p_attn = self.drop(p_attn)
+        output = torch.matmul(p_attn, value)
+        if self.window_size is not None:
+            relative_weights = self._absolute_position_to_relative_position(p_attn)
+            value_relative_embeddings = self._get_relative_embeddings(
+                self.emb_rel_v, t_s
+            )
+            output = output + self._matmul_with_relative_values(
+                relative_weights, value_relative_embeddings
+            )
+        output = (
+            output.transpose(2, 3).contiguous().view(b, d, t_t)
+        )  # [b, n_h, t_t, d_k] -> [b, d, t_t]
+        return output, p_attn
+
+    def _matmul_with_relative_values(self, x, y):
+        """
+        x: [b, h, l, m]
+        y: [h or 1, m, d]
+        ret: [b, h, l, d]
+        """
+        ret = torch.matmul(x, y.unsqueeze(0))
+        return ret
+
+    def _matmul_with_relative_keys(self, x, y):
+        """
+        x: [b, h, l, d]
+        y: [h or 1, m, d]
+        ret: [b, h, l, m]
+        """
+        ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
+        return ret
+
+    def _get_relative_embeddings(self, relative_embeddings, length: int):
+        max_relative_position = 2 * self.window_size + 1
+        # Pad first before slice to avoid using cond ops.
+        pad_length: int = max(length - (self.window_size + 1), 0)
+        slice_start_position = max((self.window_size + 1) - length, 0)
+        slice_end_position = slice_start_position + 2 * length - 1
+        if pad_length > 0:
+            padded_relative_embeddings = F.pad(
+                relative_embeddings,
+                # commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]),
+                [0, 0, pad_length, pad_length, 0, 0],
+            )
+        else:
+            padded_relative_embeddings = relative_embeddings
+        used_relative_embeddings = padded_relative_embeddings[
+            :, slice_start_position:slice_end_position
+        ]
+        return used_relative_embeddings
+
+    def _relative_position_to_absolute_position(self, x):
+        """
+        x: [b, h, l, 2*l-1]
+        ret: [b, h, l, l]
+        """
+        batch, heads, length, _ = x.size()
+        # Concat columns of pad to shift from relative to absolute indexing.
+        x = F.pad(
+            x,
+            #   commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]])
+            [0, 1, 0, 0, 0, 0, 0, 0],
+        )
+
+        # Concat extra elements so to add up to shape (len+1, 2*len-1).
+        x_flat = x.view([batch, heads, length * 2 * length])
+        x_flat = F.pad(
+            x_flat,
+            # commons.convert_pad_shape([[0, 0], [0, 0], [0, int(length) - 1]])
+            [0, int(length) - 1, 0, 0, 0, 0],
+        )
+
+        # Reshape and slice out the padded elements.
+        x_final = x_flat.view([batch, heads, length + 1, 2 * length - 1])[
+            :, :, :length, length - 1 :
+        ]
+        return x_final
+
+    def _absolute_position_to_relative_position(self, x):
+        """
+        x: [b, h, l, l]
+        ret: [b, h, l, 2*l-1]
+        """
+        batch, heads, length, _ = x.size()
+        # padd along column
+        x = F.pad(
+            x,
+            # commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, int(length) - 1]])
+            [0, int(length) - 1, 0, 0, 0, 0, 0, 0],
+        )
+        x_flat = x.view([batch, heads, int(length**2) + int(length * (length - 1))])
+        # add 0's in the beginning that will skew the elements after reshape
+        x_flat = F.pad(
+            x_flat,
+            #    commons.convert_pad_shape([[0, 0], [0, 0], [int(length), 0]])
+            [length, 0, 0, 0, 0, 0],
+        )
+        x_final = x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:]
+        return x_final
+
+    def _attention_bias_proximal(self, length: int):
+        """Bias for self-attention to encourage attention to close positions.
+        Args:
+          length: an integer scalar.
+        Returns:
+          a Tensor with shape [1, 1, length, length]
+        """
+        r = torch.arange(length, dtype=torch.float32)
+        diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
+        return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
+
+
+class FFN(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        filter_channels,
+        kernel_size,
+        p_dropout=0.0,
+        activation: str = None,
+        causal=False,
+    ):
+        super(FFN, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.activation = activation
+        self.causal = causal
+        self.is_activation = True if activation == "gelu" else False
+        # if causal:
+        #     self.padding = self._causal_padding
+        # else:
+        #     self.padding = self._same_padding
+
+        self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
+        self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
+        self.drop = nn.Dropout(p_dropout)
+
+    def padding(self, x: torch.Tensor, x_mask: torch.Tensor) -> torch.Tensor:
+        if self.causal:
+            padding = self._causal_padding(x * x_mask)
+        else:
+            padding = self._same_padding(x * x_mask)
+        return padding
+
+    def forward(self, x: torch.Tensor, x_mask: torch.Tensor):
+        x = self.conv_1(self.padding(x, x_mask))
+        if self.is_activation:
+            x = x * torch.sigmoid(1.702 * x)
+        else:
+            x = torch.relu(x)
+        x = self.drop(x)
+
+        x = self.conv_2(self.padding(x, x_mask))
+        return x * x_mask
+
+    def _causal_padding(self, x):
+        if self.kernel_size == 1:
+            return x
+        pad_l: int = self.kernel_size - 1
+        pad_r: int = 0
+        # padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+        x = F.pad(
+            x,
+            #   commons.convert_pad_shape(padding)
+            [pad_l, pad_r, 0, 0, 0, 0],
+        )
+        return x
+
+    def _same_padding(self, x):
+        if self.kernel_size == 1:
+            return x
+        pad_l: int = (self.kernel_size - 1) // 2
+        pad_r: int = self.kernel_size // 2
+        # padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+        x = F.pad(
+            x,
+            #   commons.convert_pad_shape(padding)
+            [pad_l, pad_r, 0, 0, 0, 0],
+        )
+        return x

infer/lib/infer_pack/commons.py

@@ -0,0 +1,172 @@
+from typing import List, Optional
+import math
+
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+# def convert_pad_shape(pad_shape):
+#     l = pad_shape[::-1]
+#     pad_shape = [item for sublist in l for item in sublist]
+#     return pad_shape
+
+
+def kl_divergence(m_p, logs_p, m_q, logs_q):
+    """KL(P||Q)"""
+    kl = (logs_q - logs_p) - 0.5
+    kl += (
+        0.5 * (torch.exp(2.0 * logs_p) + ((m_p - m_q) ** 2)) * torch.exp(-2.0 * logs_q)
+    )
+    return kl
+
+
+def rand_gumbel(shape):
+    """Sample from the Gumbel distribution, protect from overflows."""
+    uniform_samples = torch.rand(shape) * 0.99998 + 0.00001
+    return -torch.log(-torch.log(uniform_samples))
+
+
+def rand_gumbel_like(x):
+    g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)
+    return g
+
+
+def slice_segments(x, ids_str, segment_size=4):
+    ret = torch.zeros_like(x[:, :, :segment_size])
+    for i in range(x.size(0)):
+        idx_str = ids_str[i]
+        idx_end = idx_str + segment_size
+        ret[i] = x[i, :, idx_str:idx_end]
+    return ret
+
+
+def slice_segments2(x, ids_str, segment_size=4):
+    ret = torch.zeros_like(x[:, :segment_size])
+    for i in range(x.size(0)):
+        idx_str = ids_str[i]
+        idx_end = idx_str + segment_size
+        ret[i] = x[i, idx_str:idx_end]
+    return ret
+
+
+def rand_slice_segments(x, x_lengths=None, segment_size=4):
+    b, d, t = x.size()
+    if x_lengths is None:
+        x_lengths = t
+    ids_str_max = x_lengths - segment_size + 1
+    ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
+    ret = slice_segments(x, ids_str, segment_size)
+    return ret, ids_str
+
+
+def get_timing_signal_1d(length, channels, min_timescale=1.0, max_timescale=1.0e4):
+    position = torch.arange(length, dtype=torch.float)
+    num_timescales = channels // 2
+    log_timescale_increment = math.log(float(max_timescale) / float(min_timescale)) / (
+        num_timescales - 1
+    )
+    inv_timescales = min_timescale * torch.exp(
+        torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment
+    )
+    scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
+    signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
+    signal = F.pad(signal, [0, 0, 0, channels % 2])
+    signal = signal.view(1, channels, length)
+    return signal
+
+
+def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
+    b, channels, length = x.size()
+    signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
+    return x + signal.to(dtype=x.dtype, device=x.device)
+
+
+def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):
+    b, channels, length = x.size()
+    signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
+    return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)
+
+
+def subsequent_mask(length):
+    mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
+    return mask
+
+
+@torch.jit.script
+def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
+    n_channels_int = n_channels[0]
+    in_act = input_a + input_b
+    t_act = torch.tanh(in_act[:, :n_channels_int, :])
+    s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
+    acts = t_act * s_act
+    return acts
+
+
+# def convert_pad_shape(pad_shape):
+#     l = pad_shape[::-1]
+#     pad_shape = [item for sublist in l for item in sublist]
+#     return pad_shape
+
+
+def convert_pad_shape(pad_shape: List[List[int]]) -> List[int]:
+    return torch.tensor(pad_shape).flip(0).reshape(-1).int().tolist()
+
+
+def shift_1d(x):
+    x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
+    return x
+
+
+def sequence_mask(length: torch.Tensor, max_length: Optional[int] = None):
+    if max_length is None:
+        max_length = length.max()
+    x = torch.arange(max_length, dtype=length.dtype, device=length.device)
+    return x.unsqueeze(0) < length.unsqueeze(1)
+
+
+def generate_path(duration, mask):
+    """
+    duration: [b, 1, t_x]
+    mask: [b, 1, t_y, t_x]
+    """
+    device = duration.device
+
+    b, _, t_y, t_x = mask.shape
+    cum_duration = torch.cumsum(duration, -1)
+
+    cum_duration_flat = cum_duration.view(b * t_x)
+    path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
+    path = path.view(b, t_x, t_y)
+    path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
+    path = path.unsqueeze(1).transpose(2, 3) * mask
+    return path
+
+
+def clip_grad_value_(parameters, clip_value, norm_type=2):
+    if isinstance(parameters, torch.Tensor):
+        parameters = [parameters]
+    parameters = list(filter(lambda p: p.grad is not None, parameters))
+    norm_type = float(norm_type)
+    if clip_value is not None:
+        clip_value = float(clip_value)
+
+    total_norm = 0
+    for p in parameters:
+        param_norm = p.grad.data.norm(norm_type)
+        total_norm += param_norm.item() ** norm_type
+        if clip_value is not None:
+            p.grad.data.clamp_(min=-clip_value, max=clip_value)
+    total_norm = total_norm ** (1.0 / norm_type)
+    return total_norm

infer/lib/infer_pack/models_onnx.py

@@ -0,0 +1,821 @@
+import math
+import logging
+
+logger = logging.getLogger(__name__)
+
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import AvgPool1d, Conv1d, Conv2d, ConvTranspose1d
+from torch.nn import functional as F
+from torch.nn.utils import remove_weight_norm, spectral_norm, weight_norm
+
+from infer.lib.infer_pack import attentions, commons, modules
+from infer.lib.infer_pack.commons import get_padding, init_weights
+
+
+class TextEncoder256(nn.Module):
+    def __init__(
+        self,
+        out_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        f0=True,
+    ):
+        super().__init__()
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.emb_phone = nn.Linear(256, hidden_channels)
+        self.lrelu = nn.LeakyReLU(0.1, inplace=True)
+        if f0 == True:
+            self.emb_pitch = nn.Embedding(256, hidden_channels)  # pitch 256
+        self.encoder = attentions.Encoder(
+            hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, phone, pitch, lengths):
+        if pitch == None:
+            x = self.emb_phone(phone)
+        else:
+            x = self.emb_phone(phone) + self.emb_pitch(pitch)
+        x = x * math.sqrt(self.hidden_channels)  # [b, t, h]
+        x = self.lrelu(x)
+        x = torch.transpose(x, 1, -1)  # [b, h, t]
+        x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.encoder(x * x_mask, x_mask)
+        stats = self.proj(x) * x_mask
+
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        return m, logs, x_mask
+
+
+class TextEncoder768(nn.Module):
+    def __init__(
+        self,
+        out_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        f0=True,
+    ):
+        super().__init__()
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.emb_phone = nn.Linear(768, hidden_channels)
+        self.lrelu = nn.LeakyReLU(0.1, inplace=True)
+        if f0 == True:
+            self.emb_pitch = nn.Embedding(256, hidden_channels)  # pitch 256
+        self.encoder = attentions.Encoder(
+            hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, phone, pitch, lengths):
+        if pitch == None:
+            x = self.emb_phone(phone)
+        else:
+            x = self.emb_phone(phone) + self.emb_pitch(pitch)
+        x = x * math.sqrt(self.hidden_channels)  # [b, t, h]
+        x = self.lrelu(x)
+        x = torch.transpose(x, 1, -1)  # [b, h, t]
+        x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.encoder(x * x_mask, x_mask)
+        stats = self.proj(x) * x_mask
+
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        return m, logs, x_mask
+
+
+class ResidualCouplingBlock(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        n_flows=4,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.flows = nn.ModuleList()
+        for i in range(n_flows):
+            self.flows.append(
+                modules.ResidualCouplingLayer(
+                    channels,
+                    hidden_channels,
+                    kernel_size,
+                    dilation_rate,
+                    n_layers,
+                    gin_channels=gin_channels,
+                    mean_only=True,
+                )
+            )
+            self.flows.append(modules.Flip())
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        if not reverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        else:
+            for flow in reversed(self.flows):
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        return x
+
+    def remove_weight_norm(self):
+        for i in range(self.n_flows):
+            self.flows[i * 2].remove_weight_norm()
+
+
+class PosteriorEncoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+
+        self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
+        self.enc = modules.WN(
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            gin_channels=gin_channels,
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, x, x_lengths, g=None):
+        x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.pre(x) * x_mask
+        x = self.enc(x, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
+        return z, m, logs, x_mask
+
+    def remove_weight_norm(self):
+        self.enc.remove_weight_norm()
+
+
+class Generator(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gin_channels=0,
+    ):
+        super(Generator, self).__init__()
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.conv_pre = Conv1d(
+            initial_channel, upsample_initial_channel, 7, 1, padding=3
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(init_weights)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+    def forward(self, x, g=None):
+        x = self.conv_pre(x)
+        if g is not None:
+            x = x + self.cond(g)
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+
+
+class SineGen(torch.nn.Module):
+    """Definition of sine generator
+    SineGen(samp_rate, harmonic_num = 0,
+            sine_amp = 0.1, noise_std = 0.003,
+            voiced_threshold = 0,
+            flag_for_pulse=False)
+    samp_rate: sampling rate in Hz
+    harmonic_num: number of harmonic overtones (default 0)
+    sine_amp: amplitude of sine-wavefrom (default 0.1)
+    noise_std: std of Gaussian noise (default 0.003)
+    voiced_thoreshold: F0 threshold for U/V classification (default 0)
+    flag_for_pulse: this SinGen is used inside PulseGen (default False)
+    Note: when flag_for_pulse is True, the first time step of a voiced
+        segment is always sin(np.pi) or cos(0)
+    """
+
+    def __init__(
+        self,
+        samp_rate,
+        harmonic_num=0,
+        sine_amp=0.1,
+        noise_std=0.003,
+        voiced_threshold=0,
+        flag_for_pulse=False,
+    ):
+        super(SineGen, self).__init__()
+        self.sine_amp = sine_amp
+        self.noise_std = noise_std
+        self.harmonic_num = harmonic_num
+        self.dim = self.harmonic_num + 1
+        self.sampling_rate = samp_rate
+        self.voiced_threshold = voiced_threshold
+
+    def _f02uv(self, f0):
+        # generate uv signal
+        uv = torch.ones_like(f0)
+        uv = uv * (f0 > self.voiced_threshold)
+        return uv
+
+    def forward(self, f0, upp):
+        """sine_tensor, uv = forward(f0)
+        input F0: tensor(batchsize=1, length, dim=1)
+                  f0 for unvoiced steps should be 0
+        output sine_tensor: tensor(batchsize=1, length, dim)
+        output uv: tensor(batchsize=1, length, 1)
+        """
+        with torch.no_grad():
+            f0 = f0[:, None].transpose(1, 2)
+            f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, device=f0.device)
+            # fundamental component
+            f0_buf[:, :, 0] = f0[:, :, 0]
+            for idx in np.arange(self.harmonic_num):
+                f0_buf[:, :, idx + 1] = f0_buf[:, :, 0] * (
+                    idx + 2
+                )  # idx + 2: the (idx+1)-th overtone, (idx+2)-th harmonic
+            rad_values = (f0_buf / self.sampling_rate) % 1  ###%1意味着n_har的乘积无法后处理优化
+            rand_ini = torch.rand(
+                f0_buf.shape[0], f0_buf.shape[2], device=f0_buf.device
+            )
+            rand_ini[:, 0] = 0
+            rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
+            tmp_over_one = torch.cumsum(rad_values, 1)  # % 1  #####%1意味着后面的cumsum无法再优化
+            tmp_over_one *= upp
+            tmp_over_one = F.interpolate(
+                tmp_over_one.transpose(2, 1),
+                scale_factor=upp,
+                mode="linear",
+                align_corners=True,
+            ).transpose(2, 1)
+            rad_values = F.interpolate(
+                rad_values.transpose(2, 1), scale_factor=upp, mode="nearest"
+            ).transpose(
+                2, 1
+            )  #######
+            tmp_over_one %= 1
+            tmp_over_one_idx = (tmp_over_one[:, 1:, :] - tmp_over_one[:, :-1, :]) < 0
+            cumsum_shift = torch.zeros_like(rad_values)
+            cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
+            sine_waves = torch.sin(
+                torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * np.pi
+            )
+            sine_waves = sine_waves * self.sine_amp
+            uv = self._f02uv(f0)
+            uv = F.interpolate(
+                uv.transpose(2, 1), scale_factor=upp, mode="nearest"
+            ).transpose(2, 1)
+            noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
+            noise = noise_amp * torch.randn_like(sine_waves)
+            sine_waves = sine_waves * uv + noise
+        return sine_waves, uv, noise
+
+
+class SourceModuleHnNSF(torch.nn.Module):
+    """SourceModule for hn-nsf
+    SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
+                 add_noise_std=0.003, voiced_threshod=0)
+    sampling_rate: sampling_rate in Hz
+    harmonic_num: number of harmonic above F0 (default: 0)
+    sine_amp: amplitude of sine source signal (default: 0.1)
+    add_noise_std: std of additive Gaussian noise (default: 0.003)
+        note that amplitude of noise in unvoiced is decided
+        by sine_amp
+    voiced_threshold: threhold to set U/V given F0 (default: 0)
+    Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+    F0_sampled (batchsize, length, 1)
+    Sine_source (batchsize, length, 1)
+    noise_source (batchsize, length 1)
+    uv (batchsize, length, 1)
+    """
+
+    def __init__(
+        self,
+        sampling_rate,
+        harmonic_num=0,
+        sine_amp=0.1,
+        add_noise_std=0.003,
+        voiced_threshod=0,
+        is_half=True,
+    ):
+        super(SourceModuleHnNSF, self).__init__()
+
+        self.sine_amp = sine_amp
+        self.noise_std = add_noise_std
+        self.is_half = is_half
+        # to produce sine waveforms
+        self.l_sin_gen = SineGen(
+            sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshod
+        )
+
+        # to merge source harmonics into a single excitation
+        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
+        self.l_tanh = torch.nn.Tanh()
+
+    def forward(self, x, upp=None):
+        sine_wavs, uv, _ = self.l_sin_gen(x, upp)
+        if self.is_half:
+            sine_wavs = sine_wavs.half()
+        sine_merge = self.l_tanh(self.l_linear(sine_wavs))
+        return sine_merge, None, None  # noise, uv
+
+
+class GeneratorNSF(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gin_channels,
+        sr,
+        is_half=False,
+    ):
+        super(GeneratorNSF, self).__init__()
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+
+        self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates))
+        self.m_source = SourceModuleHnNSF(
+            sampling_rate=sr, harmonic_num=0, is_half=is_half
+        )
+        self.noise_convs = nn.ModuleList()
+        self.conv_pre = Conv1d(
+            initial_channel, upsample_initial_channel, 7, 1, padding=3
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            c_cur = upsample_initial_channel // (2 ** (i + 1))
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+            if i + 1 < len(upsample_rates):
+                stride_f0 = np.prod(upsample_rates[i + 1 :])
+                self.noise_convs.append(
+                    Conv1d(
+                        1,
+                        c_cur,
+                        kernel_size=stride_f0 * 2,
+                        stride=stride_f0,
+                        padding=stride_f0 // 2,
+                    )
+                )
+            else:
+                self.noise_convs.append(Conv1d(1, c_cur, kernel_size=1))
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(init_weights)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+        self.upp = np.prod(upsample_rates)
+
+    def forward(self, x, f0, g=None):
+        har_source, noi_source, uv = self.m_source(f0, self.upp)
+        har_source = har_source.transpose(1, 2)
+        x = self.conv_pre(x)
+        if g is not None:
+            x = x + self.cond(g)
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            x = self.ups[i](x)
+            x_source = self.noise_convs[i](har_source)
+            x = x + x_source
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+
+
+sr2sr = {
+    "32k": 32000,
+    "40k": 40000,
+    "48k": 48000,
+}
+
+
+class SynthesizerTrnMsNSFsidM(nn.Module):
+    def __init__(
+        self,
+        spec_channels,
+        segment_size,
+        inter_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        spk_embed_dim,
+        gin_channels,
+        sr,
+        version,
+        **kwargs,
+    ):
+        super().__init__()
+        if type(sr) == type("strr"):
+            sr = sr2sr[sr]
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.gin_channels = gin_channels
+        # self.hop_length = hop_length#
+        self.spk_embed_dim = spk_embed_dim
+        if version == "v1":
+            self.enc_p = TextEncoder256(
+                inter_channels,
+                hidden_channels,
+                filter_channels,
+                n_heads,
+                n_layers,
+                kernel_size,
+                p_dropout,
+            )
+        else:
+            self.enc_p = TextEncoder768(
+                inter_channels,
+                hidden_channels,
+                filter_channels,
+                n_heads,
+                n_layers,
+                kernel_size,
+                p_dropout,
+            )
+        self.dec = GeneratorNSF(
+            inter_channels,
+            resblock,
+            resblock_kernel_sizes,
+            resblock_dilation_sizes,
+            upsample_rates,
+            upsample_initial_channel,
+            upsample_kernel_sizes,
+            gin_channels=gin_channels,
+            sr=sr,
+            is_half=kwargs["is_half"],
+        )
+        self.enc_q = PosteriorEncoder(
+            spec_channels,
+            inter_channels,
+            hidden_channels,
+            5,
+            1,
+            16,
+            gin_channels=gin_channels,
+        )
+        self.flow = ResidualCouplingBlock(
+            inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels
+        )
+        self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
+        self.speaker_map = None
+        logger.debug(
+            f"gin_channels: {gin_channels}, self.spk_embed_dim: {self.spk_embed_dim}"
+        )
+
+    def remove_weight_norm(self):
+        self.dec.remove_weight_norm()
+        self.flow.remove_weight_norm()
+        self.enc_q.remove_weight_norm()
+
+    def construct_spkmixmap(self, n_speaker):
+        self.speaker_map = torch.zeros((n_speaker, 1, 1, self.gin_channels))
+        for i in range(n_speaker):
+            self.speaker_map[i] = self.emb_g(torch.LongTensor([[i]]))
+        self.speaker_map = self.speaker_map.unsqueeze(0)
+
+    def forward(self, phone, phone_lengths, pitch, nsff0, g, rnd, max_len=None):
+        if self.speaker_map is not None:  # [N, S]  *  [S, B, 1, H]
+            g = g.reshape((g.shape[0], g.shape[1], 1, 1, 1))  # [N, S, B, 1, 1]
+            g = g * self.speaker_map  # [N, S, B, 1, H]
+            g = torch.sum(g, dim=1)  # [N, 1, B, 1, H]
+            g = g.transpose(0, -1).transpose(0, -2).squeeze(0)  # [B, H, N]
+        else:
+            g = g.unsqueeze(0)
+            g = self.emb_g(g).transpose(1, 2)
+
+        m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
+        z_p = (m_p + torch.exp(logs_p) * rnd) * x_mask
+        z = self.flow(z_p, x_mask, g=g, reverse=True)
+        o = self.dec((z * x_mask)[:, :, :max_len], nsff0, g=g)
+        return o
+
+
+class MultiPeriodDiscriminator(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminator, self).__init__()
+        periods = [2, 3, 5, 7, 11, 17]
+        # periods = [3, 5, 7, 11, 17, 23, 37]
+
+        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
+        discs = discs + [
+            DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
+        ]
+        self.discriminators = nn.ModuleList(discs)
+
+    def forward(self, y, y_hat):
+        y_d_rs = []  #
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            # for j in range(len(fmap_r)):
+            #     print(i,j,y.shape,y_hat.shape,fmap_r[j].shape,fmap_g[j].shape)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class MultiPeriodDiscriminatorV2(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminatorV2, self).__init__()
+        # periods = [2, 3, 5, 7, 11, 17]
+        periods = [2, 3, 5, 7, 11, 17, 23, 37]
+
+        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
+        discs = discs + [
+            DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
+        ]
+        self.discriminators = nn.ModuleList(discs)
+
+    def forward(self, y, y_hat):
+        y_d_rs = []  #
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            # for j in range(len(fmap_r)):
+            #     print(i,j,y.shape,y_hat.shape,fmap_r[j].shape,fmap_g[j].shape)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class DiscriminatorS(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(DiscriminatorS, self).__init__()
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(Conv1d(1, 16, 15, 1, padding=7)),
+                norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
+                norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
+                norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
+                norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
+                norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
+            ]
+        )
+        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
+
+    def forward(self, x):
+        fmap = []
+
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class DiscriminatorP(torch.nn.Module):
+    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
+        super(DiscriminatorP, self).__init__()
+        self.period = period
+        self.use_spectral_norm = use_spectral_norm
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(
+                    Conv2d(
+                        1,
+                        32,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        32,
+                        128,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        128,
+                        512,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        512,
+                        1024,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        1024,
+                        1024,
+                        (kernel_size, 1),
+                        1,
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+            ]
+        )
+        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
+
+    def forward(self, x):
+        fmap = []
+
+        # 1d to 2d
+        b, c, t = x.shape
+        if t % self.period != 0:  # pad first
+            n_pad = self.period - (t % self.period)
+            x = F.pad(x, (0, n_pad), "reflect")
+            t = t + n_pad
+        x = x.view(b, c, t // self.period, self.period)
+
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap

infer/lib/infer_pack/modules.py

@@ -0,0 +1,615 @@
+import copy
+import math
+from typing import Optional, Tuple
+
+import numpy as np
+import scipy
+import torch
+from torch import nn
+from torch.nn import AvgPool1d, Conv1d, Conv2d, ConvTranspose1d
+from torch.nn import functional as F
+from torch.nn.utils import remove_weight_norm, weight_norm
+
+from infer.lib.infer_pack import commons
+from infer.lib.infer_pack.commons import get_padding, init_weights
+from infer.lib.infer_pack.transforms import piecewise_rational_quadratic_transform
+
+LRELU_SLOPE = 0.1
+
+
+class LayerNorm(nn.Module):
+    def __init__(self, channels, eps=1e-5):
+        super(LayerNorm, self).__init__()
+        self.channels = channels
+        self.eps = eps
+
+        self.gamma = nn.Parameter(torch.ones(channels))
+        self.beta = nn.Parameter(torch.zeros(channels))
+
+    def forward(self, x):
+        x = x.transpose(1, -1)
+        x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
+        return x.transpose(1, -1)
+
+
+class ConvReluNorm(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        hidden_channels,
+        out_channels,
+        kernel_size,
+        n_layers,
+        p_dropout,
+    ):
+        super(ConvReluNorm, self).__init__()
+        self.in_channels = in_channels
+        self.hidden_channels = hidden_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.p_dropout = float(p_dropout)
+        assert n_layers > 1, "Number of layers should be larger than 0."
+
+        self.conv_layers = nn.ModuleList()
+        self.norm_layers = nn.ModuleList()
+        self.conv_layers.append(
+            nn.Conv1d(
+                in_channels, hidden_channels, kernel_size, padding=kernel_size // 2
+            )
+        )
+        self.norm_layers.append(LayerNorm(hidden_channels))
+        self.relu_drop = nn.Sequential(nn.ReLU(), nn.Dropout(float(p_dropout)))
+        for _ in range(n_layers - 1):
+            self.conv_layers.append(
+                nn.Conv1d(
+                    hidden_channels,
+                    hidden_channels,
+                    kernel_size,
+                    padding=kernel_size // 2,
+                )
+            )
+            self.norm_layers.append(LayerNorm(hidden_channels))
+        self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
+        self.proj.weight.data.zero_()
+        self.proj.bias.data.zero_()
+
+    def forward(self, x, x_mask):
+        x_org = x
+        for i in range(self.n_layers):
+            x = self.conv_layers[i](x * x_mask)
+            x = self.norm_layers[i](x)
+            x = self.relu_drop(x)
+        x = x_org + self.proj(x)
+        return x * x_mask
+
+
+class DDSConv(nn.Module):
+    """
+    Dialted and Depth-Separable Convolution
+    """
+
+    def __init__(self, channels, kernel_size, n_layers, p_dropout=0.0):
+        super(DDSConv, self).__init__()
+        self.channels = channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.p_dropout = float(p_dropout)
+
+        self.drop = nn.Dropout(float(p_dropout))
+        self.convs_sep = nn.ModuleList()
+        self.convs_1x1 = nn.ModuleList()
+        self.norms_1 = nn.ModuleList()
+        self.norms_2 = nn.ModuleList()
+        for i in range(n_layers):
+            dilation = kernel_size**i
+            padding = (kernel_size * dilation - dilation) // 2
+            self.convs_sep.append(
+                nn.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    groups=channels,
+                    dilation=dilation,
+                    padding=padding,
+                )
+            )
+            self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
+            self.norms_1.append(LayerNorm(channels))
+            self.norms_2.append(LayerNorm(channels))
+
+    def forward(self, x, x_mask, g: Optional[torch.Tensor] = None):
+        if g is not None:
+            x = x + g
+        for i in range(self.n_layers):
+            y = self.convs_sep[i](x * x_mask)
+            y = self.norms_1[i](y)
+            y = F.gelu(y)
+            y = self.convs_1x1[i](y)
+            y = self.norms_2[i](y)
+            y = F.gelu(y)
+            y = self.drop(y)
+            x = x + y
+        return x * x_mask
+
+
+class WN(torch.nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+        p_dropout=0,
+    ):
+        super(WN, self).__init__()
+        assert kernel_size % 2 == 1
+        self.hidden_channels = hidden_channels
+        self.kernel_size = (kernel_size,)
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+        self.p_dropout = float(p_dropout)
+
+        self.in_layers = torch.nn.ModuleList()
+        self.res_skip_layers = torch.nn.ModuleList()
+        self.drop = nn.Dropout(float(p_dropout))
+
+        if gin_channels != 0:
+            cond_layer = torch.nn.Conv1d(
+                gin_channels, 2 * hidden_channels * n_layers, 1
+            )
+            self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name="weight")
+
+        for i in range(n_layers):
+            dilation = dilation_rate**i
+            padding = int((kernel_size * dilation - dilation) / 2)
+            in_layer = torch.nn.Conv1d(
+                hidden_channels,
+                2 * hidden_channels,
+                kernel_size,
+                dilation=dilation,
+                padding=padding,
+            )
+            in_layer = torch.nn.utils.weight_norm(in_layer, name="weight")
+            self.in_layers.append(in_layer)
+
+            # last one is not necessary
+            if i < n_layers - 1:
+                res_skip_channels = 2 * hidden_channels
+            else:
+                res_skip_channels = hidden_channels
+
+            res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
+            res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name="weight")
+            self.res_skip_layers.append(res_skip_layer)
+
+    def forward(
+        self, x: torch.Tensor, x_mask: torch.Tensor, g: Optional[torch.Tensor] = None
+    ):
+        output = torch.zeros_like(x)
+        n_channels_tensor = torch.IntTensor([self.hidden_channels])
+
+        if g is not None:
+            g = self.cond_layer(g)
+
+        for i, (in_layer, res_skip_layer) in enumerate(
+            zip(self.in_layers, self.res_skip_layers)
+        ):
+            x_in = in_layer(x)
+            if g is not None:
+                cond_offset = i * 2 * self.hidden_channels
+                g_l = g[:, cond_offset : cond_offset + 2 * self.hidden_channels, :]
+            else:
+                g_l = torch.zeros_like(x_in)
+
+            acts = commons.fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor)
+            acts = self.drop(acts)
+
+            res_skip_acts = res_skip_layer(acts)
+            if i < self.n_layers - 1:
+                res_acts = res_skip_acts[:, : self.hidden_channels, :]
+                x = (x + res_acts) * x_mask
+                output = output + res_skip_acts[:, self.hidden_channels :, :]
+            else:
+                output = output + res_skip_acts
+        return output * x_mask
+
+    def remove_weight_norm(self):
+        if self.gin_channels != 0:
+            torch.nn.utils.remove_weight_norm(self.cond_layer)
+        for l in self.in_layers:
+            torch.nn.utils.remove_weight_norm(l)
+        for l in self.res_skip_layers:
+            torch.nn.utils.remove_weight_norm(l)
+
+    def __prepare_scriptable__(self):
+        if self.gin_channels != 0:
+            for hook in self.cond_layer._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(self.cond_layer)
+        for l in self.in_layers:
+            for hook in l._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(l)
+        for l in self.res_skip_layers:
+            for hook in l._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(l)
+        return self
+
+
+class ResBlock1(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock1, self).__init__()
+        self.convs1 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[2],
+                        padding=get_padding(kernel_size, dilation[2]),
+                    )
+                ),
+            ]
+        )
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+            ]
+        )
+        self.convs2.apply(init_weights)
+        self.lrelu_slope = LRELU_SLOPE
+
+    def forward(self, x: torch.Tensor, x_mask: Optional[torch.Tensor] = None):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, self.lrelu_slope)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, self.lrelu_slope)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c2(xt)
+            x = xt + x
+        if x_mask is not None:
+            x = x * x_mask
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+
+    def __prepare_scriptable__(self):
+        for l in self.convs1:
+            for hook in l._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(l)
+        for l in self.convs2:
+            for hook in l._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(l)
+        return self
+
+
+class ResBlock2(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
+        super(ResBlock2, self).__init__()
+        self.convs = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+            ]
+        )
+        self.convs.apply(init_weights)
+        self.lrelu_slope = LRELU_SLOPE
+
+    def forward(self, x, x_mask: Optional[torch.Tensor] = None):
+        for c in self.convs:
+            xt = F.leaky_relu(x, self.lrelu_slope)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c(xt)
+            x = xt + x
+        if x_mask is not None:
+            x = x * x_mask
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs:
+            remove_weight_norm(l)
+
+    def __prepare_scriptable__(self):
+        for l in self.convs:
+            for hook in l._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(l)
+        return self
+
+
+class Log(nn.Module):
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        g: Optional[torch.Tensor] = None,
+        reverse: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        if not reverse:
+            y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
+            logdet = torch.sum(-y, [1, 2])
+            return y, logdet
+        else:
+            x = torch.exp(x) * x_mask
+            return x
+
+
+class Flip(nn.Module):
+    # torch.jit.script() Compiled functions \
+    # can't take variable number of arguments or \
+    # use keyword-only arguments with defaults
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        g: Optional[torch.Tensor] = None,
+        reverse: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        x = torch.flip(x, [1])
+        if not reverse:
+            logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
+            return x, logdet
+        else:
+            return x, torch.zeros([1], device=x.device)
+
+
+class ElementwiseAffine(nn.Module):
+    def __init__(self, channels):
+        super(ElementwiseAffine, self).__init__()
+        self.channels = channels
+        self.m = nn.Parameter(torch.zeros(channels, 1))
+        self.logs = nn.Parameter(torch.zeros(channels, 1))
+
+    def forward(self, x, x_mask, reverse=False, **kwargs):
+        if not reverse:
+            y = self.m + torch.exp(self.logs) * x
+            y = y * x_mask
+            logdet = torch.sum(self.logs * x_mask, [1, 2])
+            return y, logdet
+        else:
+            x = (x - self.m) * torch.exp(-self.logs) * x_mask
+            return x
+
+
+class ResidualCouplingLayer(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        p_dropout=0,
+        gin_channels=0,
+        mean_only=False,
+    ):
+        assert channels % 2 == 0, "channels should be divisible by 2"
+        super(ResidualCouplingLayer, self).__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.half_channels = channels // 2
+        self.mean_only = mean_only
+
+        self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
+        self.enc = WN(
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            p_dropout=float(p_dropout),
+            gin_channels=gin_channels,
+        )
+        self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
+        self.post.weight.data.zero_()
+        self.post.bias.data.zero_()
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        g: Optional[torch.Tensor] = None,
+        reverse: bool = False,
+    ):
+        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+        h = self.pre(x0) * x_mask
+        h = self.enc(h, x_mask, g=g)
+        stats = self.post(h) * x_mask
+        if not self.mean_only:
+            m, logs = torch.split(stats, [self.half_channels] * 2, 1)
+        else:
+            m = stats
+            logs = torch.zeros_like(m)
+
+        if not reverse:
+            x1 = m + x1 * torch.exp(logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            logdet = torch.sum(logs, [1, 2])
+            return x, logdet
+        else:
+            x1 = (x1 - m) * torch.exp(-logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            return x, torch.zeros([1])
+
+    def remove_weight_norm(self):
+        self.enc.remove_weight_norm()
+
+    def __prepare_scriptable__(self):
+        for hook in self.enc._forward_pre_hooks.values():
+            if (
+                hook.__module__ == "torch.nn.utils.weight_norm"
+                and hook.__class__.__name__ == "WeightNorm"
+            ):
+                torch.nn.utils.remove_weight_norm(self.enc)
+        return self
+
+
+class ConvFlow(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        filter_channels,
+        kernel_size,
+        n_layers,
+        num_bins=10,
+        tail_bound=5.0,
+    ):
+        super(ConvFlow, self).__init__()
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.num_bins = num_bins
+        self.tail_bound = tail_bound
+        self.half_channels = in_channels // 2
+
+        self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
+        self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.0)
+        self.proj = nn.Conv1d(
+            filter_channels, self.half_channels * (num_bins * 3 - 1), 1
+        )
+        self.proj.weight.data.zero_()
+        self.proj.bias.data.zero_()
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        g: Optional[torch.Tensor] = None,
+        reverse=False,
+    ):
+        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+        h = self.pre(x0)
+        h = self.convs(h, x_mask, g=g)
+        h = self.proj(h) * x_mask
+
+        b, c, t = x0.shape
+        h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2)  # [b, cx?, t] -> [b, c, t, ?]
+
+        unnormalized_widths = h[..., : self.num_bins] / math.sqrt(self.filter_channels)
+        unnormalized_heights = h[..., self.num_bins : 2 * self.num_bins] / math.sqrt(
+            self.filter_channels
+        )
+        unnormalized_derivatives = h[..., 2 * self.num_bins :]
+
+        x1, logabsdet = piecewise_rational_quadratic_transform(
+            x1,
+            unnormalized_widths,
+            unnormalized_heights,
+            unnormalized_derivatives,
+            inverse=reverse,
+            tails="linear",
+            tail_bound=self.tail_bound,
+        )
+
+        x = torch.cat([x0, x1], 1) * x_mask
+        logdet = torch.sum(logabsdet * x_mask, [1, 2])
+        if not reverse:
+            return x, logdet
+        else:
+            return x

infer/lib/infer_pack/modules/F0Predictor/DioF0Predictor.py

@@ -0,0 +1,91 @@
+import numpy as np
+import pyworld
+
+from infer.lib.infer_pack.modules.F0Predictor.F0Predictor import F0Predictor
+
+
+class DioF0Predictor(F0Predictor):
+    def __init__(self, hop_length=512, f0_min=50, f0_max=1100, sampling_rate=44100):
+        self.hop_length = hop_length
+        self.f0_min = f0_min
+        self.f0_max = f0_max
+        self.sampling_rate = sampling_rate
+
+    def interpolate_f0(self, f0):
+        """
+        对F0进行插值处理
+        """
+
+        data = np.reshape(f0, (f0.size, 1))
+
+        vuv_vector = np.zeros((data.size, 1), dtype=np.float32)
+        vuv_vector[data > 0.0] = 1.0
+        vuv_vector[data <= 0.0] = 0.0
+
+        ip_data = data
+
+        frame_number = data.size
+        last_value = 0.0
+        for i in range(frame_number):
+            if data[i] <= 0.0:
+                j = i + 1
+                for j in range(i + 1, frame_number):
+                    if data[j] > 0.0:
+                        break
+                if j < frame_number - 1:
+                    if last_value > 0.0:
+                        step = (data[j] - data[i - 1]) / float(j - i)
+                        for k in range(i, j):
+                            ip_data[k] = data[i - 1] + step * (k - i + 1)
+                    else:
+                        for k in range(i, j):
+                            ip_data[k] = data[j]
+                else:
+                    for k in range(i, frame_number):
+                        ip_data[k] = last_value
+            else:
+                ip_data[i] = data[i]  # 这里可能存在一个没有必要的拷贝
+                last_value = data[i]
+
+        return ip_data[:, 0], vuv_vector[:, 0]
+
+    def resize_f0(self, x, target_len):
+        source = np.array(x)
+        source[source < 0.001] = np.nan
+        target = np.interp(
+            np.arange(0, len(source) * target_len, len(source)) / target_len,
+            np.arange(0, len(source)),
+            source,
+        )
+        res = np.nan_to_num(target)
+        return res
+
+    def compute_f0(self, wav, p_len=None):
+        if p_len is None:
+            p_len = wav.shape[0] // self.hop_length
+        f0, t = pyworld.dio(
+            wav.astype(np.double),
+            fs=self.sampling_rate,
+            f0_floor=self.f0_min,
+            f0_ceil=self.f0_max,
+            frame_period=1000 * self.hop_length / self.sampling_rate,
+        )
+        f0 = pyworld.stonemask(wav.astype(np.double), f0, t, self.sampling_rate)
+        for index, pitch in enumerate(f0):
+            f0[index] = round(pitch, 1)
+        return self.interpolate_f0(self.resize_f0(f0, p_len))[0]
+
+    def compute_f0_uv(self, wav, p_len=None):
+        if p_len is None:
+            p_len = wav.shape[0] // self.hop_length
+        f0, t = pyworld.dio(
+            wav.astype(np.double),
+            fs=self.sampling_rate,
+            f0_floor=self.f0_min,
+            f0_ceil=self.f0_max,
+            frame_period=1000 * self.hop_length / self.sampling_rate,
+        )
+        f0 = pyworld.stonemask(wav.astype(np.double), f0, t, self.sampling_rate)
+        for index, pitch in enumerate(f0):
+            f0[index] = round(pitch, 1)
+        return self.interpolate_f0(self.resize_f0(f0, p_len))

infer/lib/infer_pack/modules/F0Predictor/F0Predictor.py

@@ -0,0 +1,16 @@
+class F0Predictor(object):
+    def compute_f0(self, wav, p_len):
+        """
+        input: wav:[signal_length]
+               p_len:int
+        output: f0:[signal_length//hop_length]
+        """
+        pass
+
+    def compute_f0_uv(self, wav, p_len):
+        """
+        input: wav:[signal_length]
+               p_len:int
+        output: f0:[signal_length//hop_length],uv:[signal_length//hop_length]
+        """
+        pass

infer/lib/infer_pack/modules/F0Predictor/HarvestF0Predictor.py

@@ -0,0 +1,87 @@
+import numpy as np
+import pyworld
+
+from infer.lib.infer_pack.modules.F0Predictor.F0Predictor import F0Predictor
+
+
+class HarvestF0Predictor(F0Predictor):
+    def __init__(self, hop_length=512, f0_min=50, f0_max=1100, sampling_rate=44100):
+        self.hop_length = hop_length
+        self.f0_min = f0_min
+        self.f0_max = f0_max
+        self.sampling_rate = sampling_rate
+
+    def interpolate_f0(self, f0):
+        """
+        对F0进行插值处理
+        """
+
+        data = np.reshape(f0, (f0.size, 1))
+
+        vuv_vector = np.zeros((data.size, 1), dtype=np.float32)
+        vuv_vector[data > 0.0] = 1.0
+        vuv_vector[data <= 0.0] = 0.0
+
+        ip_data = data
+
+        frame_number = data.size
+        last_value = 0.0
+        for i in range(frame_number):
+            if data[i] <= 0.0:
+                j = i + 1
+                for j in range(i + 1, frame_number):
+                    if data[j] > 0.0:
+                        break
+                if j < frame_number - 1:
+                    if last_value > 0.0:
+                        step = (data[j] - data[i - 1]) / float(j - i)
+                        for k in range(i, j):
+                            ip_data[k] = data[i - 1] + step * (k - i + 1)
+                    else:
+                        for k in range(i, j):
+                            ip_data[k] = data[j]
+                else:
+                    for k in range(i, frame_number):
+                        ip_data[k] = last_value
+            else:
+                ip_data[i] = data[i]  # 这里可能存在一个没有必要的拷贝
+                last_value = data[i]
+
+        return ip_data[:, 0], vuv_vector[:, 0]
+
+    def resize_f0(self, x, target_len):
+        source = np.array(x)
+        source[source < 0.001] = np.nan
+        target = np.interp(
+            np.arange(0, len(source) * target_len, len(source)) / target_len,
+            np.arange(0, len(source)),
+            source,
+        )
+        res = np.nan_to_num(target)
+        return res
+
+    def compute_f0(self, wav, p_len=None):
+        if p_len is None:
+            p_len = wav.shape[0] // self.hop_length
+        f0, t = pyworld.harvest(
+            wav.astype(np.double),
+            fs=self.sampling_rate,
+            f0_ceil=self.f0_max,
+            f0_floor=self.f0_min,
+            frame_period=1000 * self.hop_length / self.sampling_rate,
+        )
+        f0 = pyworld.stonemask(wav.astype(np.double), f0, t, self.fs)
+        return self.interpolate_f0(self.resize_f0(f0, p_len))[0]
+
+    def compute_f0_uv(self, wav, p_len=None):
+        if p_len is None:
+            p_len = wav.shape[0] // self.hop_length
+        f0, t = pyworld.harvest(
+            wav.astype(np.double),
+            fs=self.sampling_rate,
+            f0_floor=self.f0_min,
+            f0_ceil=self.f0_max,
+            frame_period=1000 * self.hop_length / self.sampling_rate,
+        )
+        f0 = pyworld.stonemask(wav.astype(np.double), f0, t, self.sampling_rate)
+        return self.interpolate_f0(self.resize_f0(f0, p_len))

infer/lib/infer_pack/modules/F0Predictor/PMF0Predictor.py

@@ -0,0 +1,98 @@
+import numpy as np
+import parselmouth
+
+from infer.lib.infer_pack.modules.F0Predictor.F0Predictor import F0Predictor
+
+
+class PMF0Predictor(F0Predictor):
+    def __init__(self, hop_length=512, f0_min=50, f0_max=1100, sampling_rate=44100):
+        self.hop_length = hop_length
+        self.f0_min = f0_min
+        self.f0_max = f0_max
+        self.sampling_rate = sampling_rate
+
+    def interpolate_f0(self, f0):
+        """
+        对F0进行插值处理
+        """
+
+        data = np.reshape(f0, (f0.size, 1))
+
+        vuv_vector = np.zeros((data.size, 1), dtype=np.float32)
+        vuv_vector[data > 0.0] = 1.0
+        vuv_vector[data <= 0.0] = 0.0
+
+        ip_data = data
+
+        frame_number = data.size
+        last_value = 0.0
+        for i in range(frame_number):
+            if data[i] <= 0.0:
+                j = i + 1
+                for j in range(i + 1, frame_number):
+                    if data[j] > 0.0:
+                        break
+                if j < frame_number - 1:
+                    if last_value > 0.0:
+                        step = (data[j] - data[i - 1]) / float(j - i)
+                        for k in range(i, j):
+                            ip_data[k] = data[i - 1] + step * (k - i + 1)
+                    else:
+                        for k in range(i, j):
+                            ip_data[k] = data[j]
+                else:
+                    for k in range(i, frame_number):
+                        ip_data[k] = last_value
+            else:
+                ip_data[i] = data[i]  # 这里可能存在一个没有必要的拷贝
+                last_value = data[i]
+
+        return ip_data[:, 0], vuv_vector[:, 0]
+
+    def compute_f0(self, wav, p_len=None):
+        x = wav
+        if p_len is None:
+            p_len = x.shape[0] // self.hop_length
+        else:
+            assert abs(p_len - x.shape[0] // self.hop_length) < 4, "pad length error"
+        time_step = self.hop_length / self.sampling_rate * 1000
+        f0 = (
+            parselmouth.Sound(x, self.sampling_rate)
+            .to_pitch_ac(
+                time_step=time_step / 1000,
+                voicing_threshold=0.6,
+                pitch_floor=self.f0_min,
+                pitch_ceiling=self.f0_max,
+            )
+            .selected_array["frequency"]
+        )
+
+        pad_size = (p_len - len(f0) + 1) // 2
+        if pad_size > 0 or p_len - len(f0) - pad_size > 0:
+            f0 = np.pad(f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant")
+        f0, uv = self.interpolate_f0(f0)
+        return f0
+
+    def compute_f0_uv(self, wav, p_len=None):
+        x = wav
+        if p_len is None:
+            p_len = x.shape[0] // self.hop_length
+        else:
+            assert abs(p_len - x.shape[0] // self.hop_length) < 4, "pad length error"
+        time_step = self.hop_length / self.sampling_rate * 1000
+        f0 = (
+            parselmouth.Sound(x, self.sampling_rate)
+            .to_pitch_ac(
+                time_step=time_step / 1000,
+                voicing_threshold=0.6,
+                pitch_floor=self.f0_min,
+                pitch_ceiling=self.f0_max,
+            )
+            .selected_array["frequency"]
+        )
+
+        pad_size = (p_len - len(f0) + 1) // 2
+        if pad_size > 0 or p_len - len(f0) - pad_size > 0:
+            f0 = np.pad(f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant")
+        f0, uv = self.interpolate_f0(f0)
+        return f0, uv

infer/lib/infer_pack/onnx_inference.py

@@ -0,0 +1,149 @@
+import librosa
+import numpy as np
+import onnxruntime
+import soundfile
+
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class ContentVec:
+    def __init__(self, vec_path="pretrained/vec-768-layer-12.onnx", device=None):
+        logger.info("Load model(s) from {}".format(vec_path))
+        if device == "cpu" or device is None:
+            providers = ["CPUExecutionProvider"]
+        elif device == "cuda":
+            providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
+        elif device == "dml":
+            providers = ["DmlExecutionProvider"]
+        else:
+            raise RuntimeError("Unsportted Device")
+        self.model = onnxruntime.InferenceSession(vec_path, providers=providers)
+
+    def __call__(self, wav):
+        return self.forward(wav)
+
+    def forward(self, wav):
+        feats = wav
+        if feats.ndim == 2:  # double channels
+            feats = feats.mean(-1)
+        assert feats.ndim == 1, feats.ndim
+        feats = np.expand_dims(np.expand_dims(feats, 0), 0)
+        onnx_input = {self.model.get_inputs()[0].name: feats}
+        logits = self.model.run(None, onnx_input)[0]
+        return logits.transpose(0, 2, 1)
+
+
+def get_f0_predictor(f0_predictor, hop_length, sampling_rate, **kargs):
+    if f0_predictor == "pm":
+        from lib.infer_pack.modules.F0Predictor.PMF0Predictor import PMF0Predictor
+
+        f0_predictor_object = PMF0Predictor(
+            hop_length=hop_length, sampling_rate=sampling_rate
+        )
+    elif f0_predictor == "harvest":
+        from lib.infer_pack.modules.F0Predictor.HarvestF0Predictor import (
+            HarvestF0Predictor,
+        )
+
+        f0_predictor_object = HarvestF0Predictor(
+            hop_length=hop_length, sampling_rate=sampling_rate
+        )
+    elif f0_predictor == "dio":
+        from lib.infer_pack.modules.F0Predictor.DioF0Predictor import DioF0Predictor
+
+        f0_predictor_object = DioF0Predictor(
+            hop_length=hop_length, sampling_rate=sampling_rate
+        )
+    else:
+        raise Exception("Unknown f0 predictor")
+    return f0_predictor_object
+
+
+class OnnxRVC:
+    def __init__(
+        self,
+        model_path,
+        sr=40000,
+        hop_size=512,
+        vec_path="vec-768-layer-12",
+        device="cpu",
+    ):
+        vec_path = f"pretrained/{vec_path}.onnx"
+        self.vec_model = ContentVec(vec_path, device)
+        if device == "cpu" or device is None:
+            providers = ["CPUExecutionProvider"]
+        elif device == "cuda":
+            providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
+        elif device == "dml":
+            providers = ["DmlExecutionProvider"]
+        else:
+            raise RuntimeError("Unsportted Device")
+        self.model = onnxruntime.InferenceSession(model_path, providers=providers)
+        self.sampling_rate = sr
+        self.hop_size = hop_size
+
+    def forward(self, hubert, hubert_length, pitch, pitchf, ds, rnd):
+        onnx_input = {
+            self.model.get_inputs()[0].name: hubert,
+            self.model.get_inputs()[1].name: hubert_length,
+            self.model.get_inputs()[2].name: pitch,
+            self.model.get_inputs()[3].name: pitchf,
+            self.model.get_inputs()[4].name: ds,
+            self.model.get_inputs()[5].name: rnd,
+        }
+        return (self.model.run(None, onnx_input)[0] * 32767).astype(np.int16)
+
+    def inference(
+        self,
+        raw_path,
+        sid,
+        f0_method="dio",
+        f0_up_key=0,
+        pad_time=0.5,
+        cr_threshold=0.02,
+    ):
+        f0_min = 50
+        f0_max = 1100
+        f0_mel_min = 1127 * np.log(1 + f0_min / 700)
+        f0_mel_max = 1127 * np.log(1 + f0_max / 700)
+        f0_predictor = get_f0_predictor(
+            f0_method,
+            hop_length=self.hop_size,
+            sampling_rate=self.sampling_rate,
+            threshold=cr_threshold,
+        )
+        wav, sr = librosa.load(raw_path, sr=self.sampling_rate)
+        org_length = len(wav)
+        if org_length / sr > 50.0:
+            raise RuntimeError("Reached Max Length")
+
+        wav16k = librosa.resample(wav, orig_sr=self.sampling_rate, target_sr=16000)
+        wav16k = wav16k
+
+        hubert = self.vec_model(wav16k)
+        hubert = np.repeat(hubert, 2, axis=2).transpose(0, 2, 1).astype(np.float32)
+        hubert_length = hubert.shape[1]
+
+        pitchf = f0_predictor.compute_f0(wav, hubert_length)
+        pitchf = pitchf * 2 ** (f0_up_key / 12)
+        pitch = pitchf.copy()
+        f0_mel = 1127 * np.log(1 + pitch / 700)
+        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (
+            f0_mel_max - f0_mel_min
+        ) + 1
+        f0_mel[f0_mel <= 1] = 1
+        f0_mel[f0_mel > 255] = 255
+        pitch = np.rint(f0_mel).astype(np.int64)
+
+        pitchf = pitchf.reshape(1, len(pitchf)).astype(np.float32)
+        pitch = pitch.reshape(1, len(pitch))
+        ds = np.array([sid]).astype(np.int64)
+
+        rnd = np.random.randn(1, 192, hubert_length).astype(np.float32)
+        hubert_length = np.array([hubert_length]).astype(np.int64)
+
+        out_wav = self.forward(hubert, hubert_length, pitch, pitchf, ds, rnd).squeeze()
+        out_wav = np.pad(out_wav, (0, 2 * self.hop_size), "constant")
+        return out_wav[0:org_length]

infer/lib/infer_pack/transforms.py

@@ -0,0 +1,207 @@
+import numpy as np
+import torch
+from torch.nn import functional as F
+
+DEFAULT_MIN_BIN_WIDTH = 1e-3
+DEFAULT_MIN_BIN_HEIGHT = 1e-3
+DEFAULT_MIN_DERIVATIVE = 1e-3
+
+
+def piecewise_rational_quadratic_transform(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    tails=None,
+    tail_bound=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    if tails is None:
+        spline_fn = rational_quadratic_spline
+        spline_kwargs = {}
+    else:
+        spline_fn = unconstrained_rational_quadratic_spline
+        spline_kwargs = {"tails": tails, "tail_bound": tail_bound}
+
+    outputs, logabsdet = spline_fn(
+        inputs=inputs,
+        unnormalized_widths=unnormalized_widths,
+        unnormalized_heights=unnormalized_heights,
+        unnormalized_derivatives=unnormalized_derivatives,
+        inverse=inverse,
+        min_bin_width=min_bin_width,
+        min_bin_height=min_bin_height,
+        min_derivative=min_derivative,
+        **spline_kwargs
+    )
+    return outputs, logabsdet
+
+
+def searchsorted(bin_locations, inputs, eps=1e-6):
+    bin_locations[..., -1] += eps
+    return torch.sum(inputs[..., None] >= bin_locations, dim=-1) - 1
+
+
+def unconstrained_rational_quadratic_spline(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    tails="linear",
+    tail_bound=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    inside_interval_mask = (inputs >= -tail_bound) & (inputs <= tail_bound)
+    outside_interval_mask = ~inside_interval_mask
+
+    outputs = torch.zeros_like(inputs)
+    logabsdet = torch.zeros_like(inputs)
+
+    if tails == "linear":
+        unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1))
+        constant = np.log(np.exp(1 - min_derivative) - 1)
+        unnormalized_derivatives[..., 0] = constant
+        unnormalized_derivatives[..., -1] = constant
+
+        outputs[outside_interval_mask] = inputs[outside_interval_mask]
+        logabsdet[outside_interval_mask] = 0
+    else:
+        raise RuntimeError("{} tails are not implemented.".format(tails))
+
+    (
+        outputs[inside_interval_mask],
+        logabsdet[inside_interval_mask],
+    ) = rational_quadratic_spline(
+        inputs=inputs[inside_interval_mask],
+        unnormalized_widths=unnormalized_widths[inside_interval_mask, :],
+        unnormalized_heights=unnormalized_heights[inside_interval_mask, :],
+        unnormalized_derivatives=unnormalized_derivatives[inside_interval_mask, :],
+        inverse=inverse,
+        left=-tail_bound,
+        right=tail_bound,
+        bottom=-tail_bound,
+        top=tail_bound,
+        min_bin_width=min_bin_width,
+        min_bin_height=min_bin_height,
+        min_derivative=min_derivative,
+    )
+
+    return outputs, logabsdet
+
+
+def rational_quadratic_spline(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    left=0.0,
+    right=1.0,
+    bottom=0.0,
+    top=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    if torch.min(inputs) < left or torch.max(inputs) > right:
+        raise ValueError("Input to a transform is not within its domain")
+
+    num_bins = unnormalized_widths.shape[-1]
+
+    if min_bin_width * num_bins > 1.0:
+        raise ValueError("Minimal bin width too large for the number of bins")
+    if min_bin_height * num_bins > 1.0:
+        raise ValueError("Minimal bin height too large for the number of bins")
+
+    widths = F.softmax(unnormalized_widths, dim=-1)
+    widths = min_bin_width + (1 - min_bin_width * num_bins) * widths
+    cumwidths = torch.cumsum(widths, dim=-1)
+    cumwidths = F.pad(cumwidths, pad=(1, 0), mode="constant", value=0.0)
+    cumwidths = (right - left) * cumwidths + left
+    cumwidths[..., 0] = left
+    cumwidths[..., -1] = right
+    widths = cumwidths[..., 1:] - cumwidths[..., :-1]
+
+    derivatives = min_derivative + F.softplus(unnormalized_derivatives)
+
+    heights = F.softmax(unnormalized_heights, dim=-1)
+    heights = min_bin_height + (1 - min_bin_height * num_bins) * heights
+    cumheights = torch.cumsum(heights, dim=-1)
+    cumheights = F.pad(cumheights, pad=(1, 0), mode="constant", value=0.0)
+    cumheights = (top - bottom) * cumheights + bottom
+    cumheights[..., 0] = bottom
+    cumheights[..., -1] = top
+    heights = cumheights[..., 1:] - cumheights[..., :-1]
+
+    if inverse:
+        bin_idx = searchsorted(cumheights, inputs)[..., None]
+    else:
+        bin_idx = searchsorted(cumwidths, inputs)[..., None]
+
+    input_cumwidths = cumwidths.gather(-1, bin_idx)[..., 0]
+    input_bin_widths = widths.gather(-1, bin_idx)[..., 0]
+
+    input_cumheights = cumheights.gather(-1, bin_idx)[..., 0]
+    delta = heights / widths
+    input_delta = delta.gather(-1, bin_idx)[..., 0]
+
+    input_derivatives = derivatives.gather(-1, bin_idx)[..., 0]
+    input_derivatives_plus_one = derivatives[..., 1:].gather(-1, bin_idx)[..., 0]
+
+    input_heights = heights.gather(-1, bin_idx)[..., 0]
+
+    if inverse:
+        a = (inputs - input_cumheights) * (
+            input_derivatives + input_derivatives_plus_one - 2 * input_delta
+        ) + input_heights * (input_delta - input_derivatives)
+        b = input_heights * input_derivatives - (inputs - input_cumheights) * (
+            input_derivatives + input_derivatives_plus_one - 2 * input_delta
+        )
+        c = -input_delta * (inputs - input_cumheights)
+
+        discriminant = b.pow(2) - 4 * a * c
+        assert (discriminant >= 0).all()
+
+        root = (2 * c) / (-b - torch.sqrt(discriminant))
+        outputs = root * input_bin_widths + input_cumwidths
+
+        theta_one_minus_theta = root * (1 - root)
+        denominator = input_delta + (
+            (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
+            * theta_one_minus_theta
+        )
+        derivative_numerator = input_delta.pow(2) * (
+            input_derivatives_plus_one * root.pow(2)
+            + 2 * input_delta * theta_one_minus_theta
+            + input_derivatives * (1 - root).pow(2)
+        )
+        logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
+
+        return outputs, -logabsdet
+    else:
+        theta = (inputs - input_cumwidths) / input_bin_widths
+        theta_one_minus_theta = theta * (1 - theta)
+
+        numerator = input_heights * (
+            input_delta * theta.pow(2) + input_derivatives * theta_one_minus_theta
+        )
+        denominator = input_delta + (
+            (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
+            * theta_one_minus_theta
+        )
+        outputs = input_cumheights + numerator / denominator
+
+        derivative_numerator = input_delta.pow(2) * (
+            input_derivatives_plus_one * theta.pow(2)
+            + 2 * input_delta * theta_one_minus_theta
+            + input_derivatives * (1 - theta).pow(2)
+        )
+        logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
+
+        return outputs, logabsdet

infer/lib/jit/__init__.py

@@ -0,0 +1,163 @@
+from io import BytesIO
+import pickle
+import time
+import torch
+from tqdm import tqdm
+from collections import OrderedDict
+
+
+def load_inputs(path, device, is_half=False):
+    parm = torch.load(path, map_location=torch.device("cpu"))
+    for key in parm.keys():
+        parm[key] = parm[key].to(device)
+        if is_half and parm[key].dtype == torch.float32:
+            parm[key] = parm[key].half()
+        elif not is_half and parm[key].dtype == torch.float16:
+            parm[key] = parm[key].float()
+    return parm
+
+
+def benchmark(
+    model, inputs_path, device=torch.device("cpu"), epoch=1000, is_half=False
+):
+    parm = load_inputs(inputs_path, device, is_half)
+    total_ts = 0.0
+    bar = tqdm(range(epoch))
+    for i in bar:
+        start_time = time.perf_counter()
+        o = model(**parm)
+        total_ts += time.perf_counter() - start_time
+    print(f"num_epoch: {epoch} | avg time(ms): {(total_ts*1000)/epoch}")
+
+
+def jit_warm_up(model, inputs_path, device=torch.device("cpu"), epoch=5, is_half=False):
+    benchmark(model, inputs_path, device, epoch=epoch, is_half=is_half)
+
+
+def to_jit_model(
+    model_path,
+    model_type: str,
+    mode: str = "trace",
+    inputs_path: str = None,
+    device=torch.device("cpu"),
+    is_half=False,
+):
+    model = None
+    if model_type.lower() == "synthesizer":
+        from .get_synthesizer import get_synthesizer
+
+        model, _ = get_synthesizer(model_path, device)
+        model.forward = model.infer
+    elif model_type.lower() == "rmvpe":
+        from .get_rmvpe import get_rmvpe
+
+        model = get_rmvpe(model_path, device)
+    elif model_type.lower() == "hubert":
+        from .get_hubert import get_hubert_model
+
+        model = get_hubert_model(model_path, device)
+        model.forward = model.infer
+    else:
+        raise ValueError(f"No model type named {model_type}")
+    model = model.eval()
+    model = model.half() if is_half else model.float()
+    if mode == "trace":
+        assert not inputs_path
+        inputs = load_inputs(inputs_path, device, is_half)
+        model_jit = torch.jit.trace(model, example_kwarg_inputs=inputs)
+    elif mode == "script":
+        model_jit = torch.jit.script(model)
+    model_jit.to(device)
+    model_jit = model_jit.half() if is_half else model_jit.float()
+    # model = model.half() if is_half else model.float()
+    return (model, model_jit)
+
+
+def export(
+    model: torch.nn.Module,
+    mode: str = "trace",
+    inputs: dict = None,
+    device=torch.device("cpu"),
+    is_half: bool = False,
+) -> dict:
+    model = model.half() if is_half else model.float()
+    model.eval()
+    if mode == "trace":
+        assert inputs is not None
+        model_jit = torch.jit.trace(model, example_kwarg_inputs=inputs)
+    elif mode == "script":
+        model_jit = torch.jit.script(model)
+    model_jit.to(device)
+    model_jit = model_jit.half() if is_half else model_jit.float()
+    buffer = BytesIO()
+    # model_jit=model_jit.cpu()
+    torch.jit.save(model_jit, buffer)
+    del model_jit
+    cpt = OrderedDict()
+    cpt["model"] = buffer.getvalue()
+    cpt["is_half"] = is_half
+    return cpt
+
+
+def load(path: str):
+    with open(path, "rb") as f:
+        return pickle.load(f)
+
+
+def save(ckpt: dict, save_path: str):
+    with open(save_path, "wb") as f:
+        pickle.dump(ckpt, f)
+
+
+def rmvpe_jit_export(
+    model_path: str,
+    mode: str = "script",
+    inputs_path: str = None,
+    save_path: str = None,
+    device=torch.device("cpu"),
+    is_half=False,
+):
+    if not save_path:
+        save_path = model_path.rstrip(".pth")
+        save_path += ".half.jit" if is_half else ".jit"
+    if "cuda" in str(device) and ":" not in str(device):
+        device = torch.device("cuda:0")
+    from .get_rmvpe import get_rmvpe
+
+    model = get_rmvpe(model_path, device)
+    inputs = None
+    if mode == "trace":
+        inputs = load_inputs(inputs_path, device, is_half)
+    ckpt = export(model, mode, inputs, device, is_half)
+    ckpt["device"] = str(device)
+    save(ckpt, save_path)
+    return ckpt
+
+
+def synthesizer_jit_export(
+    model_path: str,
+    mode: str = "script",
+    inputs_path: str = None,
+    save_path: str = None,
+    device=torch.device("cpu"),
+    is_half=False,
+):
+    if not save_path:
+        save_path = model_path.rstrip(".pth")
+        save_path += ".half.jit" if is_half else ".jit"
+    if "cuda" in str(device) and ":" not in str(device):
+        device = torch.device("cuda:0")
+    from .get_synthesizer import get_synthesizer
+
+    model, cpt = get_synthesizer(model_path, device)
+    assert isinstance(cpt, dict)
+    model.forward = model.infer
+    inputs = None
+    if mode == "trace":
+        inputs = load_inputs(inputs_path, device, is_half)
+    ckpt = export(model, mode, inputs, device, is_half)
+    cpt.pop("weight")
+    cpt["model"] = ckpt["model"]
+    cpt["device"] = device
+    save(cpt, save_path)
+    return cpt

infer/lib/jit/get_hubert.py

@@ -0,0 +1,342 @@
+import math
+import random
+from typing import Optional, Tuple
+from fairseq.checkpoint_utils import load_model_ensemble_and_task
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+# from fairseq.data.data_utils import compute_mask_indices
+from fairseq.utils import index_put
+
+
+# @torch.jit.script
+def pad_to_multiple(x, multiple, dim=-1, value=0):
+    # Inspired from https://github.com/lucidrains/local-attention/blob/master/local_attention/local_attention.py#L41
+    if x is None:
+        return None, 0
+    tsz = x.size(dim)
+    m = tsz / multiple
+    remainder = math.ceil(m) * multiple - tsz
+    if int(tsz % multiple) == 0:
+        return x, 0
+    pad_offset = (0,) * (-1 - dim) * 2
+
+    return F.pad(x, (*pad_offset, 0, remainder), value=value), remainder
+
+
+def extract_features(
+    self,
+    x,
+    padding_mask=None,
+    tgt_layer=None,
+    min_layer=0,
+):
+    if padding_mask is not None:
+        x = index_put(x, padding_mask, 0)
+
+    x_conv = self.pos_conv(x.transpose(1, 2))
+    x_conv = x_conv.transpose(1, 2)
+    x = x + x_conv
+
+    if not self.layer_norm_first:
+        x = self.layer_norm(x)
+
+    # pad to the sequence length dimension
+    x, pad_length = pad_to_multiple(x, self.required_seq_len_multiple, dim=-2, value=0)
+    if pad_length > 0 and padding_mask is None:
+        padding_mask = x.new_zeros((x.size(0), x.size(1)), dtype=torch.bool)
+        padding_mask[:, -pad_length:] = True
+    else:
+        padding_mask, _ = pad_to_multiple(
+            padding_mask, self.required_seq_len_multiple, dim=-1, value=True
+        )
+    x = F.dropout(x, p=self.dropout, training=self.training)
+
+    # B x T x C -> T x B x C
+    x = x.transpose(0, 1)
+
+    layer_results = []
+    r = None
+    for i, layer in enumerate(self.layers):
+        dropout_probability = np.random.random() if self.layerdrop > 0 else 1
+        if not self.training or (dropout_probability > self.layerdrop):
+            x, (z, lr) = layer(
+                x, self_attn_padding_mask=padding_mask, need_weights=False
+            )
+            if i >= min_layer:
+                layer_results.append((x, z, lr))
+        if i == tgt_layer:
+            r = x
+            break
+
+    if r is not None:
+        x = r
+
+    # T x B x C -> B x T x C
+    x = x.transpose(0, 1)
+
+    # undo paddding
+    if pad_length > 0:
+        x = x[:, :-pad_length]
+
+        def undo_pad(a, b, c):
+            return (
+                a[:-pad_length],
+                b[:-pad_length] if b is not None else b,
+                c[:-pad_length],
+            )
+
+        layer_results = [undo_pad(*u) for u in layer_results]
+
+    return x, layer_results
+
+
+def compute_mask_indices(
+    shape: Tuple[int, int],
+    padding_mask: Optional[torch.Tensor],
+    mask_prob: float,
+    mask_length: int,
+    mask_type: str = "static",
+    mask_other: float = 0.0,
+    min_masks: int = 0,
+    no_overlap: bool = False,
+    min_space: int = 0,
+    require_same_masks: bool = True,
+    mask_dropout: float = 0.0,
+) -> torch.Tensor:
+    """
+    Computes random mask spans for a given shape
+
+    Args:
+        shape: the the shape for which to compute masks.
+            should be of size 2 where first element is batch size and 2nd is timesteps
+        padding_mask: optional padding mask of the same size as shape, which will prevent masking padded elements
+        mask_prob: probability for each token to be chosen as start of the span to be masked. this will be multiplied by
+            number of timesteps divided by length of mask span to mask approximately this percentage of all elements.
+            however due to overlaps, the actual number will be smaller (unless no_overlap is True)
+        mask_type: how to compute mask lengths
+            static = fixed size
+            uniform = sample from uniform distribution [mask_other, mask_length*2]
+            normal = sample from normal distribution with mean mask_length and stdev mask_other. mask is min 1 element
+            poisson = sample from possion distribution with lambda = mask length
+        min_masks: minimum number of masked spans
+        no_overlap: if false, will switch to an alternative recursive algorithm that prevents spans from overlapping
+        min_space: only used if no_overlap is True, this is how many elements to keep unmasked between spans
+        require_same_masks: if true, will randomly drop out masks until same amount of masks remains in each sample
+        mask_dropout: randomly dropout this percentage of masks in each example
+    """
+
+    bsz, all_sz = shape
+    mask = torch.full((bsz, all_sz), False)
+
+    all_num_mask = int(
+        # add a random number for probabilistic rounding
+        mask_prob * all_sz / float(mask_length)
+        + torch.rand([1]).item()
+    )
+
+    all_num_mask = max(min_masks, all_num_mask)
+
+    mask_idcs = []
+    for i in range(bsz):
+        if padding_mask is not None:
+            sz = all_sz - padding_mask[i].long().sum().item()
+            num_mask = int(mask_prob * sz / float(mask_length) + np.random.rand())
+            num_mask = max(min_masks, num_mask)
+        else:
+            sz = all_sz
+            num_mask = all_num_mask
+
+        if mask_type == "static":
+            lengths = torch.full([num_mask], mask_length)
+        elif mask_type == "uniform":
+            lengths = torch.randint(mask_other, mask_length * 2 + 1, size=[num_mask])
+        elif mask_type == "normal":
+            lengths = torch.normal(mask_length, mask_other, size=[num_mask])
+            lengths = [max(1, int(round(x))) for x in lengths]
+        else:
+            raise Exception("unknown mask selection " + mask_type)
+
+        if sum(lengths) == 0:
+            lengths[0] = min(mask_length, sz - 1)
+
+        if no_overlap:
+            mask_idc = []
+
+            def arrange(s, e, length, keep_length):
+                span_start = torch.randint(low=s, high=e - length, size=[1]).item()
+                mask_idc.extend(span_start + i for i in range(length))
+
+                new_parts = []
+                if span_start - s - min_space >= keep_length:
+                    new_parts.append((s, span_start - min_space + 1))
+                if e - span_start - length - min_space > keep_length:
+                    new_parts.append((span_start + length + min_space, e))
+                return new_parts
+
+            parts = [(0, sz)]
+            min_length = min(lengths)
+            for length in sorted(lengths, reverse=True):
+                t = [e - s if e - s >= length + min_space else 0 for s, e in parts]
+                lens = torch.asarray(t, dtype=torch.int)
+                l_sum = torch.sum(lens)
+                if l_sum == 0:
+                    break
+                probs = lens / torch.sum(lens)
+                c = torch.multinomial(probs.float(), len(parts)).item()
+                s, e = parts.pop(c)
+                parts.extend(arrange(s, e, length, min_length))
+            mask_idc = torch.asarray(mask_idc)
+        else:
+            min_len = min(lengths)
+            if sz - min_len <= num_mask:
+                min_len = sz - num_mask - 1
+            mask_idc = torch.asarray(
+                random.sample([i for i in range(sz - min_len)], num_mask)
+            )
+            mask_idc = torch.asarray(
+                [
+                    mask_idc[j] + offset
+                    for j in range(len(mask_idc))
+                    for offset in range(lengths[j])
+                ]
+            )
+
+        mask_idcs.append(torch.unique(mask_idc[mask_idc < sz]))
+
+    min_len = min([len(m) for m in mask_idcs])
+    for i, mask_idc in enumerate(mask_idcs):
+        if isinstance(mask_idc, torch.Tensor):
+            mask_idc = torch.asarray(mask_idc, dtype=torch.float)
+        if len(mask_idc) > min_len and require_same_masks:
+            mask_idc = torch.asarray(
+                random.sample([i for i in range(mask_idc)], min_len)
+            )
+        if mask_dropout > 0:
+            num_holes = int(round(len(mask_idc) * mask_dropout))
+            mask_idc = torch.asarray(
+                random.sample([i for i in range(mask_idc)], len(mask_idc) - num_holes)
+            )
+
+        mask[i, mask_idc.int()] = True
+
+    return mask
+
+
+def apply_mask(self, x, padding_mask, target_list):
+    B, T, C = x.shape
+    torch.zeros_like(x)
+    if self.mask_prob > 0:
+        mask_indices = compute_mask_indices(
+            (B, T),
+            padding_mask,
+            self.mask_prob,
+            self.mask_length,
+            self.mask_selection,
+            self.mask_other,
+            min_masks=2,
+            no_overlap=self.no_mask_overlap,
+            min_space=self.mask_min_space,
+        )
+        mask_indices = mask_indices.to(x.device)
+        x[mask_indices] = self.mask_emb
+    else:
+        mask_indices = None
+
+    if self.mask_channel_prob > 0:
+        mask_channel_indices = compute_mask_indices(
+            (B, C),
+            None,
+            self.mask_channel_prob,
+            self.mask_channel_length,
+            self.mask_channel_selection,
+            self.mask_channel_other,
+            no_overlap=self.no_mask_channel_overlap,
+            min_space=self.mask_channel_min_space,
+        )
+        mask_channel_indices = (
+            mask_channel_indices.to(x.device).unsqueeze(1).expand(-1, T, -1)
+        )
+        x[mask_channel_indices] = 0
+
+    return x, mask_indices
+
+
+def get_hubert_model(
+    model_path="assets/hubert/hubert_base.pt", device=torch.device("cpu")
+):
+    models, _, _ = load_model_ensemble_and_task(
+        [model_path],
+        suffix="",
+    )
+    hubert_model = models[0]
+    hubert_model = hubert_model.to(device)
+
+    def _apply_mask(x, padding_mask, target_list):
+        return apply_mask(hubert_model, x, padding_mask, target_list)
+
+    hubert_model.apply_mask = _apply_mask
+
+    def _extract_features(
+        x,
+        padding_mask=None,
+        tgt_layer=None,
+        min_layer=0,
+    ):
+        return extract_features(
+            hubert_model.encoder,
+            x,
+            padding_mask=padding_mask,
+            tgt_layer=tgt_layer,
+            min_layer=min_layer,
+        )
+
+    hubert_model.encoder.extract_features = _extract_features
+
+    hubert_model._forward = hubert_model.forward
+
+    def hubert_extract_features(
+        self,
+        source: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+        mask: bool = False,
+        ret_conv: bool = False,
+        output_layer: Optional[int] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        res = self._forward(
+            source,
+            padding_mask=padding_mask,
+            mask=mask,
+            features_only=True,
+            output_layer=output_layer,
+        )
+        feature = res["features"] if ret_conv else res["x"]
+        return feature, res["padding_mask"]
+
+    def _hubert_extract_features(
+        source: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+        mask: bool = False,
+        ret_conv: bool = False,
+        output_layer: Optional[int] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        return hubert_extract_features(
+            hubert_model, source, padding_mask, mask, ret_conv, output_layer
+        )
+
+    hubert_model.extract_features = _hubert_extract_features
+
+    def infer(source, padding_mask, output_layer: torch.Tensor):
+        output_layer = output_layer.item()
+        logits = hubert_model.extract_features(
+            source=source, padding_mask=padding_mask, output_layer=output_layer
+        )
+        feats = hubert_model.final_proj(logits[0]) if output_layer == 9 else logits[0]
+        return feats
+
+    hubert_model.infer = infer
+    # hubert_model.forward=infer
+    # hubert_model.forward
+
+    return hubert_model

infer/lib/jit/get_rmvpe.py

@@ -0,0 +1,12 @@
+import torch
+
+
+def get_rmvpe(model_path="assets/rmvpe/rmvpe.pt", device=torch.device("cpu")):
+    from infer.lib.rmvpe import E2E
+
+    model = E2E(4, 1, (2, 2))
+    ckpt = torch.load(model_path, map_location=device)
+    model.load_state_dict(ckpt)
+    model.eval()
+    model = model.to(device)
+    return model

infer/lib/jit/get_synthesizer.py

@@ -0,0 +1,38 @@
+import torch
+
+
+def get_synthesizer(pth_path, device=torch.device("cpu")):
+    from infer.lib.infer_pack.models import (
+        SynthesizerTrnMs256NSFsid,
+        SynthesizerTrnMs256NSFsid_nono,
+        SynthesizerTrnMs768NSFsid,
+        SynthesizerTrnMs768NSFsid_nono,
+    )
+
+    cpt = torch.load(pth_path, map_location=torch.device("cpu"))
+    # tgt_sr = cpt["config"][-1]
+    cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
+    if_f0 = cpt.get("f0", 1)
+    version = cpt.get("version", "v1")
+    if version == "v1":
+        if if_f0 == 1:
+            net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=False)
+        else:
+            net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
+    elif version == "v2":
+        if if_f0 == 1:
+            net_g = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=False)
+        else:
+            net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
+    del net_g.enc_q
+    # net_g.forward = net_g.infer
+    # ckpt = {}
+    # ckpt["config"] = cpt["config"]
+    # ckpt["f0"] = if_f0
+    # ckpt["version"] = version
+    # ckpt["info"] = cpt.get("info", "0epoch")
+    net_g.load_state_dict(cpt["weight"], strict=False)
+    net_g = net_g.float()
+    net_g.eval().to(device)
+    net_g.remove_weight_norm()
+    return net_g, cpt

infer/lib/rmvpe.py

@@ -0,0 +1,670 @@
+from io import BytesIO
+import os
+from typing import List, Optional, Tuple
+import numpy as np
+import torch
+
+from infer.lib import jit
+
+try:
+    # Fix "Torch not compiled with CUDA enabled"
+    import intel_extension_for_pytorch as ipex  # pylint: disable=import-error, unused-import
+
+    if torch.xpu.is_available():
+        from infer.modules.ipex import ipex_init
+
+        ipex_init()
+except Exception:  # pylint: disable=broad-exception-caught
+    pass
+import torch.nn as nn
+import torch.nn.functional as F
+from librosa.util import normalize, pad_center, tiny
+from scipy.signal import get_window
+
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class STFT(torch.nn.Module):
+    def __init__(
+        self, filter_length=1024, hop_length=512, win_length=None, window="hann"
+    ):
+        """
+        This module implements an STFT using 1D convolution and 1D transpose convolutions.
+        This is a bit tricky so there are some cases that probably won't work as working
+        out the same sizes before and after in all overlap add setups is tough. Right now,
+        this code should work with hop lengths that are half the filter length (50% overlap
+        between frames).
+
+        Keyword Arguments:
+            filter_length {int} -- Length of filters used (default: {1024})
+            hop_length {int} -- Hop length of STFT (restrict to 50% overlap between frames) (default: {512})
+            win_length {[type]} -- Length of the window function applied to each frame (if not specified, it
+                equals the filter length). (default: {None})
+            window {str} -- Type of window to use (options are bartlett, hann, hamming, blackman, blackmanharris)
+                (default: {'hann'})
+        """
+        super(STFT, self).__init__()
+        self.filter_length = filter_length
+        self.hop_length = hop_length
+        self.win_length = win_length if win_length else filter_length
+        self.window = window
+        self.forward_transform = None
+        self.pad_amount = int(self.filter_length / 2)
+        fourier_basis = np.fft.fft(np.eye(self.filter_length))
+
+        cutoff = int((self.filter_length / 2 + 1))
+        fourier_basis = np.vstack(
+            [np.real(fourier_basis[:cutoff, :]), np.imag(fourier_basis[:cutoff, :])]
+        )
+        forward_basis = torch.FloatTensor(fourier_basis)
+        inverse_basis = torch.FloatTensor(np.linalg.pinv(fourier_basis))
+
+        assert filter_length >= self.win_length
+        # get window and zero center pad it to filter_length
+        fft_window = get_window(window, self.win_length, fftbins=True)
+        fft_window = pad_center(fft_window, size=filter_length)
+        fft_window = torch.from_numpy(fft_window).float()
+
+        # window the bases
+        forward_basis *= fft_window
+        inverse_basis = (inverse_basis.T * fft_window).T
+
+        self.register_buffer("forward_basis", forward_basis.float())
+        self.register_buffer("inverse_basis", inverse_basis.float())
+        self.register_buffer("fft_window", fft_window.float())
+
+    def transform(self, input_data, return_phase=False):
+        """Take input data (audio) to STFT domain.
+
+        Arguments:
+            input_data {tensor} -- Tensor of floats, with shape (num_batch, num_samples)
+
+        Returns:
+            magnitude {tensor} -- Magnitude of STFT with shape (num_batch,
+                num_frequencies, num_frames)
+            phase {tensor} -- Phase of STFT with shape (num_batch,
+                num_frequencies, num_frames)
+        """
+        input_data = F.pad(
+            input_data,
+            (self.pad_amount, self.pad_amount),
+            mode="reflect",
+        )
+        forward_transform = input_data.unfold(
+            1, self.filter_length, self.hop_length
+        ).permute(0, 2, 1)
+        forward_transform = torch.matmul(self.forward_basis, forward_transform)
+        cutoff = int((self.filter_length / 2) + 1)
+        real_part = forward_transform[:, :cutoff, :]
+        imag_part = forward_transform[:, cutoff:, :]
+        magnitude = torch.sqrt(real_part**2 + imag_part**2)
+        if return_phase:
+            phase = torch.atan2(imag_part.data, real_part.data)
+            return magnitude, phase
+        else:
+            return magnitude
+
+    def inverse(self, magnitude, phase):
+        """Call the inverse STFT (iSTFT), given magnitude and phase tensors produced
+        by the ```transform``` function.
+
+        Arguments:
+            magnitude {tensor} -- Magnitude of STFT with shape (num_batch,
+                num_frequencies, num_frames)
+            phase {tensor} -- Phase of STFT with shape (num_batch,
+                num_frequencies, num_frames)
+
+        Returns:
+            inverse_transform {tensor} -- Reconstructed audio given magnitude and phase. Of
+                shape (num_batch, num_samples)
+        """
+        cat = torch.cat(
+            [magnitude * torch.cos(phase), magnitude * torch.sin(phase)], dim=1
+        )
+        fold = torch.nn.Fold(
+            output_size=(1, (cat.size(-1) - 1) * self.hop_length + self.filter_length),
+            kernel_size=(1, self.filter_length),
+            stride=(1, self.hop_length),
+        )
+        inverse_transform = torch.matmul(self.inverse_basis, cat)
+        inverse_transform = fold(inverse_transform)[
+            :, 0, 0, self.pad_amount : -self.pad_amount
+        ]
+        window_square_sum = (
+            self.fft_window.pow(2).repeat(cat.size(-1), 1).T.unsqueeze(0)
+        )
+        window_square_sum = fold(window_square_sum)[
+            :, 0, 0, self.pad_amount : -self.pad_amount
+        ]
+        inverse_transform /= window_square_sum
+        return inverse_transform
+
+    def forward(self, input_data):
+        """Take input data (audio) to STFT domain and then back to audio.
+
+        Arguments:
+            input_data {tensor} -- Tensor of floats, with shape (num_batch, num_samples)
+
+        Returns:
+            reconstruction {tensor} -- Reconstructed audio given magnitude and phase. Of
+                shape (num_batch, num_samples)
+        """
+        self.magnitude, self.phase = self.transform(input_data, return_phase=True)
+        reconstruction = self.inverse(self.magnitude, self.phase)
+        return reconstruction
+
+
+from time import time as ttime
+
+
+class BiGRU(nn.Module):
+    def __init__(self, input_features, hidden_features, num_layers):
+        super(BiGRU, self).__init__()
+        self.gru = nn.GRU(
+            input_features,
+            hidden_features,
+            num_layers=num_layers,
+            batch_first=True,
+            bidirectional=True,
+        )
+
+    def forward(self, x):
+        return self.gru(x)[0]
+
+
+class ConvBlockRes(nn.Module):
+    def __init__(self, in_channels, out_channels, momentum=0.01):
+        super(ConvBlockRes, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=(3, 3),
+                stride=(1, 1),
+                padding=(1, 1),
+                bias=False,
+            ),
+            nn.BatchNorm2d(out_channels, momentum=momentum),
+            nn.ReLU(),
+            nn.Conv2d(
+                in_channels=out_channels,
+                out_channels=out_channels,
+                kernel_size=(3, 3),
+                stride=(1, 1),
+                padding=(1, 1),
+                bias=False,
+            ),
+            nn.BatchNorm2d(out_channels, momentum=momentum),
+            nn.ReLU(),
+        )
+        # self.shortcut:Optional[nn.Module] = None
+        if in_channels != out_channels:
+            self.shortcut = nn.Conv2d(in_channels, out_channels, (1, 1))
+
+    def forward(self, x: torch.Tensor):
+        if not hasattr(self, "shortcut"):
+            return self.conv(x) + x
+        else:
+            return self.conv(x) + self.shortcut(x)
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        in_size,
+        n_encoders,
+        kernel_size,
+        n_blocks,
+        out_channels=16,
+        momentum=0.01,
+    ):
+        super(Encoder, self).__init__()
+        self.n_encoders = n_encoders
+        self.bn = nn.BatchNorm2d(in_channels, momentum=momentum)
+        self.layers = nn.ModuleList()
+        self.latent_channels = []
+        for i in range(self.n_encoders):
+            self.layers.append(
+                ResEncoderBlock(
+                    in_channels, out_channels, kernel_size, n_blocks, momentum=momentum
+                )
+            )
+            self.latent_channels.append([out_channels, in_size])
+            in_channels = out_channels
+            out_channels *= 2
+            in_size //= 2
+        self.out_size = in_size
+        self.out_channel = out_channels
+
+    def forward(self, x: torch.Tensor):
+        concat_tensors: List[torch.Tensor] = []
+        x = self.bn(x)
+        for i, layer in enumerate(self.layers):
+            t, x = layer(x)
+            concat_tensors.append(t)
+        return x, concat_tensors
+
+
+class ResEncoderBlock(nn.Module):
+    def __init__(
+        self, in_channels, out_channels, kernel_size, n_blocks=1, momentum=0.01
+    ):
+        super(ResEncoderBlock, self).__init__()
+        self.n_blocks = n_blocks
+        self.conv = nn.ModuleList()
+        self.conv.append(ConvBlockRes(in_channels, out_channels, momentum))
+        for i in range(n_blocks - 1):
+            self.conv.append(ConvBlockRes(out_channels, out_channels, momentum))
+        self.kernel_size = kernel_size
+        if self.kernel_size is not None:
+            self.pool = nn.AvgPool2d(kernel_size=kernel_size)
+
+    def forward(self, x):
+        for i, conv in enumerate(self.conv):
+            x = conv(x)
+        if self.kernel_size is not None:
+            return x, self.pool(x)
+        else:
+            return x
+
+
+class Intermediate(nn.Module):  #
+    def __init__(self, in_channels, out_channels, n_inters, n_blocks, momentum=0.01):
+        super(Intermediate, self).__init__()
+        self.n_inters = n_inters
+        self.layers = nn.ModuleList()
+        self.layers.append(
+            ResEncoderBlock(in_channels, out_channels, None, n_blocks, momentum)
+        )
+        for i in range(self.n_inters - 1):
+            self.layers.append(
+                ResEncoderBlock(out_channels, out_channels, None, n_blocks, momentum)
+            )
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = layer(x)
+        return x
+
+
+class ResDecoderBlock(nn.Module):
+    def __init__(self, in_channels, out_channels, stride, n_blocks=1, momentum=0.01):
+        super(ResDecoderBlock, self).__init__()
+        out_padding = (0, 1) if stride == (1, 2) else (1, 1)
+        self.n_blocks = n_blocks
+        self.conv1 = nn.Sequential(
+            nn.ConvTranspose2d(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=(3, 3),
+                stride=stride,
+                padding=(1, 1),
+                output_padding=out_padding,
+                bias=False,
+            ),
+            nn.BatchNorm2d(out_channels, momentum=momentum),
+            nn.ReLU(),
+        )
+        self.conv2 = nn.ModuleList()
+        self.conv2.append(ConvBlockRes(out_channels * 2, out_channels, momentum))
+        for i in range(n_blocks - 1):
+            self.conv2.append(ConvBlockRes(out_channels, out_channels, momentum))
+
+    def forward(self, x, concat_tensor):
+        x = self.conv1(x)
+        x = torch.cat((x, concat_tensor), dim=1)
+        for i, conv2 in enumerate(self.conv2):
+            x = conv2(x)
+        return x
+
+
+class Decoder(nn.Module):
+    def __init__(self, in_channels, n_decoders, stride, n_blocks, momentum=0.01):
+        super(Decoder, self).__init__()
+        self.layers = nn.ModuleList()
+        self.n_decoders = n_decoders
+        for i in range(self.n_decoders):
+            out_channels = in_channels // 2
+            self.layers.append(
+                ResDecoderBlock(in_channels, out_channels, stride, n_blocks, momentum)
+            )
+            in_channels = out_channels
+
+    def forward(self, x: torch.Tensor, concat_tensors: List[torch.Tensor]):
+        for i, layer in enumerate(self.layers):
+            x = layer(x, concat_tensors[-1 - i])
+        return x
+
+
+class DeepUnet(nn.Module):
+    def __init__(
+        self,
+        kernel_size,
+        n_blocks,
+        en_de_layers=5,
+        inter_layers=4,
+        in_channels=1,
+        en_out_channels=16,
+    ):
+        super(DeepUnet, self).__init__()
+        self.encoder = Encoder(
+            in_channels, 128, en_de_layers, kernel_size, n_blocks, en_out_channels
+        )
+        self.intermediate = Intermediate(
+            self.encoder.out_channel // 2,
+            self.encoder.out_channel,
+            inter_layers,
+            n_blocks,
+        )
+        self.decoder = Decoder(
+            self.encoder.out_channel, en_de_layers, kernel_size, n_blocks
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x, concat_tensors = self.encoder(x)
+        x = self.intermediate(x)
+        x = self.decoder(x, concat_tensors)
+        return x
+
+
+class E2E(nn.Module):
+    def __init__(
+        self,
+        n_blocks,
+        n_gru,
+        kernel_size,
+        en_de_layers=5,
+        inter_layers=4,
+        in_channels=1,
+        en_out_channels=16,
+    ):
+        super(E2E, self).__init__()
+        self.unet = DeepUnet(
+            kernel_size,
+            n_blocks,
+            en_de_layers,
+            inter_layers,
+            in_channels,
+            en_out_channels,
+        )
+        self.cnn = nn.Conv2d(en_out_channels, 3, (3, 3), padding=(1, 1))
+        if n_gru:
+            self.fc = nn.Sequential(
+                BiGRU(3 * 128, 256, n_gru),
+                nn.Linear(512, 360),
+                nn.Dropout(0.25),
+                nn.Sigmoid(),
+            )
+        else:
+            self.fc = nn.Sequential(
+                nn.Linear(3 * nn.N_MELS, nn.N_CLASS), nn.Dropout(0.25), nn.Sigmoid()
+            )
+
+    def forward(self, mel):
+        # print(mel.shape)
+        mel = mel.transpose(-1, -2).unsqueeze(1)
+        x = self.cnn(self.unet(mel)).transpose(1, 2).flatten(-2)
+        x = self.fc(x)
+        # print(x.shape)
+        return x
+
+
+from librosa.filters import mel
+
+
+class MelSpectrogram(torch.nn.Module):
+    def __init__(
+        self,
+        is_half,
+        n_mel_channels,
+        sampling_rate,
+        win_length,
+        hop_length,
+        n_fft=None,
+        mel_fmin=0,
+        mel_fmax=None,
+        clamp=1e-5,
+    ):
+        super().__init__()
+        n_fft = win_length if n_fft is None else n_fft
+        self.hann_window = {}
+        mel_basis = mel(
+            sr=sampling_rate,
+            n_fft=n_fft,
+            n_mels=n_mel_channels,
+            fmin=mel_fmin,
+            fmax=mel_fmax,
+            htk=True,
+        )
+        mel_basis = torch.from_numpy(mel_basis).float()
+        self.register_buffer("mel_basis", mel_basis)
+        self.n_fft = win_length if n_fft is None else n_fft
+        self.hop_length = hop_length
+        self.win_length = win_length
+        self.sampling_rate = sampling_rate
+        self.n_mel_channels = n_mel_channels
+        self.clamp = clamp
+        self.is_half = is_half
+
+    def forward(self, audio, keyshift=0, speed=1, center=True):
+        factor = 2 ** (keyshift / 12)
+        n_fft_new = int(np.round(self.n_fft * factor))
+        win_length_new = int(np.round(self.win_length * factor))
+        hop_length_new = int(np.round(self.hop_length * speed))
+        keyshift_key = str(keyshift) + "_" + str(audio.device)
+        if keyshift_key not in self.hann_window:
+            self.hann_window[keyshift_key] = torch.hann_window(win_length_new).to(
+                audio.device
+            )
+        if "privateuseone" in str(audio.device):
+            if not hasattr(self, "stft"):
+                self.stft = STFT(
+                    filter_length=n_fft_new,
+                    hop_length=hop_length_new,
+                    win_length=win_length_new,
+                    window="hann",
+                ).to(audio.device)
+            magnitude = self.stft.transform(audio)
+        else:
+            fft = torch.stft(
+                audio,
+                n_fft=n_fft_new,
+                hop_length=hop_length_new,
+                win_length=win_length_new,
+                window=self.hann_window[keyshift_key],
+                center=center,
+                return_complex=True,
+            )
+            magnitude = torch.sqrt(fft.real.pow(2) + fft.imag.pow(2))
+        if keyshift != 0:
+            size = self.n_fft // 2 + 1
+            resize = magnitude.size(1)
+            if resize < size:
+                magnitude = F.pad(magnitude, (0, 0, 0, size - resize))
+            magnitude = magnitude[:, :size, :] * self.win_length / win_length_new
+        mel_output = torch.matmul(self.mel_basis, magnitude)
+        if self.is_half == True:
+            mel_output = mel_output.half()
+        log_mel_spec = torch.log(torch.clamp(mel_output, min=self.clamp))
+        return log_mel_spec
+
+
+class RMVPE:
+    def __init__(self, model_path: str, is_half, device=None, use_jit=False):
+        self.resample_kernel = {}
+        self.resample_kernel = {}
+        self.is_half = is_half
+        if device is None:
+            device = "cuda:0" if torch.cuda.is_available() else "cpu"
+        self.device = device
+        self.mel_extractor = MelSpectrogram(
+            is_half, 128, 16000, 1024, 160, None, 30, 8000
+        ).to(device)
+        if "privateuseone" in str(device):
+            import onnxruntime as ort
+
+            ort_session = ort.InferenceSession(
+                "%s/rmvpe.onnx" % os.environ["rmvpe_root"],
+                providers=["DmlExecutionProvider"],
+            )
+            self.model = ort_session
+        else:
+            if str(self.device) == "cuda":
+                self.device = torch.device("cuda:0")
+
+            def get_jit_model():
+                jit_model_path = model_path.rstrip(".pth")
+                jit_model_path += ".half.jit" if is_half else ".jit"
+                reload = False
+                if os.path.exists(jit_model_path):
+                    ckpt = jit.load(jit_model_path)
+                    model_device = ckpt["device"]
+                    if model_device != str(self.device):
+                        reload = True
+                else:
+                    reload = True
+
+                if reload:
+                    ckpt = jit.rmvpe_jit_export(
+                        model_path=model_path,
+                        mode="script",
+                        inputs_path=None,
+                        save_path=jit_model_path,
+                        device=device,
+                        is_half=is_half,
+                    )
+                model = torch.jit.load(BytesIO(ckpt["model"]), map_location=device)
+                return model
+
+            def get_default_model():
+                model = E2E(4, 1, (2, 2))
+                ckpt = torch.load(model_path, map_location="cpu")
+                model.load_state_dict(ckpt)
+                model.eval()
+                if is_half:
+                    model = model.half()
+                else:
+                    model = model.float()
+                return model
+
+            if use_jit:
+                if is_half and "cpu" in str(self.device):
+                    logger.warning(
+                        "Use default rmvpe model. \
+                                 Jit is not supported on the CPU for half floating point"
+                    )
+                    self.model = get_default_model()
+                else:
+                    self.model = get_jit_model()
+            else:
+                self.model = get_default_model()
+
+            self.model = self.model.to(device)
+        cents_mapping = 20 * np.arange(360) + 1997.3794084376191
+        self.cents_mapping = np.pad(cents_mapping, (4, 4))  # 368
+
+    def mel2hidden(self, mel):
+        with torch.no_grad():
+            n_frames = mel.shape[-1]
+            n_pad = 32 * ((n_frames - 1) // 32 + 1) - n_frames
+            if n_pad > 0:
+                mel = F.pad(mel, (0, n_pad), mode="constant")
+            if "privateuseone" in str(self.device):
+                onnx_input_name = self.model.get_inputs()[0].name
+                onnx_outputs_names = self.model.get_outputs()[0].name
+                hidden = self.model.run(
+                    [onnx_outputs_names],
+                    input_feed={onnx_input_name: mel.cpu().numpy()},
+                )[0]
+            else:
+                mel = mel.half() if self.is_half else mel.float()
+                hidden = self.model(mel)
+            return hidden[:, :n_frames]
+
+    def decode(self, hidden, thred=0.03):
+        cents_pred = self.to_local_average_cents(hidden, thred=thred)
+        f0 = 10 * (2 ** (cents_pred / 1200))
+        f0[f0 == 10] = 0
+        # f0 = np.array([10 * (2 ** (cent_pred / 1200)) if cent_pred else 0 for cent_pred in cents_pred])
+        return f0
+
+    def infer_from_audio(self, audio, thred=0.03):
+        # torch.cuda.synchronize()
+        # t0 = ttime()
+        if not torch.is_tensor(audio):
+            audio = torch.from_numpy(audio)
+        mel = self.mel_extractor(
+            audio.float().to(self.device).unsqueeze(0), center=True
+        )
+        # print(123123123,mel.device.type)
+        # torch.cuda.synchronize()
+        # t1 = ttime()
+        hidden = self.mel2hidden(mel)
+        # torch.cuda.synchronize()
+        # t2 = ttime()
+        # print(234234,hidden.device.type)
+        if "privateuseone" not in str(self.device):
+            hidden = hidden.squeeze(0).cpu().numpy()
+        else:
+            hidden = hidden[0]
+        if self.is_half == True:
+            hidden = hidden.astype("float32")
+
+        f0 = self.decode(hidden, thred=thred)
+        # torch.cuda.synchronize()
+        # t3 = ttime()
+        # print("hmvpe:%s\t%s\t%s\t%s"%(t1-t0,t2-t1,t3-t2,t3-t0))
+        return f0
+
+    def to_local_average_cents(self, salience, thred=0.05):
+        # t0 = ttime()
+        center = np.argmax(salience, axis=1)  # 帧长#index
+        salience = np.pad(salience, ((0, 0), (4, 4)))  # 帧长,368
+        # t1 = ttime()
+        center += 4
+        todo_salience = []
+        todo_cents_mapping = []
+        starts = center - 4
+        ends = center + 5
+        for idx in range(salience.shape[0]):
+            todo_salience.append(salience[:, starts[idx] : ends[idx]][idx])
+            todo_cents_mapping.append(self.cents_mapping[starts[idx] : ends[idx]])
+        # t2 = ttime()
+        todo_salience = np.array(todo_salience)  # 帧长，9
+        todo_cents_mapping = np.array(todo_cents_mapping)  # 帧长，9
+        product_sum = np.sum(todo_salience * todo_cents_mapping, 1)
+        weight_sum = np.sum(todo_salience, 1)  # 帧长
+        devided = product_sum / weight_sum  # 帧长
+        # t3 = ttime()
+        maxx = np.max(salience, axis=1)  # 帧长
+        devided[maxx <= thred] = 0
+        # t4 = ttime()
+        # print("decode:%s\t%s\t%s\t%s" % (t1 - t0, t2 - t1, t3 - t2, t4 - t3))
+        return devided
+
+
+if __name__ == "__main__":
+    import librosa
+    import soundfile as sf
+
+    audio, sampling_rate = sf.read(r"C:\Users\liujing04\Desktop\Z\冬之花clip1.wav")
+    if len(audio.shape) > 1:
+        audio = librosa.to_mono(audio.transpose(1, 0))
+    audio_bak = audio.copy()
+    if sampling_rate != 16000:
+        audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000)
+    model_path = r"D:\BaiduNetdiskDownload\RVC-beta-v2-0727AMD_realtime\rmvpe.pt"
+    thred = 0.03  # 0.01
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    rmvpe = RMVPE(model_path, is_half=False, device=device)
+    t0 = ttime()
+    f0 = rmvpe.infer_from_audio(audio, thred=thred)
+    # f0 = rmvpe.infer_from_audio(audio, thred=thred)
+    # f0 = rmvpe.infer_from_audio(audio, thred=thred)
+    # f0 = rmvpe.infer_from_audio(audio, thred=thred)
+    # f0 = rmvpe.infer_from_audio(audio, thred=thred)
+    t1 = ttime()
+    logger.info("%s %.2f", f0.shape, t1 - t0)

infer/lib/slicer2.py

@@ -0,0 +1,260 @@
+import numpy as np
+
+
+# This function is obtained from librosa.
+def get_rms(
+    y,
+    frame_length=2048,
+    hop_length=512,
+    pad_mode="constant",
+):
+    padding = (int(frame_length // 2), int(frame_length // 2))
+    y = np.pad(y, padding, mode=pad_mode)
+
+    axis = -1
+    # put our new within-frame axis at the end for now
+    out_strides = y.strides + tuple([y.strides[axis]])
+    # Reduce the shape on the framing axis
+    x_shape_trimmed = list(y.shape)
+    x_shape_trimmed[axis] -= frame_length - 1
+    out_shape = tuple(x_shape_trimmed) + tuple([frame_length])
+    xw = np.lib.stride_tricks.as_strided(y, shape=out_shape, strides=out_strides)
+    if axis < 0:
+        target_axis = axis - 1
+    else:
+        target_axis = axis + 1
+    xw = np.moveaxis(xw, -1, target_axis)
+    # Downsample along the target axis
+    slices = [slice(None)] * xw.ndim
+    slices[axis] = slice(0, None, hop_length)
+    x = xw[tuple(slices)]
+
+    # Calculate power
+    power = np.mean(np.abs(x) ** 2, axis=-2, keepdims=True)
+
+    return np.sqrt(power)
+
+
+class Slicer:
+    def __init__(
+        self,
+        sr: int,
+        threshold: float = -40.0,
+        min_length: int = 5000,
+        min_interval: int = 300,
+        hop_size: int = 20,
+        max_sil_kept: int = 5000,
+    ):
+        if not min_length >= min_interval >= hop_size:
+            raise ValueError(
+                "The following condition must be satisfied: min_length >= min_interval >= hop_size"
+            )
+        if not max_sil_kept >= hop_size:
+            raise ValueError(
+                "The following condition must be satisfied: max_sil_kept >= hop_size"
+            )
+        min_interval = sr * min_interval / 1000
+        self.threshold = 10 ** (threshold / 20.0)
+        self.hop_size = round(sr * hop_size / 1000)
+        self.win_size = min(round(min_interval), 4 * self.hop_size)
+        self.min_length = round(sr * min_length / 1000 / self.hop_size)
+        self.min_interval = round(min_interval / self.hop_size)
+        self.max_sil_kept = round(sr * max_sil_kept / 1000 / self.hop_size)
+
+    def _apply_slice(self, waveform, begin, end):
+        if len(waveform.shape) > 1:
+            return waveform[
+                :, begin * self.hop_size : min(waveform.shape[1], end * self.hop_size)
+            ]
+        else:
+            return waveform[
+                begin * self.hop_size : min(waveform.shape[0], end * self.hop_size)
+            ]
+
+    # @timeit
+    def slice(self, waveform):
+        if len(waveform.shape) > 1:
+            samples = waveform.mean(axis=0)
+        else:
+            samples = waveform
+        if samples.shape[0] <= self.min_length:
+            return [waveform]
+        rms_list = get_rms(
+            y=samples, frame_length=self.win_size, hop_length=self.hop_size
+        ).squeeze(0)
+        sil_tags = []
+        silence_start = None
+        clip_start = 0
+        for i, rms in enumerate(rms_list):
+            # Keep looping while frame is silent.
+            if rms < self.threshold:
+                # Record start of silent frames.
+                if silence_start is None:
+                    silence_start = i
+                continue
+            # Keep looping while frame is not silent and silence start has not been recorded.
+            if silence_start is None:
+                continue
+            # Clear recorded silence start if interval is not enough or clip is too short
+            is_leading_silence = silence_start == 0 and i > self.max_sil_kept
+            need_slice_middle = (
+                i - silence_start >= self.min_interval
+                and i - clip_start >= self.min_length
+            )
+            if not is_leading_silence and not need_slice_middle:
+                silence_start = None
+                continue
+            # Need slicing. Record the range of silent frames to be removed.
+            if i - silence_start <= self.max_sil_kept:
+                pos = rms_list[silence_start : i + 1].argmin() + silence_start
+                if silence_start == 0:
+                    sil_tags.append((0, pos))
+                else:
+                    sil_tags.append((pos, pos))
+                clip_start = pos
+            elif i - silence_start <= self.max_sil_kept * 2:
+                pos = rms_list[
+                    i - self.max_sil_kept : silence_start + self.max_sil_kept + 1
+                ].argmin()
+                pos += i - self.max_sil_kept
+                pos_l = (
+                    rms_list[
+                        silence_start : silence_start + self.max_sil_kept + 1
+                    ].argmin()
+                    + silence_start
+                )
+                pos_r = (
+                    rms_list[i - self.max_sil_kept : i + 1].argmin()
+                    + i
+                    - self.max_sil_kept
+                )
+                if silence_start == 0:
+                    sil_tags.append((0, pos_r))
+                    clip_start = pos_r
+                else:
+                    sil_tags.append((min(pos_l, pos), max(pos_r, pos)))
+                    clip_start = max(pos_r, pos)
+            else:
+                pos_l = (
+                    rms_list[
+                        silence_start : silence_start + self.max_sil_kept + 1
+                    ].argmin()
+                    + silence_start
+                )
+                pos_r = (
+                    rms_list[i - self.max_sil_kept : i + 1].argmin()
+                    + i
+                    - self.max_sil_kept
+                )
+                if silence_start == 0:
+                    sil_tags.append((0, pos_r))
+                else:
+                    sil_tags.append((pos_l, pos_r))
+                clip_start = pos_r
+            silence_start = None
+        # Deal with trailing silence.
+        total_frames = rms_list.shape[0]
+        if (
+            silence_start is not None
+            and total_frames - silence_start >= self.min_interval
+        ):
+            silence_end = min(total_frames, silence_start + self.max_sil_kept)
+            pos = rms_list[silence_start : silence_end + 1].argmin() + silence_start
+            sil_tags.append((pos, total_frames + 1))
+        # Apply and return slices.
+        if len(sil_tags) == 0:
+            return [waveform]
+        else:
+            chunks = []
+            if sil_tags[0][0] > 0:
+                chunks.append(self._apply_slice(waveform, 0, sil_tags[0][0]))
+            for i in range(len(sil_tags) - 1):
+                chunks.append(
+                    self._apply_slice(waveform, sil_tags[i][1], sil_tags[i + 1][0])
+                )
+            if sil_tags[-1][1] < total_frames:
+                chunks.append(
+                    self._apply_slice(waveform, sil_tags[-1][1], total_frames)
+                )
+            return chunks
+
+
+def main():
+    import os.path
+    from argparse import ArgumentParser
+
+    import librosa
+    import soundfile
+
+    parser = ArgumentParser()
+    parser.add_argument("audio", type=str, help="The audio to be sliced")
+    parser.add_argument(
+        "--out", type=str, help="Output directory of the sliced audio clips"
+    )
+    parser.add_argument(
+        "--db_thresh",
+        type=float,
+        required=False,
+        default=-40,
+        help="The dB threshold for silence detection",
+    )
+    parser.add_argument(
+        "--min_length",
+        type=int,
+        required=False,
+        default=5000,
+        help="The minimum milliseconds required for each sliced audio clip",
+    )
+    parser.add_argument(
+        "--min_interval",
+        type=int,
+        required=False,
+        default=300,
+        help="The minimum milliseconds for a silence part to be sliced",
+    )
+    parser.add_argument(
+        "--hop_size",
+        type=int,
+        required=False,
+        default=10,
+        help="Frame length in milliseconds",
+    )
+    parser.add_argument(
+        "--max_sil_kept",
+        type=int,
+        required=False,
+        default=500,
+        help="The maximum silence length kept around the sliced clip, presented in milliseconds",
+    )
+    args = parser.parse_args()
+    out = args.out
+    if out is None:
+        out = os.path.dirname(os.path.abspath(args.audio))
+    audio, sr = librosa.load(args.audio, sr=None, mono=False)
+    slicer = Slicer(
+        sr=sr,
+        threshold=args.db_thresh,
+        min_length=args.min_length,
+        min_interval=args.min_interval,
+        hop_size=args.hop_size,
+        max_sil_kept=args.max_sil_kept,
+    )
+    chunks = slicer.slice(audio)
+    if not os.path.exists(out):
+        os.makedirs(out)
+    for i, chunk in enumerate(chunks):
+        if len(chunk.shape) > 1:
+            chunk = chunk.T
+        soundfile.write(
+            os.path.join(
+                out,
+                f"%s_%d.wav"
+                % (os.path.basename(args.audio).rsplit(".", maxsplit=1)[0], i),
+            ),
+            chunk,
+            sr,
+        )
+
+
+if __name__ == "__main__":
+    main()

infer/lib/train/data_utils.py

@@ -0,0 +1,517 @@
+import os
+import traceback
+import logging
+
+logger = logging.getLogger(__name__)
+
+import numpy as np
+import torch
+import torch.utils.data
+
+from infer.lib.train.mel_processing import spectrogram_torch
+from infer.lib.train.utils import load_filepaths_and_text, load_wav_to_torch
+
+
+class TextAudioLoaderMultiNSFsid(torch.utils.data.Dataset):
+    """
+    1) loads audio, text pairs
+    2) normalizes text and converts them to sequences of integers
+    3) computes spectrograms from audio files.
+    """
+
+    def __init__(self, audiopaths_and_text, hparams):
+        self.audiopaths_and_text = load_filepaths_and_text(audiopaths_and_text)
+        self.max_wav_value = hparams.max_wav_value
+        self.sampling_rate = hparams.sampling_rate
+        self.filter_length = hparams.filter_length
+        self.hop_length = hparams.hop_length
+        self.win_length = hparams.win_length
+        self.sampling_rate = hparams.sampling_rate
+        self.min_text_len = getattr(hparams, "min_text_len", 1)
+        self.max_text_len = getattr(hparams, "max_text_len", 5000)
+        self._filter()
+
+    def _filter(self):
+        """
+        Filter text & store spec lengths
+        """
+        # Store spectrogram lengths for Bucketing
+        # wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
+        # spec_length = wav_length // hop_length
+        audiopaths_and_text_new = []
+        lengths = []
+        for audiopath, text, pitch, pitchf, dv in self.audiopaths_and_text:
+            if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
+                audiopaths_and_text_new.append([audiopath, text, pitch, pitchf, dv])
+                lengths.append(os.path.getsize(audiopath) // (3 * self.hop_length))
+        self.audiopaths_and_text = audiopaths_and_text_new
+        self.lengths = lengths
+
+    def get_sid(self, sid):
+        sid = torch.LongTensor([int(sid)])
+        return sid
+
+    def get_audio_text_pair(self, audiopath_and_text):
+        # separate filename and text
+        file = audiopath_and_text[0]
+        phone = audiopath_and_text[1]
+        pitch = audiopath_and_text[2]
+        pitchf = audiopath_and_text[3]
+        dv = audiopath_and_text[4]
+
+        phone, pitch, pitchf = self.get_labels(phone, pitch, pitchf)
+        spec, wav = self.get_audio(file)
+        dv = self.get_sid(dv)
+
+        len_phone = phone.size()[0]
+        len_spec = spec.size()[-1]
+        # print(123,phone.shape,pitch.shape,spec.shape)
+        if len_phone != len_spec:
+            len_min = min(len_phone, len_spec)
+            # amor
+            len_wav = len_min * self.hop_length
+
+            spec = spec[:, :len_min]
+            wav = wav[:, :len_wav]
+
+            phone = phone[:len_min, :]
+            pitch = pitch[:len_min]
+            pitchf = pitchf[:len_min]
+
+        return (spec, wav, phone, pitch, pitchf, dv)
+
+    def get_labels(self, phone, pitch, pitchf):
+        phone = np.load(phone)
+        phone = np.repeat(phone, 2, axis=0)
+        pitch = np.load(pitch)
+        pitchf = np.load(pitchf)
+        n_num = min(phone.shape[0], 900)  # DistributedBucketSampler
+        # print(234,phone.shape,pitch.shape)
+        phone = phone[:n_num, :]
+        pitch = pitch[:n_num]
+        pitchf = pitchf[:n_num]
+        phone = torch.FloatTensor(phone)
+        pitch = torch.LongTensor(pitch)
+        pitchf = torch.FloatTensor(pitchf)
+        return phone, pitch, pitchf
+
+    def get_audio(self, filename):
+        audio, sampling_rate = load_wav_to_torch(filename)
+        if sampling_rate != self.sampling_rate:
+            raise ValueError(
+                "{} SR doesn't match target {} SR".format(
+                    sampling_rate, self.sampling_rate
+                )
+            )
+        audio_norm = audio
+        #        audio_norm = audio / self.max_wav_value
+        #        audio_norm = audio / np.abs(audio).max()
+
+        audio_norm = audio_norm.unsqueeze(0)
+        spec_filename = filename.replace(".wav", ".spec.pt")
+        if os.path.exists(spec_filename):
+            try:
+                spec = torch.load(spec_filename)
+            except:
+                logger.warning("%s %s", spec_filename, traceback.format_exc())
+                spec = spectrogram_torch(
+                    audio_norm,
+                    self.filter_length,
+                    self.sampling_rate,
+                    self.hop_length,
+                    self.win_length,
+                    center=False,
+                )
+                spec = torch.squeeze(spec, 0)
+                torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
+        else:
+            spec = spectrogram_torch(
+                audio_norm,
+                self.filter_length,
+                self.sampling_rate,
+                self.hop_length,
+                self.win_length,
+                center=False,
+            )
+            spec = torch.squeeze(spec, 0)
+            torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
+        return spec, audio_norm
+
+    def __getitem__(self, index):
+        return self.get_audio_text_pair(self.audiopaths_and_text[index])
+
+    def __len__(self):
+        return len(self.audiopaths_and_text)
+
+
+class TextAudioCollateMultiNSFsid:
+    """Zero-pads model inputs and targets"""
+
+    def __init__(self, return_ids=False):
+        self.return_ids = return_ids
+
+    def __call__(self, batch):
+        """Collate's training batch from normalized text and aduio
+        PARAMS
+        ------
+        batch: [text_normalized, spec_normalized, wav_normalized]
+        """
+        # Right zero-pad all one-hot text sequences to max input length
+        _, ids_sorted_decreasing = torch.sort(
+            torch.LongTensor([x[0].size(1) for x in batch]), dim=0, descending=True
+        )
+
+        max_spec_len = max([x[0].size(1) for x in batch])
+        max_wave_len = max([x[1].size(1) for x in batch])
+        spec_lengths = torch.LongTensor(len(batch))
+        wave_lengths = torch.LongTensor(len(batch))
+        spec_padded = torch.FloatTensor(len(batch), batch[0][0].size(0), max_spec_len)
+        wave_padded = torch.FloatTensor(len(batch), 1, max_wave_len)
+        spec_padded.zero_()
+        wave_padded.zero_()
+
+        max_phone_len = max([x[2].size(0) for x in batch])
+        phone_lengths = torch.LongTensor(len(batch))
+        phone_padded = torch.FloatTensor(
+            len(batch), max_phone_len, batch[0][2].shape[1]
+        )  # (spec, wav, phone, pitch)
+        pitch_padded = torch.LongTensor(len(batch), max_phone_len)
+        pitchf_padded = torch.FloatTensor(len(batch), max_phone_len)
+        phone_padded.zero_()
+        pitch_padded.zero_()
+        pitchf_padded.zero_()
+        # dv = torch.FloatTensor(len(batch), 256)#gin=256
+        sid = torch.LongTensor(len(batch))
+
+        for i in range(len(ids_sorted_decreasing)):
+            row = batch[ids_sorted_decreasing[i]]
+
+            spec = row[0]
+            spec_padded[i, :, : spec.size(1)] = spec
+            spec_lengths[i] = spec.size(1)
+
+            wave = row[1]
+            wave_padded[i, :, : wave.size(1)] = wave
+            wave_lengths[i] = wave.size(1)
+
+            phone = row[2]
+            phone_padded[i, : phone.size(0), :] = phone
+            phone_lengths[i] = phone.size(0)
+
+            pitch = row[3]
+            pitch_padded[i, : pitch.size(0)] = pitch
+            pitchf = row[4]
+            pitchf_padded[i, : pitchf.size(0)] = pitchf
+
+            # dv[i] = row[5]
+            sid[i] = row[5]
+
+        return (
+            phone_padded,
+            phone_lengths,
+            pitch_padded,
+            pitchf_padded,
+            spec_padded,
+            spec_lengths,
+            wave_padded,
+            wave_lengths,
+            # dv
+            sid,
+        )
+
+
+class TextAudioLoader(torch.utils.data.Dataset):
+    """
+    1) loads audio, text pairs
+    2) normalizes text and converts them to sequences of integers
+    3) computes spectrograms from audio files.
+    """
+
+    def __init__(self, audiopaths_and_text, hparams):
+        self.audiopaths_and_text = load_filepaths_and_text(audiopaths_and_text)
+        self.max_wav_value = hparams.max_wav_value
+        self.sampling_rate = hparams.sampling_rate
+        self.filter_length = hparams.filter_length
+        self.hop_length = hparams.hop_length
+        self.win_length = hparams.win_length
+        self.sampling_rate = hparams.sampling_rate
+        self.min_text_len = getattr(hparams, "min_text_len", 1)
+        self.max_text_len = getattr(hparams, "max_text_len", 5000)
+        self._filter()
+
+    def _filter(self):
+        """
+        Filter text & store spec lengths
+        """
+        # Store spectrogram lengths for Bucketing
+        # wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
+        # spec_length = wav_length // hop_length
+        audiopaths_and_text_new = []
+        lengths = []
+        for audiopath, text, dv in self.audiopaths_and_text:
+            if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
+                audiopaths_and_text_new.append([audiopath, text, dv])
+                lengths.append(os.path.getsize(audiopath) // (3 * self.hop_length))
+        self.audiopaths_and_text = audiopaths_and_text_new
+        self.lengths = lengths
+
+    def get_sid(self, sid):
+        sid = torch.LongTensor([int(sid)])
+        return sid
+
+    def get_audio_text_pair(self, audiopath_and_text):
+        # separate filename and text
+        file = audiopath_and_text[0]
+        phone = audiopath_and_text[1]
+        dv = audiopath_and_text[2]
+
+        phone = self.get_labels(phone)
+        spec, wav = self.get_audio(file)
+        dv = self.get_sid(dv)
+
+        len_phone = phone.size()[0]
+        len_spec = spec.size()[-1]
+        if len_phone != len_spec:
+            len_min = min(len_phone, len_spec)
+            len_wav = len_min * self.hop_length
+            spec = spec[:, :len_min]
+            wav = wav[:, :len_wav]
+            phone = phone[:len_min, :]
+        return (spec, wav, phone, dv)
+
+    def get_labels(self, phone):
+        phone = np.load(phone)
+        phone = np.repeat(phone, 2, axis=0)
+        n_num = min(phone.shape[0], 900)  # DistributedBucketSampler
+        phone = phone[:n_num, :]
+        phone = torch.FloatTensor(phone)
+        return phone
+
+    def get_audio(self, filename):
+        audio, sampling_rate = load_wav_to_torch(filename)
+        if sampling_rate != self.sampling_rate:
+            raise ValueError(
+                "{} SR doesn't match target {} SR".format(
+                    sampling_rate, self.sampling_rate
+                )
+            )
+        audio_norm = audio
+        #        audio_norm = audio / self.max_wav_value
+        #        audio_norm = audio / np.abs(audio).max()
+
+        audio_norm = audio_norm.unsqueeze(0)
+        spec_filename = filename.replace(".wav", ".spec.pt")
+        if os.path.exists(spec_filename):
+            try:
+                spec = torch.load(spec_filename)
+            except:
+                logger.warning("%s %s", spec_filename, traceback.format_exc())
+                spec = spectrogram_torch(
+                    audio_norm,
+                    self.filter_length,
+                    self.sampling_rate,
+                    self.hop_length,
+                    self.win_length,
+                    center=False,
+                )
+                spec = torch.squeeze(spec, 0)
+                torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
+        else:
+            spec = spectrogram_torch(
+                audio_norm,
+                self.filter_length,
+                self.sampling_rate,
+                self.hop_length,
+                self.win_length,
+                center=False,
+            )
+            spec = torch.squeeze(spec, 0)
+            torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
+        return spec, audio_norm
+
+    def __getitem__(self, index):
+        return self.get_audio_text_pair(self.audiopaths_and_text[index])
+
+    def __len__(self):
+        return len(self.audiopaths_and_text)
+
+
+class TextAudioCollate:
+    """Zero-pads model inputs and targets"""
+
+    def __init__(self, return_ids=False):
+        self.return_ids = return_ids
+
+    def __call__(self, batch):
+        """Collate's training batch from normalized text and aduio
+        PARAMS
+        ------
+        batch: [text_normalized, spec_normalized, wav_normalized]
+        """
+        # Right zero-pad all one-hot text sequences to max input length
+        _, ids_sorted_decreasing = torch.sort(
+            torch.LongTensor([x[0].size(1) for x in batch]), dim=0, descending=True
+        )
+
+        max_spec_len = max([x[0].size(1) for x in batch])
+        max_wave_len = max([x[1].size(1) for x in batch])
+        spec_lengths = torch.LongTensor(len(batch))
+        wave_lengths = torch.LongTensor(len(batch))
+        spec_padded = torch.FloatTensor(len(batch), batch[0][0].size(0), max_spec_len)
+        wave_padded = torch.FloatTensor(len(batch), 1, max_wave_len)
+        spec_padded.zero_()
+        wave_padded.zero_()
+
+        max_phone_len = max([x[2].size(0) for x in batch])
+        phone_lengths = torch.LongTensor(len(batch))
+        phone_padded = torch.FloatTensor(
+            len(batch), max_phone_len, batch[0][2].shape[1]
+        )
+        phone_padded.zero_()
+        sid = torch.LongTensor(len(batch))
+
+        for i in range(len(ids_sorted_decreasing)):
+            row = batch[ids_sorted_decreasing[i]]
+
+            spec = row[0]
+            spec_padded[i, :, : spec.size(1)] = spec
+            spec_lengths[i] = spec.size(1)
+
+            wave = row[1]
+            wave_padded[i, :, : wave.size(1)] = wave
+            wave_lengths[i] = wave.size(1)
+
+            phone = row[2]
+            phone_padded[i, : phone.size(0), :] = phone
+            phone_lengths[i] = phone.size(0)
+
+            sid[i] = row[3]
+
+        return (
+            phone_padded,
+            phone_lengths,
+            spec_padded,
+            spec_lengths,
+            wave_padded,
+            wave_lengths,
+            sid,
+        )
+
+
+class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
+    """
+    Maintain similar input lengths in a batch.
+    Length groups are specified by boundaries.
+    Ex) boundaries = [b1, b2, b3] -> any batch is included either {x | b1 < length(x) <=b2} or {x | b2 < length(x) <= b3}.
+
+    It removes samples which are not included in the boundaries.
+    Ex) boundaries = [b1, b2, b3] -> any x s.t. length(x) <= b1 or length(x) > b3 are discarded.
+    """
+
+    def __init__(
+        self,
+        dataset,
+        batch_size,
+        boundaries,
+        num_replicas=None,
+        rank=None,
+        shuffle=True,
+    ):
+        super().__init__(dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle)
+        self.lengths = dataset.lengths
+        self.batch_size = batch_size
+        self.boundaries = boundaries
+
+        self.buckets, self.num_samples_per_bucket = self._create_buckets()
+        self.total_size = sum(self.num_samples_per_bucket)
+        self.num_samples = self.total_size // self.num_replicas
+
+    def _create_buckets(self):
+        buckets = [[] for _ in range(len(self.boundaries) - 1)]
+        for i in range(len(self.lengths)):
+            length = self.lengths[i]
+            idx_bucket = self._bisect(length)
+            if idx_bucket != -1:
+                buckets[idx_bucket].append(i)
+
+        for i in range(len(buckets) - 1, -1, -1):  #
+            if len(buckets[i]) == 0:
+                buckets.pop(i)
+                self.boundaries.pop(i + 1)
+
+        num_samples_per_bucket = []
+        for i in range(len(buckets)):
+            len_bucket = len(buckets[i])
+            total_batch_size = self.num_replicas * self.batch_size
+            rem = (
+                total_batch_size - (len_bucket % total_batch_size)
+            ) % total_batch_size
+            num_samples_per_bucket.append(len_bucket + rem)
+        return buckets, num_samples_per_bucket
+
+    def __iter__(self):
+        # deterministically shuffle based on epoch
+        g = torch.Generator()
+        g.manual_seed(self.epoch)
+
+        indices = []
+        if self.shuffle:
+            for bucket in self.buckets:
+                indices.append(torch.randperm(len(bucket), generator=g).tolist())
+        else:
+            for bucket in self.buckets:
+                indices.append(list(range(len(bucket))))
+
+        batches = []
+        for i in range(len(self.buckets)):
+            bucket = self.buckets[i]
+            len_bucket = len(bucket)
+            ids_bucket = indices[i]
+            num_samples_bucket = self.num_samples_per_bucket[i]
+
+            # add extra samples to make it evenly divisible
+            rem = num_samples_bucket - len_bucket
+            ids_bucket = (
+                ids_bucket
+                + ids_bucket * (rem // len_bucket)
+                + ids_bucket[: (rem % len_bucket)]
+            )
+
+            # subsample
+            ids_bucket = ids_bucket[self.rank :: self.num_replicas]
+
+            # batching
+            for j in range(len(ids_bucket) // self.batch_size):
+                batch = [
+                    bucket[idx]
+                    for idx in ids_bucket[
+                        j * self.batch_size : (j + 1) * self.batch_size
+                    ]
+                ]
+                batches.append(batch)
+
+        if self.shuffle:
+            batch_ids = torch.randperm(len(batches), generator=g).tolist()
+            batches = [batches[i] for i in batch_ids]
+        self.batches = batches
+
+        assert len(self.batches) * self.batch_size == self.num_samples
+        return iter(self.batches)
+
+    def _bisect(self, x, lo=0, hi=None):
+        if hi is None:
+            hi = len(self.boundaries) - 1
+
+        if hi > lo:
+            mid = (hi + lo) // 2
+            if self.boundaries[mid] < x and x <= self.boundaries[mid + 1]:
+                return mid
+            elif x <= self.boundaries[mid]:
+                return self._bisect(x, lo, mid)
+            else:
+                return self._bisect(x, mid + 1, hi)
+        else:
+            return -1
+
+    def __len__(self):
+        return self.num_samples // self.batch_size

infer/lib/train/losses.py

@@ -0,0 +1,58 @@
+import torch
+
+
+def feature_loss(fmap_r, fmap_g):
+    loss = 0
+    for dr, dg in zip(fmap_r, fmap_g):
+        for rl, gl in zip(dr, dg):
+            rl = rl.float().detach()
+            gl = gl.float()
+            loss += torch.mean(torch.abs(rl - gl))
+
+    return loss * 2
+
+
+def discriminator_loss(disc_real_outputs, disc_generated_outputs):
+    loss = 0
+    r_losses = []
+    g_losses = []
+    for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
+        dr = dr.float()
+        dg = dg.float()
+        r_loss = torch.mean((1 - dr) ** 2)
+        g_loss = torch.mean(dg**2)
+        loss += r_loss + g_loss
+        r_losses.append(r_loss.item())
+        g_losses.append(g_loss.item())
+
+    return loss, r_losses, g_losses
+
+
+def generator_loss(disc_outputs):
+    loss = 0
+    gen_losses = []
+    for dg in disc_outputs:
+        dg = dg.float()
+        l = torch.mean((1 - dg) ** 2)
+        gen_losses.append(l)
+        loss += l
+
+    return loss, gen_losses
+
+
+def kl_loss(z_p, logs_q, m_p, logs_p, z_mask):
+    """
+    z_p, logs_q: [b, h, t_t]
+    m_p, logs_p: [b, h, t_t]
+    """
+    z_p = z_p.float()
+    logs_q = logs_q.float()
+    m_p = m_p.float()
+    logs_p = logs_p.float()
+    z_mask = z_mask.float()
+
+    kl = logs_p - logs_q - 0.5
+    kl += 0.5 * ((z_p - m_p) ** 2) * torch.exp(-2.0 * logs_p)
+    kl = torch.sum(kl * z_mask)
+    l = kl / torch.sum(z_mask)
+    return l

infer/lib/train/mel_processing.py

@@ -0,0 +1,127 @@
+import torch
+import torch.utils.data
+from librosa.filters import mel as librosa_mel_fn
+import logging
+
+logger = logging.getLogger(__name__)
+
+MAX_WAV_VALUE = 32768.0
+
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+    """
+    PARAMS
+    ------
+    C: compression factor
+    """
+    return torch.log(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression_torch(x, C=1):
+    """
+    PARAMS
+    ------
+    C: compression factor used to compress
+    """
+    return torch.exp(x) / C
+
+
+def spectral_normalize_torch(magnitudes):
+    return dynamic_range_compression_torch(magnitudes)
+
+
+def spectral_de_normalize_torch(magnitudes):
+    return dynamic_range_decompression_torch(magnitudes)
+
+
+# Reusable banks
+mel_basis = {}
+hann_window = {}
+
+
+def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):
+    """Convert waveform into Linear-frequency Linear-amplitude spectrogram.
+
+    Args:
+        y             :: (B, T) - Audio waveforms
+        n_fft
+        sampling_rate
+        hop_size
+        win_size
+        center
+    Returns:
+        :: (B, Freq, Frame) - Linear-frequency Linear-amplitude spectrogram
+    """
+
+    # Window - Cache if needed
+    global hann_window
+    dtype_device = str(y.dtype) + "_" + str(y.device)
+    wnsize_dtype_device = str(win_size) + "_" + dtype_device
+    if wnsize_dtype_device not in hann_window:
+        hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(
+            dtype=y.dtype, device=y.device
+        )
+
+    # Padding
+    y = torch.nn.functional.pad(
+        y.unsqueeze(1),
+        (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)),
+        mode="reflect",
+    )
+    y = y.squeeze(1)
+
+    # Complex Spectrogram :: (B, T) -> (B, Freq, Frame, RealComplex=2)
+    spec = torch.stft(
+        y,
+        n_fft,
+        hop_length=hop_size,
+        win_length=win_size,
+        window=hann_window[wnsize_dtype_device],
+        center=center,
+        pad_mode="reflect",
+        normalized=False,
+        onesided=True,
+        return_complex=True,
+    )
+
+    # Linear-frequency Linear-amplitude spectrogram :: (B, Freq, Frame, RealComplex=2) -> (B, Freq, Frame)
+    spec = torch.sqrt(spec.real.pow(2) + spec.imag.pow(2) + 1e-6)
+    return spec
+
+
+def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
+    # MelBasis - Cache if needed
+    global mel_basis
+    dtype_device = str(spec.dtype) + "_" + str(spec.device)
+    fmax_dtype_device = str(fmax) + "_" + dtype_device
+    if fmax_dtype_device not in mel_basis:
+        mel = librosa_mel_fn(
+            sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax
+        )
+        mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(
+            dtype=spec.dtype, device=spec.device
+        )
+
+    # Mel-frequency Log-amplitude spectrogram :: (B, Freq=num_mels, Frame)
+    melspec = torch.matmul(mel_basis[fmax_dtype_device], spec)
+    melspec = spectral_normalize_torch(melspec)
+    return melspec
+
+
+def mel_spectrogram_torch(
+    y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False
+):
+    """Convert waveform into Mel-frequency Log-amplitude spectrogram.
+
+    Args:
+        y       :: (B, T)           - Waveforms
+    Returns:
+        melspec :: (B, Freq, Frame) - Mel-frequency Log-amplitude spectrogram
+    """
+    # Linear-frequency Linear-amplitude spectrogram :: (B, T) -> (B, Freq, Frame)
+    spec = spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center)
+
+    # Mel-frequency Log-amplitude spectrogram :: (B, Freq, Frame) -> (B, Freq=num_mels, Frame)
+    melspec = spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax)
+
+    return melspec

infer/lib/train/process_ckpt.py

@@ -0,0 +1,261 @@
+import os
+import sys
+import traceback
+from collections import OrderedDict
+
+import torch
+
+from i18n.i18n import I18nAuto
+
+i18n = I18nAuto()
+
+
+def savee(ckpt, sr, if_f0, name, epoch, version, hps):
+    try:
+        opt = OrderedDict()
+        opt["weight"] = {}
+        for key in ckpt.keys():
+            if "enc_q" in key:
+                continue
+            opt["weight"][key] = ckpt[key].half()
+        opt["config"] = [
+            hps.data.filter_length // 2 + 1,
+            32,
+            hps.model.inter_channels,
+            hps.model.hidden_channels,
+            hps.model.filter_channels,
+            hps.model.n_heads,
+            hps.model.n_layers,
+            hps.model.kernel_size,
+            hps.model.p_dropout,
+            hps.model.resblock,
+            hps.model.resblock_kernel_sizes,
+            hps.model.resblock_dilation_sizes,
+            hps.model.upsample_rates,
+            hps.model.upsample_initial_channel,
+            hps.model.upsample_kernel_sizes,
+            hps.model.spk_embed_dim,
+            hps.model.gin_channels,
+            hps.data.sampling_rate,
+        ]
+        opt["info"] = "%sepoch" % epoch
+        opt["sr"] = sr
+        opt["f0"] = if_f0
+        opt["version"] = version
+        torch.save(opt, "assets/weights/%s.pth" % name)
+        return "Success."
+    except:
+        return traceback.format_exc()
+
+
+def show_info(path):
+    try:
+        a = torch.load(path, map_location="cpu")
+        return "模型信息:%s\n采样率:%s\n模型是否输入音高引导:%s\n版本:%s" % (
+            a.get("info", "None"),
+            a.get("sr", "None"),
+            a.get("f0", "None"),
+            a.get("version", "None"),
+        )
+    except:
+        return traceback.format_exc()
+
+
+def extract_small_model(path, name, sr, if_f0, info, version):
+    try:
+        ckpt = torch.load(path, map_location="cpu")
+        if "model" in ckpt:
+            ckpt = ckpt["model"]
+        opt = OrderedDict()
+        opt["weight"] = {}
+        for key in ckpt.keys():
+            if "enc_q" in key:
+                continue
+            opt["weight"][key] = ckpt[key].half()
+        if sr == "40k":
+            opt["config"] = [
+                1025,
+                32,
+                192,
+                192,
+                768,
+                2,
+                6,
+                3,
+                0,
+                "1",
+                [3, 7, 11],
+                [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+                [10, 10, 2, 2],
+                512,
+                [16, 16, 4, 4],
+                109,
+                256,
+                40000,
+            ]
+        elif sr == "48k":
+            if version == "v1":
+                opt["config"] = [
+                    1025,
+                    32,
+                    192,
+                    192,
+                    768,
+                    2,
+                    6,
+                    3,
+                    0,
+                    "1",
+                    [3, 7, 11],
+                    [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+                    [10, 6, 2, 2, 2],
+                    512,
+                    [16, 16, 4, 4, 4],
+                    109,
+                    256,
+                    48000,
+                ]
+            else:
+                opt["config"] = [
+                    1025,
+                    32,
+                    192,
+                    192,
+                    768,
+                    2,
+                    6,
+                    3,
+                    0,
+                    "1",
+                    [3, 7, 11],
+                    [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+                    [12, 10, 2, 2],
+                    512,
+                    [24, 20, 4, 4],
+                    109,
+                    256,
+                    48000,
+                ]
+        elif sr == "32k":
+            if version == "v1":
+                opt["config"] = [
+                    513,
+                    32,
+                    192,
+                    192,
+                    768,
+                    2,
+                    6,
+                    3,
+                    0,
+                    "1",
+                    [3, 7, 11],
+                    [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+                    [10, 4, 2, 2, 2],
+                    512,
+                    [16, 16, 4, 4, 4],
+                    109,
+                    256,
+                    32000,
+                ]
+            else:
+                opt["config"] = [
+                    513,
+                    32,
+                    192,
+                    192,
+                    768,
+                    2,
+                    6,
+                    3,
+                    0,
+                    "1",
+                    [3, 7, 11],
+                    [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+                    [10, 8, 2, 2],
+                    512,
+                    [20, 16, 4, 4],
+                    109,
+                    256,
+                    32000,
+                ]
+        if info == "":
+            info = "Extracted model."
+        opt["info"] = info
+        opt["version"] = version
+        opt["sr"] = sr
+        opt["f0"] = int(if_f0)
+        torch.save(opt, "assets/weights/%s.pth" % name)
+        return "Success."
+    except:
+        return traceback.format_exc()
+
+
+def change_info(path, info, name):
+    try:
+        ckpt = torch.load(path, map_location="cpu")
+        ckpt["info"] = info
+        if name == "":
+            name = os.path.basename(path)
+        torch.save(ckpt, "assets/weights/%s" % name)
+        return "Success."
+    except:
+        return traceback.format_exc()
+
+
+def merge(path1, path2, alpha1, sr, f0, info, name, version):
+    try:
+
+        def extract(ckpt):
+            a = ckpt["model"]
+            opt = OrderedDict()
+            opt["weight"] = {}
+            for key in a.keys():
+                if "enc_q" in key:
+                    continue
+                opt["weight"][key] = a[key]
+            return opt
+
+        ckpt1 = torch.load(path1, map_location="cpu")
+        ckpt2 = torch.load(path2, map_location="cpu")
+        cfg = ckpt1["config"]
+        if "model" in ckpt1:
+            ckpt1 = extract(ckpt1)
+        else:
+            ckpt1 = ckpt1["weight"]
+        if "model" in ckpt2:
+            ckpt2 = extract(ckpt2)
+        else:
+            ckpt2 = ckpt2["weight"]
+        if sorted(list(ckpt1.keys())) != sorted(list(ckpt2.keys())):
+            return "Fail to merge the models. The model architectures are not the same."
+        opt = OrderedDict()
+        opt["weight"] = {}
+        for key in ckpt1.keys():
+            # try:
+            if key == "emb_g.weight" and ckpt1[key].shape != ckpt2[key].shape:
+                min_shape0 = min(ckpt1[key].shape[0], ckpt2[key].shape[0])
+                opt["weight"][key] = (
+                    alpha1 * (ckpt1[key][:min_shape0].float())
+                    + (1 - alpha1) * (ckpt2[key][:min_shape0].float())
+                ).half()
+            else:
+                opt["weight"][key] = (
+                    alpha1 * (ckpt1[key].float()) + (1 - alpha1) * (ckpt2[key].float())
+                ).half()
+        # except:
+        #     pdb.set_trace()
+        opt["config"] = cfg
+        """
+        if(sr=="40k"):opt["config"] = [1025, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10, 10, 2, 2], 512, [16, 16, 4, 4,4], 109, 256, 40000]
+        elif(sr=="48k"):opt["config"] = [1025, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10,6,2,2,2], 512, [16, 16, 4, 4], 109, 256, 48000]
+        elif(sr=="32k"):opt["config"] = [513, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10, 4, 2, 2, 2], 512, [16, 16, 4, 4,4], 109, 256, 32000]
+        """
+        opt["sr"] = sr
+        opt["f0"] = 1 if f0 == i18n("是") else 0
+        opt["version"] = version
+        opt["info"] = info
+        torch.save(opt, "assets/weights/%s.pth" % name)
+        return "Success."
+    except:
+        return traceback.format_exc()

infer/lib/train/utils.py

@@ -0,0 +1,478 @@
+import argparse
+import glob
+import json
+import logging
+import os
+import subprocess
+import sys
+import shutil
+
+import numpy as np
+import torch
+from scipy.io.wavfile import read
+
+MATPLOTLIB_FLAG = False
+
+logging.basicConfig(stream=sys.stdout, level=logging.DEBUG)
+logger = logging
+
+
+def load_checkpoint_d(checkpoint_path, combd, sbd, optimizer=None, load_opt=1):
+    assert os.path.isfile(checkpoint_path)
+    checkpoint_dict = torch.load(checkpoint_path, map_location="cpu")
+
+    ##################
+    def go(model, bkey):
+        saved_state_dict = checkpoint_dict[bkey]
+        if hasattr(model, "module"):
+            state_dict = model.module.state_dict()
+        else:
+            state_dict = model.state_dict()
+        new_state_dict = {}
+        for k, v in state_dict.items():  # 模型需要的shape
+            try:
+                new_state_dict[k] = saved_state_dict[k]
+                if saved_state_dict[k].shape != state_dict[k].shape:
+                    logger.warning(
+                        "shape-%s-mismatch. need: %s, get: %s",
+                        k,
+                        state_dict[k].shape,
+                        saved_state_dict[k].shape,
+                    )  #
+                    raise KeyError
+            except:
+                # logger.info(traceback.format_exc())
+                logger.info("%s is not in the checkpoint", k)  # pretrain缺失的
+                new_state_dict[k] = v  # 模型自带的随机值
+        if hasattr(model, "module"):
+            model.module.load_state_dict(new_state_dict, strict=False)
+        else:
+            model.load_state_dict(new_state_dict, strict=False)
+        return model
+
+    go(combd, "combd")
+    model = go(sbd, "sbd")
+    #############
+    logger.info("Loaded model weights")
+
+    iteration = checkpoint_dict["iteration"]
+    learning_rate = checkpoint_dict["learning_rate"]
+    if (
+        optimizer is not None and load_opt == 1
+    ):  ###加载不了，如果是空的的话，重新初始化，可能还会影响lr时间表的更新，因此在train文件最外围catch
+        #   try:
+        optimizer.load_state_dict(checkpoint_dict["optimizer"])
+    #   except:
+    #     traceback.print_exc()
+    logger.info("Loaded checkpoint '{}' (epoch {})".format(checkpoint_path, iteration))
+    return model, optimizer, learning_rate, iteration
+
+
+# def load_checkpoint(checkpoint_path, model, optimizer=None):
+#   assert os.path.isfile(checkpoint_path)
+#   checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
+#   iteration = checkpoint_dict['iteration']
+#   learning_rate = checkpoint_dict['learning_rate']
+#   if optimizer is not None:
+#     optimizer.load_state_dict(checkpoint_dict['optimizer'])
+#   # print(1111)
+#   saved_state_dict = checkpoint_dict['model']
+#   # print(1111)
+#
+#   if hasattr(model, 'module'):
+#     state_dict = model.module.state_dict()
+#   else:
+#     state_dict = model.state_dict()
+#   new_state_dict= {}
+#   for k, v in state_dict.items():
+#     try:
+#       new_state_dict[k] = saved_state_dict[k]
+#     except:
+#       logger.info("%s is not in the checkpoint" % k)
+#       new_state_dict[k] = v
+#   if hasattr(model, 'module'):
+#     model.module.load_state_dict(new_state_dict)
+#   else:
+#     model.load_state_dict(new_state_dict)
+#   logger.info("Loaded checkpoint '{}' (epoch {})" .format(
+#     checkpoint_path, iteration))
+#   return model, optimizer, learning_rate, iteration
+def load_checkpoint(checkpoint_path, model, optimizer=None, load_opt=1):
+    assert os.path.isfile(checkpoint_path)
+    checkpoint_dict = torch.load(checkpoint_path, map_location="cpu")
+
+    saved_state_dict = checkpoint_dict["model"]
+    if hasattr(model, "module"):
+        state_dict = model.module.state_dict()
+    else:
+        state_dict = model.state_dict()
+    new_state_dict = {}
+    for k, v in state_dict.items():  # 模型需要的shape
+        try:
+            new_state_dict[k] = saved_state_dict[k]
+            if saved_state_dict[k].shape != state_dict[k].shape:
+                logger.warning(
+                    "shape-%s-mismatch|need-%s|get-%s",
+                    k,
+                    state_dict[k].shape,
+                    saved_state_dict[k].shape,
+                )  #
+                raise KeyError
+        except:
+            # logger.info(traceback.format_exc())
+            logger.info("%s is not in the checkpoint", k)  # pretrain缺失的
+            new_state_dict[k] = v  # 模型自带的随机值
+    if hasattr(model, "module"):
+        model.module.load_state_dict(new_state_dict, strict=False)
+    else:
+        model.load_state_dict(new_state_dict, strict=False)
+    logger.info("Loaded model weights")
+
+    iteration = checkpoint_dict["iteration"]
+    learning_rate = checkpoint_dict["learning_rate"]
+    if (
+        optimizer is not None and load_opt == 1
+    ):  ###加载不了，如果是空的的话，重新初始化，可能还会影响lr时间表的更新，因此在train文件最外围catch
+        #   try:
+        optimizer.load_state_dict(checkpoint_dict["optimizer"])
+    #   except:
+    #     traceback.print_exc()
+    logger.info("Loaded checkpoint '{}' (epoch {})".format(checkpoint_path, iteration))
+    return model, optimizer, learning_rate, iteration
+
+
+def save_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path):
+    logger.info(
+        "Saving model and optimizer state at epoch {} to {}".format(
+            iteration, checkpoint_path
+        )
+    )
+    if hasattr(model, "module"):
+        state_dict = model.module.state_dict()
+    else:
+        state_dict = model.state_dict()
+    torch.save(
+        {
+            "model": state_dict,
+            "iteration": iteration,
+            "optimizer": optimizer.state_dict(),
+            "learning_rate": learning_rate,
+        },
+        checkpoint_path,
+    )
+
+
+def save_checkpoint_d(combd, sbd, optimizer, learning_rate, iteration, checkpoint_path):
+    logger.info(
+        "Saving model and optimizer state at epoch {} to {}".format(
+            iteration, checkpoint_path
+        )
+    )
+    if hasattr(combd, "module"):
+        state_dict_combd = combd.module.state_dict()
+    else:
+        state_dict_combd = combd.state_dict()
+    if hasattr(sbd, "module"):
+        state_dict_sbd = sbd.module.state_dict()
+    else:
+        state_dict_sbd = sbd.state_dict()
+    torch.save(
+        {
+            "combd": state_dict_combd,
+            "sbd": state_dict_sbd,
+            "iteration": iteration,
+            "optimizer": optimizer.state_dict(),
+            "learning_rate": learning_rate,
+        },
+        checkpoint_path,
+    )
+
+
+def summarize(
+    writer,
+    global_step,
+    scalars={},
+    histograms={},
+    images={},
+    audios={},
+    audio_sampling_rate=22050,
+):
+    for k, v in scalars.items():
+        writer.add_scalar(k, v, global_step)
+    for k, v in histograms.items():
+        writer.add_histogram(k, v, global_step)
+    for k, v in images.items():
+        writer.add_image(k, v, global_step, dataformats="HWC")
+    for k, v in audios.items():
+        writer.add_audio(k, v, global_step, audio_sampling_rate)
+
+
+def latest_checkpoint_path(dir_path, regex="G_*.pth"):
+    f_list = glob.glob(os.path.join(dir_path, regex))
+    f_list.sort(key=lambda f: int("".join(filter(str.isdigit, f))))
+    x = f_list[-1]
+    logger.debug(x)
+    return x
+
+
+def plot_spectrogram_to_numpy(spectrogram):
+    global MATPLOTLIB_FLAG
+    if not MATPLOTLIB_FLAG:
+        import matplotlib
+
+        matplotlib.use("Agg")
+        MATPLOTLIB_FLAG = True
+        mpl_logger = logging.getLogger("matplotlib")
+        mpl_logger.setLevel(logging.WARNING)
+    import matplotlib.pylab as plt
+    import numpy as np
+
+    fig, ax = plt.subplots(figsize=(10, 2))
+    im = ax.imshow(spectrogram, aspect="auto", origin="lower", interpolation="none")
+    plt.colorbar(im, ax=ax)
+    plt.xlabel("Frames")
+    plt.ylabel("Channels")
+    plt.tight_layout()
+
+    fig.canvas.draw()
+    data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep="")
+    data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
+    plt.close()
+    return data
+
+
+def plot_alignment_to_numpy(alignment, info=None):
+    global MATPLOTLIB_FLAG
+    if not MATPLOTLIB_FLAG:
+        import matplotlib
+
+        matplotlib.use("Agg")
+        MATPLOTLIB_FLAG = True
+        mpl_logger = logging.getLogger("matplotlib")
+        mpl_logger.setLevel(logging.WARNING)
+    import matplotlib.pylab as plt
+    import numpy as np
+
+    fig, ax = plt.subplots(figsize=(6, 4))
+    im = ax.imshow(
+        alignment.transpose(), aspect="auto", origin="lower", interpolation="none"
+    )
+    fig.colorbar(im, ax=ax)
+    xlabel = "Decoder timestep"
+    if info is not None:
+        xlabel += "\n\n" + info
+    plt.xlabel(xlabel)
+    plt.ylabel("Encoder timestep")
+    plt.tight_layout()
+
+    fig.canvas.draw()
+    data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep="")
+    data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
+    plt.close()
+    return data
+
+
+def load_wav_to_torch(full_path):
+    sampling_rate, data = read(full_path)
+    return torch.FloatTensor(data.astype(np.float32)), sampling_rate
+
+
+def load_filepaths_and_text(filename, split="|"):
+    with open(filename, encoding="utf-8") as f:
+        filepaths_and_text = [line.strip().split(split) for line in f]
+    return filepaths_and_text
+
+
+def get_hparams(init=True):
+    """
+    todo:
+      结尾七人组：
+        保存频率、总epoch                     done
+        bs                                    done
+        pretrainG、pretrainD                  done
+        卡号：os.en["CUDA_VISIBLE_DEVICES"]   done
+        if_latest                             done
+      模型：if_f0                             done
+      采样率：自动选择config                  done
+      是否缓存数据集进GPU:if_cache_data_in_gpu done
+
+      -m:
+        自动决定training_files路径,改掉train_nsf_load_pretrain.py里的hps.data.training_files    done
+      -c不要了
+    """
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "-se",
+        "--save_every_epoch",
+        type=int,
+        required=True,
+        help="checkpoint save frequency (epoch)",
+    )
+    parser.add_argument(
+        "-te", "--total_epoch", type=int, required=True, help="total_epoch"
+    )
+    parser.add_argument(
+        "-pg", "--pretrainG", type=str, default="", help="Pretrained Generator path"
+    )
+    parser.add_argument(
+        "-pd", "--pretrainD", type=str, default="", help="Pretrained Discriminator path"
+    )
+    parser.add_argument("-g", "--gpus", type=str, default="0", help="split by -")
+    parser.add_argument(
+        "-bs", "--batch_size", type=int, required=True, help="batch size"
+    )
+    parser.add_argument(
+        "-e", "--experiment_dir", type=str, required=True, help="experiment dir"
+    )  # -m
+    parser.add_argument(
+        "-sr", "--sample_rate", type=str, required=True, help="sample rate, 32k/40k/48k"
+    )
+    parser.add_argument(
+        "-sw",
+        "--save_every_weights",
+        type=str,
+        default="0",
+        help="save the extracted model in weights directory when saving checkpoints",
+    )
+    parser.add_argument(
+        "-v", "--version", type=str, required=True, help="model version"
+    )
+    parser.add_argument(
+        "-f0",
+        "--if_f0",
+        type=int,
+        required=True,
+        help="use f0 as one of the inputs of the model, 1 or 0",
+    )
+    parser.add_argument(
+        "-l",
+        "--if_latest",
+        type=int,
+        required=True,
+        help="if only save the latest G/D pth file, 1 or 0",
+    )
+    parser.add_argument(
+        "-c",
+        "--if_cache_data_in_gpu",
+        type=int,
+        required=True,
+        help="if caching the dataset in GPU memory, 1 or 0",
+    )
+
+    args = parser.parse_args()
+    name = args.experiment_dir
+    experiment_dir = os.path.join("./logs", args.experiment_dir)
+
+    config_save_path = os.path.join(experiment_dir, "config.json")
+    with open(config_save_path, "r") as f:
+        config = json.load(f)
+
+    hparams = HParams(**config)
+    hparams.model_dir = hparams.experiment_dir = experiment_dir
+    hparams.save_every_epoch = args.save_every_epoch
+    hparams.name = name
+    hparams.total_epoch = args.total_epoch
+    hparams.pretrainG = args.pretrainG
+    hparams.pretrainD = args.pretrainD
+    hparams.version = args.version
+    hparams.gpus = args.gpus
+    hparams.train.batch_size = args.batch_size
+    hparams.sample_rate = args.sample_rate
+    hparams.if_f0 = args.if_f0
+    hparams.if_latest = args.if_latest
+    hparams.save_every_weights = args.save_every_weights
+    hparams.if_cache_data_in_gpu = args.if_cache_data_in_gpu
+    hparams.data.training_files = "%s/filelist.txt" % experiment_dir
+    return hparams
+
+
+def get_hparams_from_dir(model_dir):
+    config_save_path = os.path.join(model_dir, "config.json")
+    with open(config_save_path, "r") as f:
+        data = f.read()
+    config = json.loads(data)
+
+    hparams = HParams(**config)
+    hparams.model_dir = model_dir
+    return hparams
+
+
+def get_hparams_from_file(config_path):
+    with open(config_path, "r") as f:
+        data = f.read()
+    config = json.loads(data)
+
+    hparams = HParams(**config)
+    return hparams
+
+
+def check_git_hash(model_dir):
+    source_dir = os.path.dirname(os.path.realpath(__file__))
+    if not os.path.exists(os.path.join(source_dir, ".git")):
+        logger.warning(
+            "{} is not a git repository, therefore hash value comparison will be ignored.".format(
+                source_dir
+            )
+        )
+        return
+
+    cur_hash = subprocess.getoutput("git rev-parse HEAD")
+
+    path = os.path.join(model_dir, "githash")
+    if os.path.exists(path):
+        saved_hash = open(path).read()
+        if saved_hash != cur_hash:
+            logger.warning(
+                "git hash values are different. {}(saved) != {}(current)".format(
+                    saved_hash[:8], cur_hash[:8]
+                )
+            )
+    else:
+        open(path, "w").write(cur_hash)
+
+
+def get_logger(model_dir, filename="train.log"):
+    global logger
+    logger = logging.getLogger(os.path.basename(model_dir))
+    logger.setLevel(logging.DEBUG)
+
+    formatter = logging.Formatter("%(asctime)s\t%(name)s\t%(levelname)s\t%(message)s")
+    if not os.path.exists(model_dir):
+        os.makedirs(model_dir)
+    h = logging.FileHandler(os.path.join(model_dir, filename))
+    h.setLevel(logging.DEBUG)
+    h.setFormatter(formatter)
+    logger.addHandler(h)
+    return logger
+
+
+class HParams:
+    def __init__(self, **kwargs):
+        for k, v in kwargs.items():
+            if type(v) == dict:
+                v = HParams(**v)
+            self[k] = v
+
+    def keys(self):
+        return self.__dict__.keys()
+
+    def items(self):
+        return self.__dict__.items()
+
+    def values(self):
+        return self.__dict__.values()
+
+    def __len__(self):
+        return len(self.__dict__)
+
+    def __getitem__(self, key):
+        return getattr(self, key)
+
+    def __setitem__(self, key, value):
+        return setattr(self, key, value)
+
+    def __contains__(self, key):
+        return key in self.__dict__
+
+    def __repr__(self):
+        return self.__dict__.__repr__()

infer/lib/uvr5_pack/lib_v5/dataset.py

@@ -0,0 +1,183 @@
+import os
+import random
+
+import numpy as np
+import torch
+import torch.utils.data
+from tqdm import tqdm
+
+from . import spec_utils
+
+
+class VocalRemoverValidationSet(torch.utils.data.Dataset):
+    def __init__(self, patch_list):
+        self.patch_list = patch_list
+
+    def __len__(self):
+        return len(self.patch_list)
+
+    def __getitem__(self, idx):
+        path = self.patch_list[idx]
+        data = np.load(path)
+
+        X, y = data["X"], data["y"]
+
+        X_mag = np.abs(X)
+        y_mag = np.abs(y)
+
+        return X_mag, y_mag
+
+
+def make_pair(mix_dir, inst_dir):
+    input_exts = [".wav", ".m4a", ".mp3", ".mp4", ".flac"]
+
+    X_list = sorted(
+        [
+            os.path.join(mix_dir, fname)
+            for fname in os.listdir(mix_dir)
+            if os.path.splitext(fname)[1] in input_exts
+        ]
+    )
+    y_list = sorted(
+        [
+            os.path.join(inst_dir, fname)
+            for fname in os.listdir(inst_dir)
+            if os.path.splitext(fname)[1] in input_exts
+        ]
+    )
+
+    filelist = list(zip(X_list, y_list))
+
+    return filelist
+
+
+def train_val_split(dataset_dir, split_mode, val_rate, val_filelist):
+    if split_mode == "random":
+        filelist = make_pair(
+            os.path.join(dataset_dir, "mixtures"),
+            os.path.join(dataset_dir, "instruments"),
+        )
+
+        random.shuffle(filelist)
+
+        if len(val_filelist) == 0:
+            val_size = int(len(filelist) * val_rate)
+            train_filelist = filelist[:-val_size]
+            val_filelist = filelist[-val_size:]
+        else:
+            train_filelist = [
+                pair for pair in filelist if list(pair) not in val_filelist
+            ]
+    elif split_mode == "subdirs":
+        if len(val_filelist) != 0:
+            raise ValueError(
+                "The `val_filelist` option is not available in `subdirs` mode"
+            )
+
+        train_filelist = make_pair(
+            os.path.join(dataset_dir, "training/mixtures"),
+            os.path.join(dataset_dir, "training/instruments"),
+        )
+
+        val_filelist = make_pair(
+            os.path.join(dataset_dir, "validation/mixtures"),
+            os.path.join(dataset_dir, "validation/instruments"),
+        )
+
+    return train_filelist, val_filelist
+
+
+def augment(X, y, reduction_rate, reduction_mask, mixup_rate, mixup_alpha):
+    perm = np.random.permutation(len(X))
+    for i, idx in enumerate(tqdm(perm)):
+        if np.random.uniform() < reduction_rate:
+            y[idx] = spec_utils.reduce_vocal_aggressively(
+                X[idx], y[idx], reduction_mask
+            )
+
+        if np.random.uniform() < 0.5:
+            # swap channel
+            X[idx] = X[idx, ::-1]
+            y[idx] = y[idx, ::-1]
+        if np.random.uniform() < 0.02:
+            # mono
+            X[idx] = X[idx].mean(axis=0, keepdims=True)
+            y[idx] = y[idx].mean(axis=0, keepdims=True)
+        if np.random.uniform() < 0.02:
+            # inst
+            X[idx] = y[idx]
+
+        if np.random.uniform() < mixup_rate and i < len(perm) - 1:
+            lam = np.random.beta(mixup_alpha, mixup_alpha)
+            X[idx] = lam * X[idx] + (1 - lam) * X[perm[i + 1]]
+            y[idx] = lam * y[idx] + (1 - lam) * y[perm[i + 1]]
+
+    return X, y
+
+
+def make_padding(width, cropsize, offset):
+    left = offset
+    roi_size = cropsize - left * 2
+    if roi_size == 0:
+        roi_size = cropsize
+    right = roi_size - (width % roi_size) + left
+
+    return left, right, roi_size
+
+
+def make_training_set(filelist, cropsize, patches, sr, hop_length, n_fft, offset):
+    len_dataset = patches * len(filelist)
+
+    X_dataset = np.zeros((len_dataset, 2, n_fft // 2 + 1, cropsize), dtype=np.complex64)
+    y_dataset = np.zeros((len_dataset, 2, n_fft // 2 + 1, cropsize), dtype=np.complex64)
+
+    for i, (X_path, y_path) in enumerate(tqdm(filelist)):
+        X, y = spec_utils.cache_or_load(X_path, y_path, sr, hop_length, n_fft)
+        coef = np.max([np.abs(X).max(), np.abs(y).max()])
+        X, y = X / coef, y / coef
+
+        l, r, roi_size = make_padding(X.shape[2], cropsize, offset)
+        X_pad = np.pad(X, ((0, 0), (0, 0), (l, r)), mode="constant")
+        y_pad = np.pad(y, ((0, 0), (0, 0), (l, r)), mode="constant")
+
+        starts = np.random.randint(0, X_pad.shape[2] - cropsize, patches)
+        ends = starts + cropsize
+        for j in range(patches):
+            idx = i * patches + j
+            X_dataset[idx] = X_pad[:, :, starts[j] : ends[j]]
+            y_dataset[idx] = y_pad[:, :, starts[j] : ends[j]]
+
+    return X_dataset, y_dataset
+
+
+def make_validation_set(filelist, cropsize, sr, hop_length, n_fft, offset):
+    patch_list = []
+    patch_dir = "cs{}_sr{}_hl{}_nf{}_of{}".format(
+        cropsize, sr, hop_length, n_fft, offset
+    )
+    os.makedirs(patch_dir, exist_ok=True)
+
+    for i, (X_path, y_path) in enumerate(tqdm(filelist)):
+        basename = os.path.splitext(os.path.basename(X_path))[0]
+
+        X, y = spec_utils.cache_or_load(X_path, y_path, sr, hop_length, n_fft)
+        coef = np.max([np.abs(X).max(), np.abs(y).max()])
+        X, y = X / coef, y / coef
+
+        l, r, roi_size = make_padding(X.shape[2], cropsize, offset)
+        X_pad = np.pad(X, ((0, 0), (0, 0), (l, r)), mode="constant")
+        y_pad = np.pad(y, ((0, 0), (0, 0), (l, r)), mode="constant")
+
+        len_dataset = int(np.ceil(X.shape[2] / roi_size))
+        for j in range(len_dataset):
+            outpath = os.path.join(patch_dir, "{}_p{}.npz".format(basename, j))
+            start = j * roi_size
+            if not os.path.exists(outpath):
+                np.savez(
+                    outpath,
+                    X=X_pad[:, :, start : start + cropsize],
+                    y=y_pad[:, :, start : start + cropsize],
+                )
+            patch_list.append(outpath)
+
+    return VocalRemoverValidationSet(patch_list)

infer/lib/uvr5_pack/lib_v5/layers.py

@@ -0,0 +1,118 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(Conv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nout,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                bias=False,
+            ),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class SeperableConv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(SeperableConv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nin,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                groups=nin,
+                bias=False,
+            ),
+            nn.Conv2d(nin, nout, kernel_size=1, bias=False),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class Encoder(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+        super(Encoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.conv2 = Conv2DBNActiv(nout, nout, ksize, stride, pad, activ=activ)
+
+    def __call__(self, x):
+        skip = self.conv1(x)
+        h = self.conv2(skip)
+
+        return h, skip
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+    ):
+        super(Decoder, self).__init__()
+        self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def __call__(self, x, skip=None):
+        x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+        if skip is not None:
+            skip = spec_utils.crop_center(skip, x)
+            x = torch.cat([x, skip], dim=1)
+        h = self.conv(x)
+
+        if self.dropout is not None:
+            h = self.dropout(h)
+
+        return h
+
+
+class ASPPModule(nn.Module):
+    def __init__(self, nin, nout, dilations=(4, 8, 16), activ=nn.ReLU):
+        super(ASPPModule, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, None)),
+            Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
+        )
+        self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+        self.conv3 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
+        )
+        self.conv4 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
+        )
+        self.conv5 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.bottleneck = nn.Sequential(
+            Conv2DBNActiv(nin * 5, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
+        )
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+        feat1 = F.interpolate(
+            self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+        )
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5), dim=1)
+        bottle = self.bottleneck(out)
+        return bottle

infer/lib/uvr5_pack/lib_v5/layers_123812KB .py

@@ -0,0 +1,118 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(Conv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nout,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                bias=False,
+            ),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class SeperableConv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(SeperableConv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nin,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                groups=nin,
+                bias=False,
+            ),
+            nn.Conv2d(nin, nout, kernel_size=1, bias=False),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class Encoder(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+        super(Encoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.conv2 = Conv2DBNActiv(nout, nout, ksize, stride, pad, activ=activ)
+
+    def __call__(self, x):
+        skip = self.conv1(x)
+        h = self.conv2(skip)
+
+        return h, skip
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+    ):
+        super(Decoder, self).__init__()
+        self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def __call__(self, x, skip=None):
+        x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+        if skip is not None:
+            skip = spec_utils.crop_center(skip, x)
+            x = torch.cat([x, skip], dim=1)
+        h = self.conv(x)
+
+        if self.dropout is not None:
+            h = self.dropout(h)
+
+        return h
+
+
+class ASPPModule(nn.Module):
+    def __init__(self, nin, nout, dilations=(4, 8, 16), activ=nn.ReLU):
+        super(ASPPModule, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, None)),
+            Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
+        )
+        self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+        self.conv3 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
+        )
+        self.conv4 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
+        )
+        self.conv5 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.bottleneck = nn.Sequential(
+            Conv2DBNActiv(nin * 5, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
+        )
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+        feat1 = F.interpolate(
+            self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+        )
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5), dim=1)
+        bottle = self.bottleneck(out)
+        return bottle

infer/lib/uvr5_pack/lib_v5/layers_123821KB.py

@@ -0,0 +1,118 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(Conv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nout,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                bias=False,
+            ),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class SeperableConv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(SeperableConv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nin,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                groups=nin,
+                bias=False,
+            ),
+            nn.Conv2d(nin, nout, kernel_size=1, bias=False),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class Encoder(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+        super(Encoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.conv2 = Conv2DBNActiv(nout, nout, ksize, stride, pad, activ=activ)
+
+    def __call__(self, x):
+        skip = self.conv1(x)
+        h = self.conv2(skip)
+
+        return h, skip
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+    ):
+        super(Decoder, self).__init__()
+        self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def __call__(self, x, skip=None):
+        x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+        if skip is not None:
+            skip = spec_utils.crop_center(skip, x)
+            x = torch.cat([x, skip], dim=1)
+        h = self.conv(x)
+
+        if self.dropout is not None:
+            h = self.dropout(h)
+
+        return h
+
+
+class ASPPModule(nn.Module):
+    def __init__(self, nin, nout, dilations=(4, 8, 16), activ=nn.ReLU):
+        super(ASPPModule, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, None)),
+            Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
+        )
+        self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+        self.conv3 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
+        )
+        self.conv4 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
+        )
+        self.conv5 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.bottleneck = nn.Sequential(
+            Conv2DBNActiv(nin * 5, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
+        )
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+        feat1 = F.interpolate(
+            self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+        )
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5), dim=1)
+        bottle = self.bottleneck(out)
+        return bottle

infer/lib/uvr5_pack/lib_v5/layers_33966KB.py

@@ -0,0 +1,126 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(Conv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nout,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                bias=False,
+            ),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class SeperableConv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(SeperableConv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nin,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                groups=nin,
+                bias=False,
+            ),
+            nn.Conv2d(nin, nout, kernel_size=1, bias=False),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class Encoder(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+        super(Encoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.conv2 = Conv2DBNActiv(nout, nout, ksize, stride, pad, activ=activ)
+
+    def __call__(self, x):
+        skip = self.conv1(x)
+        h = self.conv2(skip)
+
+        return h, skip
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+    ):
+        super(Decoder, self).__init__()
+        self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def __call__(self, x, skip=None):
+        x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+        if skip is not None:
+            skip = spec_utils.crop_center(skip, x)
+            x = torch.cat([x, skip], dim=1)
+        h = self.conv(x)
+
+        if self.dropout is not None:
+            h = self.dropout(h)
+
+        return h
+
+
+class ASPPModule(nn.Module):
+    def __init__(self, nin, nout, dilations=(4, 8, 16, 32, 64), activ=nn.ReLU):
+        super(ASPPModule, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, None)),
+            Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
+        )
+        self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+        self.conv3 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
+        )
+        self.conv4 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
+        )
+        self.conv5 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.conv6 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.conv7 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.bottleneck = nn.Sequential(
+            Conv2DBNActiv(nin * 7, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
+        )
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+        feat1 = F.interpolate(
+            self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+        )
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        feat6 = self.conv6(x)
+        feat7 = self.conv7(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5, feat6, feat7), dim=1)
+        bottle = self.bottleneck(out)
+        return bottle

infer/lib/uvr5_pack/lib_v5/layers_537227KB.py

@@ -0,0 +1,126 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(Conv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nout,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                bias=False,
+            ),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class SeperableConv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(SeperableConv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nin,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                groups=nin,
+                bias=False,
+            ),
+            nn.Conv2d(nin, nout, kernel_size=1, bias=False),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class Encoder(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+        super(Encoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.conv2 = Conv2DBNActiv(nout, nout, ksize, stride, pad, activ=activ)
+
+    def __call__(self, x):
+        skip = self.conv1(x)
+        h = self.conv2(skip)
+
+        return h, skip
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+    ):
+        super(Decoder, self).__init__()
+        self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def __call__(self, x, skip=None):
+        x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+        if skip is not None:
+            skip = spec_utils.crop_center(skip, x)
+            x = torch.cat([x, skip], dim=1)
+        h = self.conv(x)
+
+        if self.dropout is not None:
+            h = self.dropout(h)
+
+        return h
+
+
+class ASPPModule(nn.Module):
+    def __init__(self, nin, nout, dilations=(4, 8, 16, 32, 64), activ=nn.ReLU):
+        super(ASPPModule, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, None)),
+            Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
+        )
+        self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+        self.conv3 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
+        )
+        self.conv4 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
+        )
+        self.conv5 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.conv6 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.conv7 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.bottleneck = nn.Sequential(
+            Conv2DBNActiv(nin * 7, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
+        )
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+        feat1 = F.interpolate(
+            self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+        )
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        feat6 = self.conv6(x)
+        feat7 = self.conv7(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5, feat6, feat7), dim=1)
+        bottle = self.bottleneck(out)
+        return bottle

infer/lib/uvr5_pack/lib_v5/layers_537238KB.py

@@ -0,0 +1,126 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(Conv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nout,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                bias=False,
+            ),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class SeperableConv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(SeperableConv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nin,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                groups=nin,
+                bias=False,
+            ),
+            nn.Conv2d(nin, nout, kernel_size=1, bias=False),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class Encoder(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+        super(Encoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.conv2 = Conv2DBNActiv(nout, nout, ksize, stride, pad, activ=activ)
+
+    def __call__(self, x):
+        skip = self.conv1(x)
+        h = self.conv2(skip)
+
+        return h, skip
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+    ):
+        super(Decoder, self).__init__()
+        self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def __call__(self, x, skip=None):
+        x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+        if skip is not None:
+            skip = spec_utils.crop_center(skip, x)
+            x = torch.cat([x, skip], dim=1)
+        h = self.conv(x)
+
+        if self.dropout is not None:
+            h = self.dropout(h)
+
+        return h
+
+
+class ASPPModule(nn.Module):
+    def __init__(self, nin, nout, dilations=(4, 8, 16, 32, 64), activ=nn.ReLU):
+        super(ASPPModule, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, None)),
+            Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
+        )
+        self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+        self.conv3 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
+        )
+        self.conv4 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
+        )
+        self.conv5 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.conv6 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.conv7 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.bottleneck = nn.Sequential(
+            Conv2DBNActiv(nin * 7, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
+        )
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+        feat1 = F.interpolate(
+            self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+        )
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        feat6 = self.conv6(x)
+        feat7 = self.conv7(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5, feat6, feat7), dim=1)
+        bottle = self.bottleneck(out)
+        return bottle

infer/lib/uvr5_pack/lib_v5/layers_new.py

@@ -0,0 +1,125 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(Conv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nout,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                bias=False,
+            ),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class Encoder(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+        super(Encoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, stride, pad, activ=activ)
+        self.conv2 = Conv2DBNActiv(nout, nout, ksize, 1, pad, activ=activ)
+
+    def __call__(self, x):
+        h = self.conv1(x)
+        h = self.conv2(h)
+
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+    ):
+        super(Decoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        # self.conv2 = Conv2DBNActiv(nout, nout, ksize, 1, pad, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def __call__(self, x, skip=None):
+        x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+
+        if skip is not None:
+            skip = spec_utils.crop_center(skip, x)
+            x = torch.cat([x, skip], dim=1)
+
+        h = self.conv1(x)
+        # h = self.conv2(h)
+
+        if self.dropout is not None:
+            h = self.dropout(h)
+
+        return h
+
+
+class ASPPModule(nn.Module):
+    def __init__(self, nin, nout, dilations=(4, 8, 12), activ=nn.ReLU, dropout=False):
+        super(ASPPModule, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, None)),
+            Conv2DBNActiv(nin, nout, 1, 1, 0, activ=activ),
+        )
+        self.conv2 = Conv2DBNActiv(nin, nout, 1, 1, 0, activ=activ)
+        self.conv3 = Conv2DBNActiv(
+            nin, nout, 3, 1, dilations[0], dilations[0], activ=activ
+        )
+        self.conv4 = Conv2DBNActiv(
+            nin, nout, 3, 1, dilations[1], dilations[1], activ=activ
+        )
+        self.conv5 = Conv2DBNActiv(
+            nin, nout, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.bottleneck = Conv2DBNActiv(nout * 5, nout, 1, 1, 0, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+        feat1 = F.interpolate(
+            self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+        )
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5), dim=1)
+        out = self.bottleneck(out)
+
+        if self.dropout is not None:
+            out = self.dropout(out)
+
+        return out
+
+
+class LSTMModule(nn.Module):
+    def __init__(self, nin_conv, nin_lstm, nout_lstm):
+        super(LSTMModule, self).__init__()
+        self.conv = Conv2DBNActiv(nin_conv, 1, 1, 1, 0)
+        self.lstm = nn.LSTM(
+            input_size=nin_lstm, hidden_size=nout_lstm // 2, bidirectional=True
+        )
+        self.dense = nn.Sequential(
+            nn.Linear(nout_lstm, nin_lstm), nn.BatchNorm1d(nin_lstm), nn.ReLU()
+        )
+
+    def forward(self, x):
+        N, _, nbins, nframes = x.size()
+        h = self.conv(x)[:, 0]  # N, nbins, nframes
+        h = h.permute(2, 0, 1)  # nframes, N, nbins
+        h, _ = self.lstm(h)
+        h = self.dense(h.reshape(-1, h.size()[-1]))  # nframes * N, nbins
+        h = h.reshape(nframes, N, 1, nbins)
+        h = h.permute(1, 2, 3, 0)
+
+        return h

infer/lib/uvr5_pack/lib_v5/model_param_init.py

@@ -0,0 +1,69 @@
+import json
+import os
+import pathlib
+
+default_param = {}
+default_param["bins"] = 768
+default_param["unstable_bins"] = 9  # training only
+default_param["reduction_bins"] = 762  # training only
+default_param["sr"] = 44100
+default_param["pre_filter_start"] = 757
+default_param["pre_filter_stop"] = 768
+default_param["band"] = {}
+
+
+default_param["band"][1] = {
+    "sr": 11025,
+    "hl": 128,
+    "n_fft": 960,
+    "crop_start": 0,
+    "crop_stop": 245,
+    "lpf_start": 61,  # inference only
+    "res_type": "polyphase",
+}
+
+default_param["band"][2] = {
+    "sr": 44100,
+    "hl": 512,
+    "n_fft": 1536,
+    "crop_start": 24,
+    "crop_stop": 547,
+    "hpf_start": 81,  # inference only
+    "res_type": "sinc_best",
+}
+
+
+def int_keys(d):
+    r = {}
+    for k, v in d:
+        if k.isdigit():
+            k = int(k)
+        r[k] = v
+    return r
+
+
+class ModelParameters(object):
+    def __init__(self, config_path=""):
+        if ".pth" == pathlib.Path(config_path).suffix:
+            import zipfile
+
+            with zipfile.ZipFile(config_path, "r") as zip:
+                self.param = json.loads(
+                    zip.read("param.json"), object_pairs_hook=int_keys
+                )
+        elif ".json" == pathlib.Path(config_path).suffix:
+            with open(config_path, "r") as f:
+                self.param = json.loads(f.read(), object_pairs_hook=int_keys)
+        else:
+            self.param = default_param
+
+        for k in [
+            "mid_side",
+            "mid_side_b",
+            "mid_side_b2",
+            "stereo_w",
+            "stereo_n",
+            "reverse",
+        ]:
+            if not k in self.param:
+                self.param[k] = False

infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr16000_hl512.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 16000,
+			"hl": 512,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 1024,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 16000,
+	"pre_filter_start": 1023,
+	"pre_filter_stop": 1024
+}

infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 32000,
+			"hl": 512,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 1024,
+			"hpf_start": -1,
+			"res_type": "kaiser_fast"
+		}
+	},
+	"sr": 32000,
+	"pre_filter_start": 1000,
+	"pre_filter_stop": 1021
+}

infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 33075,
+			"hl": 384,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 1024,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 33075,
+	"pre_filter_start": 1000,
+	"pre_filter_stop": 1021
+}

infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl1024.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 44100,
+			"hl": 1024,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 1024,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 1023,
+	"pre_filter_stop": 1024
+}

infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl256.json

@@ -0,0 +1,19 @@
+{
+	"bins": 256,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 44100,
+			"hl": 256,
+			"n_fft": 512,
+			"crop_start": 0,
+			"crop_stop": 256,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 256,
+	"pre_filter_stop": 256
+}

infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 1024,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 1023,
+	"pre_filter_stop": 1024
+}

infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512_cut.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 700,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 1023,
+	"pre_filter_stop": 700
+}

infer/lib/uvr5_pack/lib_v5/modelparams/2band_32000.json

@@ -0,0 +1,30 @@
+{
+	"bins": 768,
+	"unstable_bins": 7,
+	"reduction_bins": 705,
+	"band": {
+		"1": {
+			"sr": 6000,
+			"hl": 66,
+			"n_fft": 512,
+			"crop_start": 0,
+			"crop_stop": 240,
+			"lpf_start": 60,
+			"lpf_stop": 118,
+			"res_type": "sinc_fastest"
+		},
+		"2": {
+			"sr": 32000,
+			"hl": 352,
+			"n_fft": 1024,
+			"crop_start": 22,
+			"crop_stop": 505,
+			"hpf_start": 44,
+			"hpf_stop": 23,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 32000,
+	"pre_filter_start": 710,
+	"pre_filter_stop": 731
+}

infer/lib/uvr5_pack/lib_v5/modelparams/2band_44100_lofi.json

@@ -0,0 +1,30 @@
+{
+	"bins": 512,
+	"unstable_bins": 7,
+	"reduction_bins": 510,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 160,
+			"n_fft": 768,
+			"crop_start": 0,
+			"crop_stop": 192,
+			"lpf_start": 41,
+			"lpf_stop": 139,
+			"res_type": "sinc_fastest"
+		},
+		"2": {
+			"sr": 44100,
+			"hl": 640,
+			"n_fft": 1024,
+			"crop_start": 10,
+			"crop_stop": 320,
+			"hpf_start": 47,
+			"hpf_stop": 15,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 510,
+	"pre_filter_stop": 512
+}

infer/lib/uvr5_pack/lib_v5/modelparams/2band_48000.json

@@ -0,0 +1,30 @@
+{
+	"bins": 768,
+	"unstable_bins": 7,
+	"reduction_bins": 705,
+	"band": {
+		"1": {
+			"sr": 6000,
+			"hl": 66,
+			"n_fft": 512,
+			"crop_start": 0,
+			"crop_stop": 240,
+			"lpf_start": 60,
+			"lpf_stop": 240,
+			"res_type": "sinc_fastest"
+		},
+		"2": {
+			"sr": 48000,
+			"hl": 528,
+			"n_fft": 1536,
+			"crop_start": 22,
+			"crop_stop": 505,
+			"hpf_start": 82,
+			"hpf_stop": 22,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 48000,
+	"pre_filter_start": 710,
+	"pre_filter_stop": 731
+}

infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100.json

@@ -0,0 +1,42 @@
+{
+	"bins": 768,
+	"unstable_bins": 5,
+	"reduction_bins": 733,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 768,
+			"crop_start": 0,
+			"crop_stop": 278,
+			"lpf_start": 28,
+			"lpf_stop": 140,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 768,
+			"crop_start": 14,
+			"crop_stop": 322,
+			"hpf_start": 70,
+			"hpf_stop": 14,
+			"lpf_start": 283,
+			"lpf_stop": 314,
+			"res_type": "polyphase"
+		},	
+		"3": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 131,
+			"crop_stop": 313,
+			"hpf_start": 154,
+			"hpf_stop": 141,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 757,
+	"pre_filter_stop": 768
+}

infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_mid.json

@@ -0,0 +1,43 @@
+{
+	"mid_side": true,
+	"bins": 768,
+	"unstable_bins": 5,
+	"reduction_bins": 733,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 768,
+			"crop_start": 0,
+			"crop_stop": 278,
+			"lpf_start": 28,
+			"lpf_stop": 140,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 768,
+			"crop_start": 14,
+			"crop_stop": 322,
+			"hpf_start": 70,
+			"hpf_stop": 14,
+			"lpf_start": 283,
+			"lpf_stop": 314,
+			"res_type": "polyphase"
+		},	
+		"3": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 131,
+			"crop_stop": 313,
+			"hpf_start": 154,
+			"hpf_stop": 141,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 757,
+	"pre_filter_stop": 768
+}

infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_msb2.json

@@ -0,0 +1,43 @@
+{
+	"mid_side_b2": true,
+	"bins": 640,
+	"unstable_bins": 7,
+	"reduction_bins": 565,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 108,
+			"n_fft": 1024,
+			"crop_start": 0,
+			"crop_stop": 187,
+			"lpf_start": 92,
+			"lpf_stop": 186,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 22050,
+			"hl": 216,
+			"n_fft": 768,
+			"crop_start": 0,
+			"crop_stop": 212,
+			"hpf_start": 68,
+			"hpf_stop": 34,
+			"lpf_start": 174,
+			"lpf_stop": 209,
+			"res_type": "polyphase"
+		},	
+		"3": {
+			"sr": 44100,
+			"hl": 432,
+			"n_fft": 640,
+			"crop_start": 66,
+			"crop_stop": 307,
+			"hpf_start": 86,
+			"hpf_stop": 72,
+			"res_type": "kaiser_fast"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 639,
+	"pre_filter_stop": 640
+}

infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json

@@ -0,0 +1,54 @@
+{
+	"bins": 768,
+	"unstable_bins": 7,
+	"reduction_bins": 668,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 1024,
+			"crop_start": 0,
+			"crop_stop": 186,
+			"lpf_start": 37,
+			"lpf_stop": 73,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 512,
+			"crop_start": 4,
+			"crop_stop": 185,			
+			"hpf_start": 36,
+			"hpf_stop": 18,
+			"lpf_start": 93,
+			"lpf_stop": 185,
+			"res_type": "polyphase"
+		},
+		"3": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 512,
+			"crop_start": 46,
+			"crop_stop": 186,
+			"hpf_start": 93,
+			"hpf_stop": 46,
+			"lpf_start": 164,
+			"lpf_stop": 186,
+			"res_type": "polyphase"
+		},	
+		"4": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 121,
+			"crop_stop": 382,
+			"hpf_start": 138,
+			"hpf_stop": 123,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 740,
+	"pre_filter_stop": 768
+}

infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_mid.json

@@ -0,0 +1,55 @@
+{
+	"bins": 768,
+	"unstable_bins": 7,
+	"mid_side": true,
+	"reduction_bins": 668,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 1024,
+			"crop_start": 0,
+			"crop_stop": 186,
+			"lpf_start": 37,
+			"lpf_stop": 73,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 512,
+			"crop_start": 4,
+			"crop_stop": 185,			
+			"hpf_start": 36,
+			"hpf_stop": 18,
+			"lpf_start": 93,
+			"lpf_stop": 185,
+			"res_type": "polyphase"
+		},
+		"3": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 512,
+			"crop_start": 46,
+			"crop_stop": 186,
+			"hpf_start": 93,
+			"hpf_stop": 46,
+			"lpf_start": 164,
+			"lpf_stop": 186,
+			"res_type": "polyphase"
+		},	
+		"4": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 121,
+			"crop_stop": 382,
+			"hpf_start": 138,
+			"hpf_stop": 123,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 740,
+	"pre_filter_stop": 768
+}

infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_msb.json

@@ -0,0 +1,55 @@
+{
+	"mid_side_b": true,
+	"bins": 768,
+	"unstable_bins": 7,
+	"reduction_bins": 668,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 1024,
+			"crop_start": 0,
+			"crop_stop": 186,
+			"lpf_start": 37,
+			"lpf_stop": 73,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 512,
+			"crop_start": 4,
+			"crop_stop": 185,			
+			"hpf_start": 36,
+			"hpf_stop": 18,
+			"lpf_start": 93,
+			"lpf_stop": 185,
+			"res_type": "polyphase"
+		},
+		"3": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 512,
+			"crop_start": 46,
+			"crop_stop": 186,
+			"hpf_start": 93,
+			"hpf_stop": 46,
+			"lpf_start": 164,
+			"lpf_stop": 186,
+			"res_type": "polyphase"
+		},	
+		"4": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 121,
+			"crop_stop": 382,
+			"hpf_start": 138,
+			"hpf_stop": 123,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 740,
+	"pre_filter_stop": 768
+}

infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_msb2.json

@@ -0,0 +1,55 @@
+{
+	"mid_side_b": true,
+	"bins": 768,
+	"unstable_bins": 7,
+	"reduction_bins": 668,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 1024,
+			"crop_start": 0,
+			"crop_stop": 186,
+			"lpf_start": 37,
+			"lpf_stop": 73,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 512,
+			"crop_start": 4,
+			"crop_stop": 185,			
+			"hpf_start": 36,
+			"hpf_stop": 18,
+			"lpf_start": 93,
+			"lpf_stop": 185,
+			"res_type": "polyphase"
+		},
+		"3": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 512,
+			"crop_start": 46,
+			"crop_stop": 186,
+			"hpf_start": 93,
+			"hpf_stop": 46,
+			"lpf_start": 164,
+			"lpf_stop": 186,
+			"res_type": "polyphase"
+		},	
+		"4": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 121,
+			"crop_stop": 382,
+			"hpf_start": 138,
+			"hpf_stop": 123,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 740,
+	"pre_filter_stop": 768
+}

infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_reverse.json

@@ -0,0 +1,55 @@
+{
+	"reverse": true,
+	"bins": 768,
+	"unstable_bins": 7,
+	"reduction_bins": 668,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 1024,
+			"crop_start": 0,
+			"crop_stop": 186,
+			"lpf_start": 37,
+			"lpf_stop": 73,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 512,
+			"crop_start": 4,
+			"crop_stop": 185,			
+			"hpf_start": 36,
+			"hpf_stop": 18,
+			"lpf_start": 93,
+			"lpf_stop": 185,
+			"res_type": "polyphase"
+		},
+		"3": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 512,
+			"crop_start": 46,
+			"crop_stop": 186,
+			"hpf_start": 93,
+			"hpf_stop": 46,
+			"lpf_start": 164,
+			"lpf_stop": 186,
+			"res_type": "polyphase"
+		},	
+		"4": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 121,
+			"crop_stop": 382,
+			"hpf_start": 138,
+			"hpf_stop": 123,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 740,
+	"pre_filter_stop": 768
+}

infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_sw.json

@@ -0,0 +1,55 @@
+{
+	"stereo_w": true,
+	"bins": 768,
+	"unstable_bins": 7,
+	"reduction_bins": 668,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 1024,
+			"crop_start": 0,
+			"crop_stop": 186,
+			"lpf_start": 37,
+			"lpf_stop": 73,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 512,
+			"crop_start": 4,
+			"crop_stop": 185,			
+			"hpf_start": 36,
+			"hpf_stop": 18,
+			"lpf_start": 93,
+			"lpf_stop": 185,
+			"res_type": "polyphase"
+		},
+		"3": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 512,
+			"crop_start": 46,
+			"crop_stop": 186,
+			"hpf_start": 93,
+			"hpf_stop": 46,
+			"lpf_start": 164,
+			"lpf_stop": 186,
+			"res_type": "polyphase"
+		},	
+		"4": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 121,
+			"crop_stop": 382,
+			"hpf_start": 138,
+			"hpf_stop": 123,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 740,
+	"pre_filter_stop": 768
+}

infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2.json

@@ -0,0 +1,54 @@
+{
+	"bins": 672,
+	"unstable_bins": 8,
+	"reduction_bins": 637,
+	"band": {
+		"1": {
+			"sr": 7350,
+			"hl": 80,
+			"n_fft": 640,
+			"crop_start": 0,
+			"crop_stop": 85,
+			"lpf_start": 25,
+			"lpf_stop": 53,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 7350,
+			"hl": 80,
+			"n_fft": 320,
+			"crop_start": 4,
+			"crop_stop": 87,
+			"hpf_start": 25,
+			"hpf_stop": 12,
+			"lpf_start": 31,
+			"lpf_stop": 62,
+			"res_type": "polyphase"
+		},		
+		"3": {
+			"sr": 14700,
+			"hl": 160,
+			"n_fft": 512,
+			"crop_start": 17,
+			"crop_stop": 216,
+			"hpf_start": 48,
+			"hpf_stop": 24,
+			"lpf_start": 139,
+			"lpf_stop": 210,
+			"res_type": "polyphase"
+		},	
+		"4": {
+			"sr": 44100,
+			"hl": 480,
+			"n_fft": 960,
+			"crop_start": 78,
+			"crop_stop": 383,
+			"hpf_start": 130,
+			"hpf_stop": 86,
+			"res_type": "kaiser_fast"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 668,
+	"pre_filter_stop": 672
+}

infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2_sn.json

@@ -0,0 +1,55 @@
+{
+	"bins": 672,
+	"unstable_bins": 8,
+	"reduction_bins": 637,
+	"band": {
+		"1": {
+			"sr": 7350,
+			"hl": 80,
+			"n_fft": 640,
+			"crop_start": 0,
+			"crop_stop": 85,
+			"lpf_start": 25,
+			"lpf_stop": 53,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 7350,
+			"hl": 80,
+			"n_fft": 320,
+			"crop_start": 4,
+			"crop_stop": 87,
+			"hpf_start": 25,
+			"hpf_stop": 12,
+			"lpf_start": 31,
+			"lpf_stop": 62,
+			"res_type": "polyphase"
+		},		
+		"3": {
+			"sr": 14700,
+			"hl": 160,
+			"n_fft": 512,
+			"crop_start": 17,
+			"crop_stop": 216,
+			"hpf_start": 48,
+			"hpf_stop": 24,
+			"lpf_start": 139,
+			"lpf_stop": 210,
+			"res_type": "polyphase"
+		},	
+		"4": {
+			"sr": 44100,
+			"hl": 480,
+			"n_fft": 960,
+			"crop_start": 78,
+			"crop_stop": 383,
+			"hpf_start": 130,
+			"hpf_stop": 86,
+			"convert_channels": "stereo_n",
+			"res_type": "kaiser_fast"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 668,
+	"pre_filter_stop": 672
+}

infer/lib/uvr5_pack/lib_v5/modelparams/4band_v3.json

@@ -0,0 +1,54 @@
+{
+	"bins": 672,
+	"unstable_bins": 8,
+	"reduction_bins": 530,
+	"band": {
+		"1": {
+			"sr": 7350,
+			"hl": 80,
+			"n_fft": 640,
+			"crop_start": 0,
+			"crop_stop": 85,
+			"lpf_start": 25,
+			"lpf_stop": 53,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 7350,
+			"hl": 80,
+			"n_fft": 320,
+			"crop_start": 4,
+			"crop_stop": 87,
+			"hpf_start": 25,
+			"hpf_stop": 12,
+			"lpf_start": 31,
+			"lpf_stop": 62,
+			"res_type": "polyphase"
+		},
+		"3": {
+			"sr": 14700,
+			"hl": 160,
+			"n_fft": 512,
+			"crop_start": 17,
+			"crop_stop": 216,
+			"hpf_start": 48,
+			"hpf_stop": 24,
+			"lpf_start": 139,
+			"lpf_stop": 210,
+			"res_type": "polyphase"
+		},
+		"4": {
+			"sr": 44100,
+			"hl": 480,
+			"n_fft": 960,
+			"crop_start": 78,
+			"crop_stop": 383,
+			"hpf_start": 130,
+			"hpf_stop": 86,
+			"res_type": "kaiser_fast"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 668,
+	"pre_filter_stop": 672
+}

infer/lib/uvr5_pack/lib_v5/modelparams/ensemble.json

@@ -0,0 +1,43 @@
+{
+	"mid_side_b2": true,
+	"bins": 1280,
+	"unstable_bins": 7,
+	"reduction_bins": 565,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 108,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 374,
+			"lpf_start": 92,
+			"lpf_stop": 186,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 22050,
+			"hl": 216,
+			"n_fft": 1536,
+			"crop_start": 0,
+			"crop_stop": 424,
+			"hpf_start": 68,
+			"hpf_stop": 34,
+			"lpf_start": 348,
+			"lpf_stop": 418,
+			"res_type": "polyphase"
+		},	
+		"3": {
+			"sr": 44100,
+			"hl": 432,
+			"n_fft": 1280,
+			"crop_start": 132,
+			"crop_stop": 614,
+			"hpf_start": 172,
+			"hpf_stop": 144,
+			"res_type": "polyphase"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 1280,
+	"pre_filter_stop": 1280
+}

infer/lib/uvr5_pack/lib_v5/nets.py

@@ -0,0 +1,123 @@
+import layers
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class BaseASPPNet(nn.Module):
+    def __init__(self, nin, ch, dilations=(4, 8, 16)):
+        super(BaseASPPNet, self).__init__()
+        self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
+        self.enc2 = layers.Encoder(ch, ch * 2, 3, 2, 1)
+        self.enc3 = layers.Encoder(ch * 2, ch * 4, 3, 2, 1)
+        self.enc4 = layers.Encoder(ch * 4, ch * 8, 3, 2, 1)
+
+        self.aspp = layers.ASPPModule(ch * 8, ch * 16, dilations)
+
+        self.dec4 = layers.Decoder(ch * (8 + 16), ch * 8, 3, 1, 1)
+        self.dec3 = layers.Decoder(ch * (4 + 8), ch * 4, 3, 1, 1)
+        self.dec2 = layers.Decoder(ch * (2 + 4), ch * 2, 3, 1, 1)
+        self.dec1 = layers.Decoder(ch * (1 + 2), ch, 3, 1, 1)
+
+    def __call__(self, x):
+        h, e1 = self.enc1(x)
+        h, e2 = self.enc2(h)
+        h, e3 = self.enc3(h)
+        h, e4 = self.enc4(h)
+
+        h = self.aspp(h)
+
+        h = self.dec4(h, e4)
+        h = self.dec3(h, e3)
+        h = self.dec2(h, e2)
+        h = self.dec1(h, e1)
+
+        return h
+
+
+class CascadedASPPNet(nn.Module):
+    def __init__(self, n_fft):
+        super(CascadedASPPNet, self).__init__()
+        self.stg1_low_band_net = BaseASPPNet(2, 16)
+        self.stg1_high_band_net = BaseASPPNet(2, 16)
+
+        self.stg2_bridge = layers.Conv2DBNActiv(18, 8, 1, 1, 0)
+        self.stg2_full_band_net = BaseASPPNet(8, 16)
+
+        self.stg3_bridge = layers.Conv2DBNActiv(34, 16, 1, 1, 0)
+        self.stg3_full_band_net = BaseASPPNet(16, 32)
+
+        self.out = nn.Conv2d(32, 2, 1, bias=False)
+        self.aux1_out = nn.Conv2d(16, 2, 1, bias=False)
+        self.aux2_out = nn.Conv2d(16, 2, 1, bias=False)
+
+        self.max_bin = n_fft // 2
+        self.output_bin = n_fft // 2 + 1
+
+        self.offset = 128
+
+    def forward(self, x, aggressiveness=None):
+        mix = x.detach()
+        x = x.clone()
+
+        x = x[:, :, : self.max_bin]
+
+        bandw = x.size()[2] // 2
+        aux1 = torch.cat(
+            [
+                self.stg1_low_band_net(x[:, :, :bandw]),
+                self.stg1_high_band_net(x[:, :, bandw:]),
+            ],
+            dim=2,
+        )
+
+        h = torch.cat([x, aux1], dim=1)
+        aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
+
+        h = torch.cat([x, aux1, aux2], dim=1)
+        h = self.stg3_full_band_net(self.stg3_bridge(h))
+
+        mask = torch.sigmoid(self.out(h))
+        mask = F.pad(
+            input=mask,
+            pad=(0, 0, 0, self.output_bin - mask.size()[2]),
+            mode="replicate",
+        )
+
+        if self.training:
+            aux1 = torch.sigmoid(self.aux1_out(aux1))
+            aux1 = F.pad(
+                input=aux1,
+                pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
+                mode="replicate",
+            )
+            aux2 = torch.sigmoid(self.aux2_out(aux2))
+            aux2 = F.pad(
+                input=aux2,
+                pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
+                mode="replicate",
+            )
+            return mask * mix, aux1 * mix, aux2 * mix
+        else:
+            if aggressiveness:
+                mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
+                    mask[:, :, : aggressiveness["split_bin"]],
+                    1 + aggressiveness["value"] / 3,
+                )
+                mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
+                    mask[:, :, aggressiveness["split_bin"] :],
+                    1 + aggressiveness["value"],
+                )
+
+            return mask * mix
+
+    def predict(self, x_mag, aggressiveness=None):
+        h = self.forward(x_mag, aggressiveness)
+
+        if self.offset > 0:
+            h = h[:, :, :, self.offset : -self.offset]
+            assert h.size()[3] > 0
+
+        return h

infer/lib/uvr5_pack/lib_v5/nets_123812KB.py

@@ -0,0 +1,122 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import layers_123821KB as layers
+
+
+class BaseASPPNet(nn.Module):
+    def __init__(self, nin, ch, dilations=(4, 8, 16)):
+        super(BaseASPPNet, self).__init__()
+        self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
+        self.enc2 = layers.Encoder(ch, ch * 2, 3, 2, 1)
+        self.enc3 = layers.Encoder(ch * 2, ch * 4, 3, 2, 1)
+        self.enc4 = layers.Encoder(ch * 4, ch * 8, 3, 2, 1)
+
+        self.aspp = layers.ASPPModule(ch * 8, ch * 16, dilations)
+
+        self.dec4 = layers.Decoder(ch * (8 + 16), ch * 8, 3, 1, 1)
+        self.dec3 = layers.Decoder(ch * (4 + 8), ch * 4, 3, 1, 1)
+        self.dec2 = layers.Decoder(ch * (2 + 4), ch * 2, 3, 1, 1)
+        self.dec1 = layers.Decoder(ch * (1 + 2), ch, 3, 1, 1)
+
+    def __call__(self, x):
+        h, e1 = self.enc1(x)
+        h, e2 = self.enc2(h)
+        h, e3 = self.enc3(h)
+        h, e4 = self.enc4(h)
+
+        h = self.aspp(h)
+
+        h = self.dec4(h, e4)
+        h = self.dec3(h, e3)
+        h = self.dec2(h, e2)
+        h = self.dec1(h, e1)
+
+        return h
+
+
+class CascadedASPPNet(nn.Module):
+    def __init__(self, n_fft):
+        super(CascadedASPPNet, self).__init__()
+        self.stg1_low_band_net = BaseASPPNet(2, 32)
+        self.stg1_high_band_net = BaseASPPNet(2, 32)
+
+        self.stg2_bridge = layers.Conv2DBNActiv(34, 16, 1, 1, 0)
+        self.stg2_full_band_net = BaseASPPNet(16, 32)
+
+        self.stg3_bridge = layers.Conv2DBNActiv(66, 32, 1, 1, 0)
+        self.stg3_full_band_net = BaseASPPNet(32, 64)
+
+        self.out = nn.Conv2d(64, 2, 1, bias=False)
+        self.aux1_out = nn.Conv2d(32, 2, 1, bias=False)
+        self.aux2_out = nn.Conv2d(32, 2, 1, bias=False)
+
+        self.max_bin = n_fft // 2
+        self.output_bin = n_fft // 2 + 1
+
+        self.offset = 128
+
+    def forward(self, x, aggressiveness=None):
+        mix = x.detach()
+        x = x.clone()
+
+        x = x[:, :, : self.max_bin]
+
+        bandw = x.size()[2] // 2
+        aux1 = torch.cat(
+            [
+                self.stg1_low_band_net(x[:, :, :bandw]),
+                self.stg1_high_band_net(x[:, :, bandw:]),
+            ],
+            dim=2,
+        )
+
+        h = torch.cat([x, aux1], dim=1)
+        aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
+
+        h = torch.cat([x, aux1, aux2], dim=1)
+        h = self.stg3_full_band_net(self.stg3_bridge(h))
+
+        mask = torch.sigmoid(self.out(h))
+        mask = F.pad(
+            input=mask,
+            pad=(0, 0, 0, self.output_bin - mask.size()[2]),
+            mode="replicate",
+        )
+
+        if self.training:
+            aux1 = torch.sigmoid(self.aux1_out(aux1))
+            aux1 = F.pad(
+                input=aux1,
+                pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
+                mode="replicate",
+            )
+            aux2 = torch.sigmoid(self.aux2_out(aux2))
+            aux2 = F.pad(
+                input=aux2,
+                pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
+                mode="replicate",
+            )
+            return mask * mix, aux1 * mix, aux2 * mix
+        else:
+            if aggressiveness:
+                mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
+                    mask[:, :, : aggressiveness["split_bin"]],
+                    1 + aggressiveness["value"] / 3,
+                )
+                mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
+                    mask[:, :, aggressiveness["split_bin"] :],
+                    1 + aggressiveness["value"],
+                )
+
+            return mask * mix
+
+    def predict(self, x_mag, aggressiveness=None):
+        h = self.forward(x_mag, aggressiveness)
+
+        if self.offset > 0:
+            h = h[:, :, :, self.offset : -self.offset]
+            assert h.size()[3] > 0
+
+        return h

infer/lib/uvr5_pack/lib_v5/nets_123821KB.py

@@ -0,0 +1,122 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import layers_123821KB as layers
+
+
+class BaseASPPNet(nn.Module):
+    def __init__(self, nin, ch, dilations=(4, 8, 16)):
+        super(BaseASPPNet, self).__init__()
+        self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
+        self.enc2 = layers.Encoder(ch, ch * 2, 3, 2, 1)
+        self.enc3 = layers.Encoder(ch * 2, ch * 4, 3, 2, 1)
+        self.enc4 = layers.Encoder(ch * 4, ch * 8, 3, 2, 1)
+
+        self.aspp = layers.ASPPModule(ch * 8, ch * 16, dilations)
+
+        self.dec4 = layers.Decoder(ch * (8 + 16), ch * 8, 3, 1, 1)
+        self.dec3 = layers.Decoder(ch * (4 + 8), ch * 4, 3, 1, 1)
+        self.dec2 = layers.Decoder(ch * (2 + 4), ch * 2, 3, 1, 1)
+        self.dec1 = layers.Decoder(ch * (1 + 2), ch, 3, 1, 1)
+
+    def __call__(self, x):
+        h, e1 = self.enc1(x)
+        h, e2 = self.enc2(h)
+        h, e3 = self.enc3(h)
+        h, e4 = self.enc4(h)
+
+        h = self.aspp(h)
+
+        h = self.dec4(h, e4)
+        h = self.dec3(h, e3)
+        h = self.dec2(h, e2)
+        h = self.dec1(h, e1)
+
+        return h
+
+
+class CascadedASPPNet(nn.Module):
+    def __init__(self, n_fft):
+        super(CascadedASPPNet, self).__init__()
+        self.stg1_low_band_net = BaseASPPNet(2, 32)
+        self.stg1_high_band_net = BaseASPPNet(2, 32)
+
+        self.stg2_bridge = layers.Conv2DBNActiv(34, 16, 1, 1, 0)
+        self.stg2_full_band_net = BaseASPPNet(16, 32)
+
+        self.stg3_bridge = layers.Conv2DBNActiv(66, 32, 1, 1, 0)
+        self.stg3_full_band_net = BaseASPPNet(32, 64)
+
+        self.out = nn.Conv2d(64, 2, 1, bias=False)
+        self.aux1_out = nn.Conv2d(32, 2, 1, bias=False)
+        self.aux2_out = nn.Conv2d(32, 2, 1, bias=False)
+
+        self.max_bin = n_fft // 2
+        self.output_bin = n_fft // 2 + 1
+
+        self.offset = 128
+
+    def forward(self, x, aggressiveness=None):
+        mix = x.detach()
+        x = x.clone()
+
+        x = x[:, :, : self.max_bin]
+
+        bandw = x.size()[2] // 2
+        aux1 = torch.cat(
+            [
+                self.stg1_low_band_net(x[:, :, :bandw]),
+                self.stg1_high_band_net(x[:, :, bandw:]),
+            ],
+            dim=2,
+        )
+
+        h = torch.cat([x, aux1], dim=1)
+        aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
+
+        h = torch.cat([x, aux1, aux2], dim=1)
+        h = self.stg3_full_band_net(self.stg3_bridge(h))
+
+        mask = torch.sigmoid(self.out(h))
+        mask = F.pad(
+            input=mask,
+            pad=(0, 0, 0, self.output_bin - mask.size()[2]),
+            mode="replicate",
+        )
+
+        if self.training:
+            aux1 = torch.sigmoid(self.aux1_out(aux1))
+            aux1 = F.pad(
+                input=aux1,
+                pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
+                mode="replicate",
+            )
+            aux2 = torch.sigmoid(self.aux2_out(aux2))
+            aux2 = F.pad(
+                input=aux2,
+                pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
+                mode="replicate",
+            )
+            return mask * mix, aux1 * mix, aux2 * mix
+        else:
+            if aggressiveness:
+                mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
+                    mask[:, :, : aggressiveness["split_bin"]],
+                    1 + aggressiveness["value"] / 3,
+                )
+                mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
+                    mask[:, :, aggressiveness["split_bin"] :],
+                    1 + aggressiveness["value"],
+                )
+
+            return mask * mix
+
+    def predict(self, x_mag, aggressiveness=None):
+        h = self.forward(x_mag, aggressiveness)
+
+        if self.offset > 0:
+            h = h[:, :, :, self.offset : -self.offset]
+            assert h.size()[3] > 0
+
+        return h

infer/lib/uvr5_pack/lib_v5/nets_33966KB.py

@@ -0,0 +1,122 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import layers_33966KB as layers
+
+
+class BaseASPPNet(nn.Module):
+    def __init__(self, nin, ch, dilations=(4, 8, 16, 32)):
+        super(BaseASPPNet, self).__init__()
+        self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
+        self.enc2 = layers.Encoder(ch, ch * 2, 3, 2, 1)
+        self.enc3 = layers.Encoder(ch * 2, ch * 4, 3, 2, 1)
+        self.enc4 = layers.Encoder(ch * 4, ch * 8, 3, 2, 1)
+
+        self.aspp = layers.ASPPModule(ch * 8, ch * 16, dilations)
+
+        self.dec4 = layers.Decoder(ch * (8 + 16), ch * 8, 3, 1, 1)
+        self.dec3 = layers.Decoder(ch * (4 + 8), ch * 4, 3, 1, 1)
+        self.dec2 = layers.Decoder(ch * (2 + 4), ch * 2, 3, 1, 1)
+        self.dec1 = layers.Decoder(ch * (1 + 2), ch, 3, 1, 1)
+
+    def __call__(self, x):
+        h, e1 = self.enc1(x)
+        h, e2 = self.enc2(h)
+        h, e3 = self.enc3(h)
+        h, e4 = self.enc4(h)
+
+        h = self.aspp(h)
+
+        h = self.dec4(h, e4)
+        h = self.dec3(h, e3)
+        h = self.dec2(h, e2)
+        h = self.dec1(h, e1)
+
+        return h
+
+
+class CascadedASPPNet(nn.Module):
+    def __init__(self, n_fft):
+        super(CascadedASPPNet, self).__init__()
+        self.stg1_low_band_net = BaseASPPNet(2, 16)
+        self.stg1_high_band_net = BaseASPPNet(2, 16)
+
+        self.stg2_bridge = layers.Conv2DBNActiv(18, 8, 1, 1, 0)
+        self.stg2_full_band_net = BaseASPPNet(8, 16)
+
+        self.stg3_bridge = layers.Conv2DBNActiv(34, 16, 1, 1, 0)
+        self.stg3_full_band_net = BaseASPPNet(16, 32)
+
+        self.out = nn.Conv2d(32, 2, 1, bias=False)
+        self.aux1_out = nn.Conv2d(16, 2, 1, bias=False)
+        self.aux2_out = nn.Conv2d(16, 2, 1, bias=False)
+
+        self.max_bin = n_fft // 2
+        self.output_bin = n_fft // 2 + 1
+
+        self.offset = 128
+
+    def forward(self, x, aggressiveness=None):
+        mix = x.detach()
+        x = x.clone()
+
+        x = x[:, :, : self.max_bin]
+
+        bandw = x.size()[2] // 2
+        aux1 = torch.cat(
+            [
+                self.stg1_low_band_net(x[:, :, :bandw]),
+                self.stg1_high_band_net(x[:, :, bandw:]),
+            ],
+            dim=2,
+        )
+
+        h = torch.cat([x, aux1], dim=1)
+        aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
+
+        h = torch.cat([x, aux1, aux2], dim=1)
+        h = self.stg3_full_band_net(self.stg3_bridge(h))
+
+        mask = torch.sigmoid(self.out(h))
+        mask = F.pad(
+            input=mask,
+            pad=(0, 0, 0, self.output_bin - mask.size()[2]),
+            mode="replicate",
+        )
+
+        if self.training:
+            aux1 = torch.sigmoid(self.aux1_out(aux1))
+            aux1 = F.pad(
+                input=aux1,
+                pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
+                mode="replicate",
+            )
+            aux2 = torch.sigmoid(self.aux2_out(aux2))
+            aux2 = F.pad(
+                input=aux2,
+                pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
+                mode="replicate",
+            )
+            return mask * mix, aux1 * mix, aux2 * mix
+        else:
+            if aggressiveness:
+                mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
+                    mask[:, :, : aggressiveness["split_bin"]],
+                    1 + aggressiveness["value"] / 3,
+                )
+                mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
+                    mask[:, :, aggressiveness["split_bin"] :],
+                    1 + aggressiveness["value"],
+                )
+
+            return mask * mix
+
+    def predict(self, x_mag, aggressiveness=None):
+        h = self.forward(x_mag, aggressiveness)
+
+        if self.offset > 0:
+            h = h[:, :, :, self.offset : -self.offset]
+            assert h.size()[3] > 0
+
+        return h

infer/lib/uvr5_pack/lib_v5/nets_537227KB.py

@@ -0,0 +1,123 @@
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import layers_537238KB as layers
+
+
+class BaseASPPNet(nn.Module):
+    def __init__(self, nin, ch, dilations=(4, 8, 16)):
+        super(BaseASPPNet, self).__init__()
+        self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
+        self.enc2 = layers.Encoder(ch, ch * 2, 3, 2, 1)
+        self.enc3 = layers.Encoder(ch * 2, ch * 4, 3, 2, 1)
+        self.enc4 = layers.Encoder(ch * 4, ch * 8, 3, 2, 1)
+
+        self.aspp = layers.ASPPModule(ch * 8, ch * 16, dilations)
+
+        self.dec4 = layers.Decoder(ch * (8 + 16), ch * 8, 3, 1, 1)
+        self.dec3 = layers.Decoder(ch * (4 + 8), ch * 4, 3, 1, 1)
+        self.dec2 = layers.Decoder(ch * (2 + 4), ch * 2, 3, 1, 1)
+        self.dec1 = layers.Decoder(ch * (1 + 2), ch, 3, 1, 1)
+
+    def __call__(self, x):
+        h, e1 = self.enc1(x)
+        h, e2 = self.enc2(h)
+        h, e3 = self.enc3(h)
+        h, e4 = self.enc4(h)
+
+        h = self.aspp(h)
+
+        h = self.dec4(h, e4)
+        h = self.dec3(h, e3)
+        h = self.dec2(h, e2)
+        h = self.dec1(h, e1)
+
+        return h
+
+
+class CascadedASPPNet(nn.Module):
+    def __init__(self, n_fft):
+        super(CascadedASPPNet, self).__init__()
+        self.stg1_low_band_net = BaseASPPNet(2, 64)
+        self.stg1_high_band_net = BaseASPPNet(2, 64)
+
+        self.stg2_bridge = layers.Conv2DBNActiv(66, 32, 1, 1, 0)
+        self.stg2_full_band_net = BaseASPPNet(32, 64)
+
+        self.stg3_bridge = layers.Conv2DBNActiv(130, 64, 1, 1, 0)
+        self.stg3_full_band_net = BaseASPPNet(64, 128)
+
+        self.out = nn.Conv2d(128, 2, 1, bias=False)
+        self.aux1_out = nn.Conv2d(64, 2, 1, bias=False)
+        self.aux2_out = nn.Conv2d(64, 2, 1, bias=False)
+
+        self.max_bin = n_fft // 2
+        self.output_bin = n_fft // 2 + 1
+
+        self.offset = 128
+
+    def forward(self, x, aggressiveness=None):
+        mix = x.detach()
+        x = x.clone()
+
+        x = x[:, :, : self.max_bin]
+
+        bandw = x.size()[2] // 2
+        aux1 = torch.cat(
+            [
+                self.stg1_low_band_net(x[:, :, :bandw]),
+                self.stg1_high_band_net(x[:, :, bandw:]),
+            ],
+            dim=2,
+        )
+
+        h = torch.cat([x, aux1], dim=1)
+        aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
+
+        h = torch.cat([x, aux1, aux2], dim=1)
+        h = self.stg3_full_band_net(self.stg3_bridge(h))
+
+        mask = torch.sigmoid(self.out(h))
+        mask = F.pad(
+            input=mask,
+            pad=(0, 0, 0, self.output_bin - mask.size()[2]),
+            mode="replicate",
+        )
+
+        if self.training:
+            aux1 = torch.sigmoid(self.aux1_out(aux1))
+            aux1 = F.pad(
+                input=aux1,
+                pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
+                mode="replicate",
+            )
+            aux2 = torch.sigmoid(self.aux2_out(aux2))
+            aux2 = F.pad(
+                input=aux2,
+                pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
+                mode="replicate",
+            )
+            return mask * mix, aux1 * mix, aux2 * mix
+        else:
+            if aggressiveness:
+                mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
+                    mask[:, :, : aggressiveness["split_bin"]],
+                    1 + aggressiveness["value"] / 3,
+                )
+                mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
+                    mask[:, :, aggressiveness["split_bin"] :],
+                    1 + aggressiveness["value"],
+                )
+
+            return mask * mix
+
+    def predict(self, x_mag, aggressiveness=None):
+        h = self.forward(x_mag, aggressiveness)
+
+        if self.offset > 0:
+            h = h[:, :, :, self.offset : -self.offset]
+            assert h.size()[3] > 0
+
+        return h

infer/lib/uvr5_pack/lib_v5/nets_537238KB.py

@@ -0,0 +1,123 @@
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import layers_537238KB as layers
+
+
+class BaseASPPNet(nn.Module):
+    def __init__(self, nin, ch, dilations=(4, 8, 16)):
+        super(BaseASPPNet, self).__init__()
+        self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
+        self.enc2 = layers.Encoder(ch, ch * 2, 3, 2, 1)
+        self.enc3 = layers.Encoder(ch * 2, ch * 4, 3, 2, 1)
+        self.enc4 = layers.Encoder(ch * 4, ch * 8, 3, 2, 1)
+
+        self.aspp = layers.ASPPModule(ch * 8, ch * 16, dilations)
+
+        self.dec4 = layers.Decoder(ch * (8 + 16), ch * 8, 3, 1, 1)
+        self.dec3 = layers.Decoder(ch * (4 + 8), ch * 4, 3, 1, 1)
+        self.dec2 = layers.Decoder(ch * (2 + 4), ch * 2, 3, 1, 1)
+        self.dec1 = layers.Decoder(ch * (1 + 2), ch, 3, 1, 1)
+
+    def __call__(self, x):
+        h, e1 = self.enc1(x)
+        h, e2 = self.enc2(h)
+        h, e3 = self.enc3(h)
+        h, e4 = self.enc4(h)
+
+        h = self.aspp(h)
+
+        h = self.dec4(h, e4)
+        h = self.dec3(h, e3)
+        h = self.dec2(h, e2)
+        h = self.dec1(h, e1)
+
+        return h
+
+
+class CascadedASPPNet(nn.Module):
+    def __init__(self, n_fft):
+        super(CascadedASPPNet, self).__init__()
+        self.stg1_low_band_net = BaseASPPNet(2, 64)
+        self.stg1_high_band_net = BaseASPPNet(2, 64)
+
+        self.stg2_bridge = layers.Conv2DBNActiv(66, 32, 1, 1, 0)
+        self.stg2_full_band_net = BaseASPPNet(32, 64)
+
+        self.stg3_bridge = layers.Conv2DBNActiv(130, 64, 1, 1, 0)
+        self.stg3_full_band_net = BaseASPPNet(64, 128)
+
+        self.out = nn.Conv2d(128, 2, 1, bias=False)
+        self.aux1_out = nn.Conv2d(64, 2, 1, bias=False)
+        self.aux2_out = nn.Conv2d(64, 2, 1, bias=False)
+
+        self.max_bin = n_fft // 2
+        self.output_bin = n_fft // 2 + 1
+
+        self.offset = 128
+
+    def forward(self, x, aggressiveness=None):
+        mix = x.detach()
+        x = x.clone()
+
+        x = x[:, :, : self.max_bin]
+
+        bandw = x.size()[2] // 2
+        aux1 = torch.cat(
+            [
+                self.stg1_low_band_net(x[:, :, :bandw]),
+                self.stg1_high_band_net(x[:, :, bandw:]),
+            ],
+            dim=2,
+        )
+
+        h = torch.cat([x, aux1], dim=1)
+        aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
+
+        h = torch.cat([x, aux1, aux2], dim=1)
+        h = self.stg3_full_band_net(self.stg3_bridge(h))
+
+        mask = torch.sigmoid(self.out(h))
+        mask = F.pad(
+            input=mask,
+            pad=(0, 0, 0, self.output_bin - mask.size()[2]),
+            mode="replicate",
+        )
+
+        if self.training:
+            aux1 = torch.sigmoid(self.aux1_out(aux1))
+            aux1 = F.pad(
+                input=aux1,
+                pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
+                mode="replicate",
+            )
+            aux2 = torch.sigmoid(self.aux2_out(aux2))
+            aux2 = F.pad(
+                input=aux2,
+                pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
+                mode="replicate",
+            )
+            return mask * mix, aux1 * mix, aux2 * mix
+        else:
+            if aggressiveness:
+                mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
+                    mask[:, :, : aggressiveness["split_bin"]],
+                    1 + aggressiveness["value"] / 3,
+                )
+                mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
+                    mask[:, :, aggressiveness["split_bin"] :],
+                    1 + aggressiveness["value"],
+                )
+
+            return mask * mix
+
+    def predict(self, x_mag, aggressiveness=None):
+        h = self.forward(x_mag, aggressiveness)
+
+        if self.offset > 0:
+            h = h[:, :, :, self.offset : -self.offset]
+            assert h.size()[3] > 0
+
+        return h

infer/lib/uvr5_pack/lib_v5/nets_61968KB.py

@@ -0,0 +1,122 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import layers_123821KB as layers
+
+
+class BaseASPPNet(nn.Module):
+    def __init__(self, nin, ch, dilations=(4, 8, 16)):
+        super(BaseASPPNet, self).__init__()
+        self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
+        self.enc2 = layers.Encoder(ch, ch * 2, 3, 2, 1)
+        self.enc3 = layers.Encoder(ch * 2, ch * 4, 3, 2, 1)
+        self.enc4 = layers.Encoder(ch * 4, ch * 8, 3, 2, 1)
+
+        self.aspp = layers.ASPPModule(ch * 8, ch * 16, dilations)
+
+        self.dec4 = layers.Decoder(ch * (8 + 16), ch * 8, 3, 1, 1)
+        self.dec3 = layers.Decoder(ch * (4 + 8), ch * 4, 3, 1, 1)
+        self.dec2 = layers.Decoder(ch * (2 + 4), ch * 2, 3, 1, 1)
+        self.dec1 = layers.Decoder(ch * (1 + 2), ch, 3, 1, 1)
+
+    def __call__(self, x):
+        h, e1 = self.enc1(x)
+        h, e2 = self.enc2(h)
+        h, e3 = self.enc3(h)
+        h, e4 = self.enc4(h)
+
+        h = self.aspp(h)
+
+        h = self.dec4(h, e4)
+        h = self.dec3(h, e3)
+        h = self.dec2(h, e2)
+        h = self.dec1(h, e1)
+
+        return h
+
+
+class CascadedASPPNet(nn.Module):
+    def __init__(self, n_fft):
+        super(CascadedASPPNet, self).__init__()
+        self.stg1_low_band_net = BaseASPPNet(2, 32)
+        self.stg1_high_band_net = BaseASPPNet(2, 32)
+
+        self.stg2_bridge = layers.Conv2DBNActiv(34, 16, 1, 1, 0)
+        self.stg2_full_band_net = BaseASPPNet(16, 32)
+
+        self.stg3_bridge = layers.Conv2DBNActiv(66, 32, 1, 1, 0)
+        self.stg3_full_band_net = BaseASPPNet(32, 64)
+
+        self.out = nn.Conv2d(64, 2, 1, bias=False)
+        self.aux1_out = nn.Conv2d(32, 2, 1, bias=False)
+        self.aux2_out = nn.Conv2d(32, 2, 1, bias=False)
+
+        self.max_bin = n_fft // 2
+        self.output_bin = n_fft // 2 + 1
+
+        self.offset = 128
+
+    def forward(self, x, aggressiveness=None):
+        mix = x.detach()
+        x = x.clone()
+
+        x = x[:, :, : self.max_bin]
+
+        bandw = x.size()[2] // 2
+        aux1 = torch.cat(
+            [
+                self.stg1_low_band_net(x[:, :, :bandw]),
+                self.stg1_high_band_net(x[:, :, bandw:]),
+            ],
+            dim=2,
+        )
+
+        h = torch.cat([x, aux1], dim=1)
+        aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
+
+        h = torch.cat([x, aux1, aux2], dim=1)
+        h = self.stg3_full_band_net(self.stg3_bridge(h))
+
+        mask = torch.sigmoid(self.out(h))
+        mask = F.pad(
+            input=mask,
+            pad=(0, 0, 0, self.output_bin - mask.size()[2]),
+            mode="replicate",
+        )
+
+        if self.training:
+            aux1 = torch.sigmoid(self.aux1_out(aux1))
+            aux1 = F.pad(
+                input=aux1,
+                pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
+                mode="replicate",
+            )
+            aux2 = torch.sigmoid(self.aux2_out(aux2))
+            aux2 = F.pad(
+                input=aux2,
+                pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
+                mode="replicate",
+            )
+            return mask * mix, aux1 * mix, aux2 * mix
+        else:
+            if aggressiveness:
+                mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
+                    mask[:, :, : aggressiveness["split_bin"]],
+                    1 + aggressiveness["value"] / 3,
+                )
+                mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
+                    mask[:, :, aggressiveness["split_bin"] :],
+                    1 + aggressiveness["value"],
+                )
+
+            return mask * mix
+
+    def predict(self, x_mag, aggressiveness=None):
+        h = self.forward(x_mag, aggressiveness)
+
+        if self.offset > 0:
+            h = h[:, :, :, self.offset : -self.offset]
+            assert h.size()[3] > 0
+
+        return h

infer/lib/uvr5_pack/lib_v5/nets_new.py

@@ -0,0 +1,133 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import layers_new
+
+
+class BaseNet(nn.Module):
+    def __init__(
+        self, nin, nout, nin_lstm, nout_lstm, dilations=((4, 2), (8, 4), (12, 6))
+    ):
+        super(BaseNet, self).__init__()
+        self.enc1 = layers_new.Conv2DBNActiv(nin, nout, 3, 1, 1)
+        self.enc2 = layers_new.Encoder(nout, nout * 2, 3, 2, 1)
+        self.enc3 = layers_new.Encoder(nout * 2, nout * 4, 3, 2, 1)
+        self.enc4 = layers_new.Encoder(nout * 4, nout * 6, 3, 2, 1)
+        self.enc5 = layers_new.Encoder(nout * 6, nout * 8, 3, 2, 1)
+
+        self.aspp = layers_new.ASPPModule(nout * 8, nout * 8, dilations, dropout=True)
+
+        self.dec4 = layers_new.Decoder(nout * (6 + 8), nout * 6, 3, 1, 1)
+        self.dec3 = layers_new.Decoder(nout * (4 + 6), nout * 4, 3, 1, 1)
+        self.dec2 = layers_new.Decoder(nout * (2 + 4), nout * 2, 3, 1, 1)
+        self.lstm_dec2 = layers_new.LSTMModule(nout * 2, nin_lstm, nout_lstm)
+        self.dec1 = layers_new.Decoder(nout * (1 + 2) + 1, nout * 1, 3, 1, 1)
+
+    def __call__(self, x):
+        e1 = self.enc1(x)
+        e2 = self.enc2(e1)
+        e3 = self.enc3(e2)
+        e4 = self.enc4(e3)
+        e5 = self.enc5(e4)
+
+        h = self.aspp(e5)
+
+        h = self.dec4(h, e4)
+        h = self.dec3(h, e3)
+        h = self.dec2(h, e2)
+        h = torch.cat([h, self.lstm_dec2(h)], dim=1)
+        h = self.dec1(h, e1)
+
+        return h
+
+
+class CascadedNet(nn.Module):
+    def __init__(self, n_fft, nout=32, nout_lstm=128):
+        super(CascadedNet, self).__init__()
+
+        self.max_bin = n_fft // 2
+        self.output_bin = n_fft // 2 + 1
+        self.nin_lstm = self.max_bin // 2
+        self.offset = 64
+
+        self.stg1_low_band_net = nn.Sequential(
+            BaseNet(2, nout // 2, self.nin_lstm // 2, nout_lstm),
+            layers_new.Conv2DBNActiv(nout // 2, nout // 4, 1, 1, 0),
+        )
+
+        self.stg1_high_band_net = BaseNet(
+            2, nout // 4, self.nin_lstm // 2, nout_lstm // 2
+        )
+
+        self.stg2_low_band_net = nn.Sequential(
+            BaseNet(nout // 4 + 2, nout, self.nin_lstm // 2, nout_lstm),
+            layers_new.Conv2DBNActiv(nout, nout // 2, 1, 1, 0),
+        )
+        self.stg2_high_band_net = BaseNet(
+            nout // 4 + 2, nout // 2, self.nin_lstm // 2, nout_lstm // 2
+        )
+
+        self.stg3_full_band_net = BaseNet(
+            3 * nout // 4 + 2, nout, self.nin_lstm, nout_lstm
+        )
+
+        self.out = nn.Conv2d(nout, 2, 1, bias=False)
+        self.aux_out = nn.Conv2d(3 * nout // 4, 2, 1, bias=False)
+
+    def forward(self, x):
+        x = x[:, :, : self.max_bin]
+
+        bandw = x.size()[2] // 2
+        l1_in = x[:, :, :bandw]
+        h1_in = x[:, :, bandw:]
+        l1 = self.stg1_low_band_net(l1_in)
+        h1 = self.stg1_high_band_net(h1_in)
+        aux1 = torch.cat([l1, h1], dim=2)
+
+        l2_in = torch.cat([l1_in, l1], dim=1)
+        h2_in = torch.cat([h1_in, h1], dim=1)
+        l2 = self.stg2_low_band_net(l2_in)
+        h2 = self.stg2_high_band_net(h2_in)
+        aux2 = torch.cat([l2, h2], dim=2)
+
+        f3_in = torch.cat([x, aux1, aux2], dim=1)
+        f3 = self.stg3_full_band_net(f3_in)
+
+        mask = torch.sigmoid(self.out(f3))
+        mask = F.pad(
+            input=mask,
+            pad=(0, 0, 0, self.output_bin - mask.size()[2]),
+            mode="replicate",
+        )
+
+        if self.training:
+            aux = torch.cat([aux1, aux2], dim=1)
+            aux = torch.sigmoid(self.aux_out(aux))
+            aux = F.pad(
+                input=aux,
+                pad=(0, 0, 0, self.output_bin - aux.size()[2]),
+                mode="replicate",
+            )
+            return mask, aux
+        else:
+            return mask
+
+    def predict_mask(self, x):
+        mask = self.forward(x)
+
+        if self.offset > 0:
+            mask = mask[:, :, :, self.offset : -self.offset]
+            assert mask.size()[3] > 0
+
+        return mask
+
+    def predict(self, x, aggressiveness=None):
+        mask = self.forward(x)
+        pred_mag = x * mask
+
+        if self.offset > 0:
+            pred_mag = pred_mag[:, :, :, self.offset : -self.offset]
+            assert pred_mag.size()[3] > 0
+
+        return pred_mag

infer/lib/uvr5_pack/lib_v5/spec_utils.py

@@ -0,0 +1,672 @@
+import hashlib
+import json
+import math
+import os
+
+import librosa
+import numpy as np
+import soundfile as sf
+from tqdm import tqdm
+
+
+def crop_center(h1, h2):
+    h1_shape = h1.size()
+    h2_shape = h2.size()
+
+    if h1_shape[3] == h2_shape[3]:
+        return h1
+    elif h1_shape[3] < h2_shape[3]:
+        raise ValueError("h1_shape[3] must be greater than h2_shape[3]")
+
+    # s_freq = (h2_shape[2] - h1_shape[2]) // 2
+    # e_freq = s_freq + h1_shape[2]
+    s_time = (h1_shape[3] - h2_shape[3]) // 2
+    e_time = s_time + h2_shape[3]
+    h1 = h1[:, :, :, s_time:e_time]
+
+    return h1
+
+
+def wave_to_spectrogram(
+    wave, hop_length, n_fft, mid_side=False, mid_side_b2=False, reverse=False
+):
+    if reverse:
+        wave_left = np.flip(np.asfortranarray(wave[0]))
+        wave_right = np.flip(np.asfortranarray(wave[1]))
+    elif mid_side:
+        wave_left = np.asfortranarray(np.add(wave[0], wave[1]) / 2)
+        wave_right = np.asfortranarray(np.subtract(wave[0], wave[1]))
+    elif mid_side_b2:
+        wave_left = np.asfortranarray(np.add(wave[1], wave[0] * 0.5))
+        wave_right = np.asfortranarray(np.subtract(wave[0], wave[1] * 0.5))
+    else:
+        wave_left = np.asfortranarray(wave[0])
+        wave_right = np.asfortranarray(wave[1])
+
+    spec_left = librosa.stft(wave_left, n_fft, hop_length=hop_length)
+    spec_right = librosa.stft(wave_right, n_fft, hop_length=hop_length)
+
+    spec = np.asfortranarray([spec_left, spec_right])
+
+    return spec
+
+
+def wave_to_spectrogram_mt(
+    wave, hop_length, n_fft, mid_side=False, mid_side_b2=False, reverse=False
+):
+    import threading
+
+    if reverse:
+        wave_left = np.flip(np.asfortranarray(wave[0]))
+        wave_right = np.flip(np.asfortranarray(wave[1]))
+    elif mid_side:
+        wave_left = np.asfortranarray(np.add(wave[0], wave[1]) / 2)
+        wave_right = np.asfortranarray(np.subtract(wave[0], wave[1]))
+    elif mid_side_b2:
+        wave_left = np.asfortranarray(np.add(wave[1], wave[0] * 0.5))
+        wave_right = np.asfortranarray(np.subtract(wave[0], wave[1] * 0.5))
+    else:
+        wave_left = np.asfortranarray(wave[0])
+        wave_right = np.asfortranarray(wave[1])
+
+    def run_thread(**kwargs):
+        global spec_left
+        spec_left = librosa.stft(**kwargs)
+
+    thread = threading.Thread(
+        target=run_thread,
+        kwargs={"y": wave_left, "n_fft": n_fft, "hop_length": hop_length},
+    )
+    thread.start()
+    spec_right = librosa.stft(wave_right, n_fft, hop_length=hop_length)
+    thread.join()
+
+    spec = np.asfortranarray([spec_left, spec_right])
+
+    return spec
+
+
+def combine_spectrograms(specs, mp):
+    l = min([specs[i].shape[2] for i in specs])
+    spec_c = np.zeros(shape=(2, mp.param["bins"] + 1, l), dtype=np.complex64)
+    offset = 0
+    bands_n = len(mp.param["band"])
+
+    for d in range(1, bands_n + 1):
+        h = mp.param["band"][d]["crop_stop"] - mp.param["band"][d]["crop_start"]
+        spec_c[:, offset : offset + h, :l] = specs[d][
+            :, mp.param["band"][d]["crop_start"] : mp.param["band"][d]["crop_stop"], :l
+        ]
+        offset += h
+
+    if offset > mp.param["bins"]:
+        raise ValueError("Too much bins")
+
+    # lowpass fiter
+    if (
+        mp.param["pre_filter_start"] > 0
+    ):  # and mp.param['band'][bands_n]['res_type'] in ['scipy', 'polyphase']:
+        if bands_n == 1:
+            spec_c = fft_lp_filter(
+                spec_c, mp.param["pre_filter_start"], mp.param["pre_filter_stop"]
+            )
+        else:
+            gp = 1
+            for b in range(
+                mp.param["pre_filter_start"] + 1, mp.param["pre_filter_stop"]
+            ):
+                g = math.pow(
+                    10, -(b - mp.param["pre_filter_start"]) * (3.5 - gp) / 20.0
+                )
+                gp = g
+                spec_c[:, b, :] *= g
+
+    return np.asfortranarray(spec_c)
+
+
+def spectrogram_to_image(spec, mode="magnitude"):
+    if mode == "magnitude":
+        if np.iscomplexobj(spec):
+            y = np.abs(spec)
+        else:
+            y = spec
+        y = np.log10(y**2 + 1e-8)
+    elif mode == "phase":
+        if np.iscomplexobj(spec):
+            y = np.angle(spec)
+        else:
+            y = spec
+
+    y -= y.min()
+    y *= 255 / y.max()
+    img = np.uint8(y)
+
+    if y.ndim == 3:
+        img = img.transpose(1, 2, 0)
+        img = np.concatenate([np.max(img, axis=2, keepdims=True), img], axis=2)
+
+    return img
+
+
+def reduce_vocal_aggressively(X, y, softmask):
+    v = X - y
+    y_mag_tmp = np.abs(y)
+    v_mag_tmp = np.abs(v)
+
+    v_mask = v_mag_tmp > y_mag_tmp
+    y_mag = np.clip(y_mag_tmp - v_mag_tmp * v_mask * softmask, 0, np.inf)
+
+    return y_mag * np.exp(1.0j * np.angle(y))
+
+
+def mask_silence(mag, ref, thres=0.2, min_range=64, fade_size=32):
+    if min_range < fade_size * 2:
+        raise ValueError("min_range must be >= fade_area * 2")
+
+    mag = mag.copy()
+
+    idx = np.where(ref.mean(axis=(0, 1)) < thres)[0]
+    starts = np.insert(idx[np.where(np.diff(idx) != 1)[0] + 1], 0, idx[0])
+    ends = np.append(idx[np.where(np.diff(idx) != 1)[0]], idx[-1])
+    uninformative = np.where(ends - starts > min_range)[0]
+    if len(uninformative) > 0:
+        starts = starts[uninformative]
+        ends = ends[uninformative]
+        old_e = None
+        for s, e in zip(starts, ends):
+            if old_e is not None and s - old_e < fade_size:
+                s = old_e - fade_size * 2
+
+            if s != 0:
+                weight = np.linspace(0, 1, fade_size)
+                mag[:, :, s : s + fade_size] += weight * ref[:, :, s : s + fade_size]
+            else:
+                s -= fade_size
+
+            if e != mag.shape[2]:
+                weight = np.linspace(1, 0, fade_size)
+                mag[:, :, e - fade_size : e] += weight * ref[:, :, e - fade_size : e]
+            else:
+                e += fade_size
+
+            mag[:, :, s + fade_size : e - fade_size] += ref[
+                :, :, s + fade_size : e - fade_size
+            ]
+            old_e = e
+
+    return mag
+
+
+def align_wave_head_and_tail(a, b):
+    l = min([a[0].size, b[0].size])
+
+    return a[:l, :l], b[:l, :l]
+
+
+def cache_or_load(mix_path, inst_path, mp):
+    mix_basename = os.path.splitext(os.path.basename(mix_path))[0]
+    inst_basename = os.path.splitext(os.path.basename(inst_path))[0]
+
+    cache_dir = "mph{}".format(
+        hashlib.sha1(json.dumps(mp.param, sort_keys=True).encode("utf-8")).hexdigest()
+    )
+    mix_cache_dir = os.path.join("cache", cache_dir)
+    inst_cache_dir = os.path.join("cache", cache_dir)
+
+    os.makedirs(mix_cache_dir, exist_ok=True)
+    os.makedirs(inst_cache_dir, exist_ok=True)
+
+    mix_cache_path = os.path.join(mix_cache_dir, mix_basename + ".npy")
+    inst_cache_path = os.path.join(inst_cache_dir, inst_basename + ".npy")
+
+    if os.path.exists(mix_cache_path) and os.path.exists(inst_cache_path):
+        X_spec_m = np.load(mix_cache_path)
+        y_spec_m = np.load(inst_cache_path)
+    else:
+        X_wave, y_wave, X_spec_s, y_spec_s = {}, {}, {}, {}
+
+        for d in range(len(mp.param["band"]), 0, -1):
+            bp = mp.param["band"][d]
+
+            if d == len(mp.param["band"]):  # high-end band
+                X_wave[d], _ = librosa.load(
+                    mix_path, bp["sr"], False, dtype=np.float32, res_type=bp["res_type"]
+                )
+                y_wave[d], _ = librosa.load(
+                    inst_path,
+                    bp["sr"],
+                    False,
+                    dtype=np.float32,
+                    res_type=bp["res_type"],
+                )
+            else:  # lower bands
+                X_wave[d] = librosa.resample(
+                    X_wave[d + 1],
+                    mp.param["band"][d + 1]["sr"],
+                    bp["sr"],
+                    res_type=bp["res_type"],
+                )
+                y_wave[d] = librosa.resample(
+                    y_wave[d + 1],
+                    mp.param["band"][d + 1]["sr"],
+                    bp["sr"],
+                    res_type=bp["res_type"],
+                )
+
+            X_wave[d], y_wave[d] = align_wave_head_and_tail(X_wave[d], y_wave[d])
+
+            X_spec_s[d] = wave_to_spectrogram(
+                X_wave[d],
+                bp["hl"],
+                bp["n_fft"],
+                mp.param["mid_side"],
+                mp.param["mid_side_b2"],
+                mp.param["reverse"],
+            )
+            y_spec_s[d] = wave_to_spectrogram(
+                y_wave[d],
+                bp["hl"],
+                bp["n_fft"],
+                mp.param["mid_side"],
+                mp.param["mid_side_b2"],
+                mp.param["reverse"],
+            )
+
+        del X_wave, y_wave
+
+        X_spec_m = combine_spectrograms(X_spec_s, mp)
+        y_spec_m = combine_spectrograms(y_spec_s, mp)
+
+        if X_spec_m.shape != y_spec_m.shape:
+            raise ValueError("The combined spectrograms are different: " + mix_path)
+
+        _, ext = os.path.splitext(mix_path)
+
+        np.save(mix_cache_path, X_spec_m)
+        np.save(inst_cache_path, y_spec_m)
+
+    return X_spec_m, y_spec_m
+
+
+def spectrogram_to_wave(spec, hop_length, mid_side, mid_side_b2, reverse):
+    spec_left = np.asfortranarray(spec[0])
+    spec_right = np.asfortranarray(spec[1])
+
+    wave_left = librosa.istft(spec_left, hop_length=hop_length)
+    wave_right = librosa.istft(spec_right, hop_length=hop_length)
+
+    if reverse:
+        return np.asfortranarray([np.flip(wave_left), np.flip(wave_right)])
+    elif mid_side:
+        return np.asfortranarray(
+            [np.add(wave_left, wave_right / 2), np.subtract(wave_left, wave_right / 2)]
+        )
+    elif mid_side_b2:
+        return np.asfortranarray(
+            [
+                np.add(wave_right / 1.25, 0.4 * wave_left),
+                np.subtract(wave_left / 1.25, 0.4 * wave_right),
+            ]
+        )
+    else:
+        return np.asfortranarray([wave_left, wave_right])
+
+
+def spectrogram_to_wave_mt(spec, hop_length, mid_side, reverse, mid_side_b2):
+    import threading
+
+    spec_left = np.asfortranarray(spec[0])
+    spec_right = np.asfortranarray(spec[1])
+
+    def run_thread(**kwargs):
+        global wave_left
+        wave_left = librosa.istft(**kwargs)
+
+    thread = threading.Thread(
+        target=run_thread, kwargs={"stft_matrix": spec_left, "hop_length": hop_length}
+    )
+    thread.start()
+    wave_right = librosa.istft(spec_right, hop_length=hop_length)
+    thread.join()
+
+    if reverse:
+        return np.asfortranarray([np.flip(wave_left), np.flip(wave_right)])
+    elif mid_side:
+        return np.asfortranarray(
+            [np.add(wave_left, wave_right / 2), np.subtract(wave_left, wave_right / 2)]
+        )
+    elif mid_side_b2:
+        return np.asfortranarray(
+            [
+                np.add(wave_right / 1.25, 0.4 * wave_left),
+                np.subtract(wave_left / 1.25, 0.4 * wave_right),
+            ]
+        )
+    else:
+        return np.asfortranarray([wave_left, wave_right])
+
+
+def cmb_spectrogram_to_wave(spec_m, mp, extra_bins_h=None, extra_bins=None):
+    wave_band = {}
+    bands_n = len(mp.param["band"])
+    offset = 0
+
+    for d in range(1, bands_n + 1):
+        bp = mp.param["band"][d]
+        spec_s = np.ndarray(
+            shape=(2, bp["n_fft"] // 2 + 1, spec_m.shape[2]), dtype=complex
+        )
+        h = bp["crop_stop"] - bp["crop_start"]
+        spec_s[:, bp["crop_start"] : bp["crop_stop"], :] = spec_m[
+            :, offset : offset + h, :
+        ]
+
+        offset += h
+        if d == bands_n:  # higher
+            if extra_bins_h:  # if --high_end_process bypass
+                max_bin = bp["n_fft"] // 2
+                spec_s[:, max_bin - extra_bins_h : max_bin, :] = extra_bins[
+                    :, :extra_bins_h, :
+                ]
+            if bp["hpf_start"] > 0:
+                spec_s = fft_hp_filter(spec_s, bp["hpf_start"], bp["hpf_stop"] - 1)
+            if bands_n == 1:
+                wave = spectrogram_to_wave(
+                    spec_s,
+                    bp["hl"],
+                    mp.param["mid_side"],
+                    mp.param["mid_side_b2"],
+                    mp.param["reverse"],
+                )
+            else:
+                wave = np.add(
+                    wave,
+                    spectrogram_to_wave(
+                        spec_s,
+                        bp["hl"],
+                        mp.param["mid_side"],
+                        mp.param["mid_side_b2"],
+                        mp.param["reverse"],
+                    ),
+                )
+        else:
+            sr = mp.param["band"][d + 1]["sr"]
+            if d == 1:  # lower
+                spec_s = fft_lp_filter(spec_s, bp["lpf_start"], bp["lpf_stop"])
+                wave = librosa.resample(
+                    spectrogram_to_wave(
+                        spec_s,
+                        bp["hl"],
+                        mp.param["mid_side"],
+                        mp.param["mid_side_b2"],
+                        mp.param["reverse"],
+                    ),
+                    bp["sr"],
+                    sr,
+                    res_type="sinc_fastest",
+                )
+            else:  # mid
+                spec_s = fft_hp_filter(spec_s, bp["hpf_start"], bp["hpf_stop"] - 1)
+                spec_s = fft_lp_filter(spec_s, bp["lpf_start"], bp["lpf_stop"])
+                wave2 = np.add(
+                    wave,
+                    spectrogram_to_wave(
+                        spec_s,
+                        bp["hl"],
+                        mp.param["mid_side"],
+                        mp.param["mid_side_b2"],
+                        mp.param["reverse"],
+                    ),
+                )
+                # wave = librosa.core.resample(wave2, bp['sr'], sr, res_type="sinc_fastest")
+                wave = librosa.core.resample(wave2, bp["sr"], sr, res_type="scipy")
+
+    return wave.T
+
+
+def fft_lp_filter(spec, bin_start, bin_stop):
+    g = 1.0
+    for b in range(bin_start, bin_stop):
+        g -= 1 / (bin_stop - bin_start)
+        spec[:, b, :] = g * spec[:, b, :]
+
+    spec[:, bin_stop:, :] *= 0
+
+    return spec
+
+
+def fft_hp_filter(spec, bin_start, bin_stop):
+    g = 1.0
+    for b in range(bin_start, bin_stop, -1):
+        g -= 1 / (bin_start - bin_stop)
+        spec[:, b, :] = g * spec[:, b, :]
+
+    spec[:, 0 : bin_stop + 1, :] *= 0
+
+    return spec
+
+
+def mirroring(a, spec_m, input_high_end, mp):
+    if "mirroring" == a:
+        mirror = np.flip(
+            np.abs(
+                spec_m[
+                    :,
+                    mp.param["pre_filter_start"]
+                    - 10
+                    - input_high_end.shape[1] : mp.param["pre_filter_start"]
+                    - 10,
+                    :,
+                ]
+            ),
+            1,
+        )
+        mirror = mirror * np.exp(1.0j * np.angle(input_high_end))
+
+        return np.where(
+            np.abs(input_high_end) <= np.abs(mirror), input_high_end, mirror
+        )
+
+    if "mirroring2" == a:
+        mirror = np.flip(
+            np.abs(
+                spec_m[
+                    :,
+                    mp.param["pre_filter_start"]
+                    - 10
+                    - input_high_end.shape[1] : mp.param["pre_filter_start"]
+                    - 10,
+                    :,
+                ]
+            ),
+            1,
+        )
+        mi = np.multiply(mirror, input_high_end * 1.7)
+
+        return np.where(np.abs(input_high_end) <= np.abs(mi), input_high_end, mi)
+
+
+def ensembling(a, specs):
+    for i in range(1, len(specs)):
+        if i == 1:
+            spec = specs[0]
+
+        ln = min([spec.shape[2], specs[i].shape[2]])
+        spec = spec[:, :, :ln]
+        specs[i] = specs[i][:, :, :ln]
+
+        if "min_mag" == a:
+            spec = np.where(np.abs(specs[i]) <= np.abs(spec), specs[i], spec)
+        if "max_mag" == a:
+            spec = np.where(np.abs(specs[i]) >= np.abs(spec), specs[i], spec)
+
+    return spec
+
+
+def stft(wave, nfft, hl):
+    wave_left = np.asfortranarray(wave[0])
+    wave_right = np.asfortranarray(wave[1])
+    spec_left = librosa.stft(wave_left, nfft, hop_length=hl)
+    spec_right = librosa.stft(wave_right, nfft, hop_length=hl)
+    spec = np.asfortranarray([spec_left, spec_right])
+
+    return spec
+
+
+def istft(spec, hl):
+    spec_left = np.asfortranarray(spec[0])
+    spec_right = np.asfortranarray(spec[1])
+
+    wave_left = librosa.istft(spec_left, hop_length=hl)
+    wave_right = librosa.istft(spec_right, hop_length=hl)
+    wave = np.asfortranarray([wave_left, wave_right])
+
+
+if __name__ == "__main__":
+    import argparse
+    import sys
+    import time
+
+    import cv2
+    from model_param_init import ModelParameters
+
+    p = argparse.ArgumentParser()
+    p.add_argument(
+        "--algorithm",
+        "-a",
+        type=str,
+        choices=["invert", "invert_p", "min_mag", "max_mag", "deep", "align"],
+        default="min_mag",
+    )
+    p.add_argument(
+        "--model_params",
+        "-m",
+        type=str,
+        default=os.path.join("modelparams", "1band_sr44100_hl512.json"),
+    )
+    p.add_argument("--output_name", "-o", type=str, default="output")
+    p.add_argument("--vocals_only", "-v", action="store_true")
+    p.add_argument("input", nargs="+")
+    args = p.parse_args()
+
+    start_time = time.time()
+
+    if args.algorithm.startswith("invert") and len(args.input) != 2:
+        raise ValueError("There should be two input files.")
+
+    if not args.algorithm.startswith("invert") and len(args.input) < 2:
+        raise ValueError("There must be at least two input files.")
+
+    wave, specs = {}, {}
+    mp = ModelParameters(args.model_params)
+
+    for i in range(len(args.input)):
+        spec = {}
+
+        for d in range(len(mp.param["band"]), 0, -1):
+            bp = mp.param["band"][d]
+
+            if d == len(mp.param["band"]):  # high-end band
+                wave[d], _ = librosa.load(
+                    args.input[i],
+                    bp["sr"],
+                    False,
+                    dtype=np.float32,
+                    res_type=bp["res_type"],
+                )
+
+                if len(wave[d].shape) == 1:  # mono to stereo
+                    wave[d] = np.array([wave[d], wave[d]])
+            else:  # lower bands
+                wave[d] = librosa.resample(
+                    wave[d + 1],
+                    mp.param["band"][d + 1]["sr"],
+                    bp["sr"],
+                    res_type=bp["res_type"],
+                )
+
+            spec[d] = wave_to_spectrogram(
+                wave[d],
+                bp["hl"],
+                bp["n_fft"],
+                mp.param["mid_side"],
+                mp.param["mid_side_b2"],
+                mp.param["reverse"],
+            )
+
+        specs[i] = combine_spectrograms(spec, mp)
+
+    del wave
+
+    if args.algorithm == "deep":
+        d_spec = np.where(np.abs(specs[0]) <= np.abs(spec[1]), specs[0], spec[1])
+        v_spec = d_spec - specs[1]
+        sf.write(
+            os.path.join("{}.wav".format(args.output_name)),
+            cmb_spectrogram_to_wave(v_spec, mp),
+            mp.param["sr"],
+        )
+
+    if args.algorithm.startswith("invert"):
+        ln = min([specs[0].shape[2], specs[1].shape[2]])
+        specs[0] = specs[0][:, :, :ln]
+        specs[1] = specs[1][:, :, :ln]
+
+        if "invert_p" == args.algorithm:
+            X_mag = np.abs(specs[0])
+            y_mag = np.abs(specs[1])
+            max_mag = np.where(X_mag >= y_mag, X_mag, y_mag)
+            v_spec = specs[1] - max_mag * np.exp(1.0j * np.angle(specs[0]))
+        else:
+            specs[1] = reduce_vocal_aggressively(specs[0], specs[1], 0.2)
+            v_spec = specs[0] - specs[1]
+
+            if not args.vocals_only:
+                X_mag = np.abs(specs[0])
+                y_mag = np.abs(specs[1])
+                v_mag = np.abs(v_spec)
+
+                X_image = spectrogram_to_image(X_mag)
+                y_image = spectrogram_to_image(y_mag)
+                v_image = spectrogram_to_image(v_mag)
+
+                cv2.imwrite("{}_X.png".format(args.output_name), X_image)
+                cv2.imwrite("{}_y.png".format(args.output_name), y_image)
+                cv2.imwrite("{}_v.png".format(args.output_name), v_image)
+
+                sf.write(
+                    "{}_X.wav".format(args.output_name),
+                    cmb_spectrogram_to_wave(specs[0], mp),
+                    mp.param["sr"],
+                )
+                sf.write(
+                    "{}_y.wav".format(args.output_name),
+                    cmb_spectrogram_to_wave(specs[1], mp),
+                    mp.param["sr"],
+                )
+
+        sf.write(
+            "{}_v.wav".format(args.output_name),
+            cmb_spectrogram_to_wave(v_spec, mp),
+            mp.param["sr"],
+        )
+    else:
+        if not args.algorithm == "deep":
+            sf.write(
+                os.path.join("ensembled", "{}.wav".format(args.output_name)),
+                cmb_spectrogram_to_wave(ensembling(args.algorithm, specs), mp),
+                mp.param["sr"],
+            )
+
+    if args.algorithm == "align":
+        trackalignment = [
+            {
+                "file1": '"{}"'.format(args.input[0]),
+                "file2": '"{}"'.format(args.input[1]),
+            }
+        ]
+
+        for i, e in tqdm(enumerate(trackalignment), desc="Performing Alignment..."):
+            os.system(f"python lib/align_tracks.py {e['file1']} {e['file2']}")
+
+    # print('Total time: {0:.{1}f}s'.format(time.time() - start_time, 1))

infer/lib/uvr5_pack/name_params.json

@@ -0,0 +1,263 @@
+{
+    "equivalent" : [
+        {
+            "model_hash_name" : [
+                {
+                    "hash_name": "47939caf0cfe52a0e81442b85b971dfd",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json",
+                    "param_name": "4band_44100"
+                },
+                {
+                    "hash_name": "4e4ecb9764c50a8c414fee6e10395bbe",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2.json",
+                    "param_name": "4band_v2"
+                },
+                {
+                    "hash_name": "ca106edd563e034bde0bdec4bb7a4b36",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2.json",
+                    "param_name": "4band_v2"
+                },
+                {
+                    "hash_name": "e60a1e84803ce4efc0a6551206cc4b71",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json",
+                    "param_name": "4band_44100"
+                },
+                {
+                    "hash_name": "a82f14e75892e55e994376edbf0c8435",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json",
+                    "param_name": "4band_44100"
+                },
+                {
+                    "hash_name": "6dd9eaa6f0420af9f1d403aaafa4cc06",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2_sn.json",
+                    "param_name": "4band_v2_sn"
+                },
+                {
+                    "hash_name": "08611fb99bd59eaa79ad27c58d137727",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2_sn.json",
+                    "param_name": "4band_v2_sn"
+                },
+                {
+                    "hash_name": "5c7bbca45a187e81abbbd351606164e5",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_msb2.json",
+                    "param_name": "3band_44100_msb2"
+                },
+                {
+                    "hash_name": "d6b2cb685a058a091e5e7098192d3233",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_msb2.json",
+                    "param_name": "3band_44100_msb2"
+                },
+                {
+                    "hash_name": "c1b9f38170a7c90e96f027992eb7c62b",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json",
+                    "param_name": "4band_44100"
+                },
+                {
+                    "hash_name": "c3448ec923fa0edf3d03a19e633faa53",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json",
+                    "param_name": "4band_44100"
+                },
+                {
+                    "hash_name": "68aa2c8093d0080704b200d140f59e54",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100.json",
+                    "param_name": "3band_44100"
+                },
+                {
+                    "hash_name": "fdc83be5b798e4bd29fe00fe6600e147",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_mid.json",
+                    "param_name": "3band_44100_mid.json"
+                },
+                {
+                    "hash_name": "2ce34bc92fd57f55db16b7a4def3d745",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_mid.json",
+                    "param_name": "3band_44100_mid.json"
+                },
+                {
+                    "hash_name": "52fdca89576f06cf4340b74a4730ee5f",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json",
+                    "param_name": "4band_44100.json"
+                },
+                {
+                    "hash_name": "41191165b05d38fc77f072fa9e8e8a30",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json",
+                    "param_name": "4band_44100.json"
+                },
+                {
+                    "hash_name": "89e83b511ad474592689e562d5b1f80e",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/2band_32000.json",
+                    "param_name": "2band_32000.json"
+                },
+                {
+                    "hash_name": "0b954da81d453b716b114d6d7c95177f",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/2band_32000.json",
+                    "param_name": "2band_32000.json"
+                }
+
+            ],
+            "v4 Models": [
+                {
+                    "hash_name": "6a00461c51c2920fd68937d4609ed6c8",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr16000_hl512.json",
+                    "param_name": "1band_sr16000_hl512"
+                },
+                {
+                    "hash_name": "0ab504864d20f1bd378fe9c81ef37140",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json",
+                    "param_name": "1band_sr32000_hl512"
+                },
+                {
+                    "hash_name": "7dd21065bf91c10f7fccb57d7d83b07f",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json",
+                    "param_name": "1band_sr32000_hl512"
+                },
+                {
+                    "hash_name": "80ab74d65e515caa3622728d2de07d23",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json",
+                    "param_name": "1band_sr32000_hl512"
+                },
+                {
+                    "hash_name": "edc115e7fc523245062200c00caa847f",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json",
+                    "param_name": "1band_sr33075_hl384"
+                },
+                {
+                    "hash_name": "28063e9f6ab5b341c5f6d3c67f2045b7",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json",
+                    "param_name": "1band_sr33075_hl384"
+                },
+                {
+                    "hash_name": "b58090534c52cbc3e9b5104bad666ef2",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json",
+                    "param_name": "1band_sr44100_hl512"
+                },
+                {
+                    "hash_name": "0cdab9947f1b0928705f518f3c78ea8f",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json",
+                    "param_name": "1band_sr44100_hl512"
+                },
+                {
+                    "hash_name": "ae702fed0238afb5346db8356fe25f13",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl1024.json",
+                    "param_name": "1band_sr44100_hl1024"
+                }
+            ]
+        }
+    ],
+    "User Models" : [
+        {
+            "1 Band": [
+                {
+                    "hash_name": "1band_sr16000_hl512",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr16000_hl512.json",
+                    "param_name": "1band_sr16000_hl512"
+                },
+                {
+                    "hash_name": "1band_sr32000_hl512",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json",
+                    "param_name": "1band_sr16000_hl512"
+                },
+                {
+                    "hash_name": "1band_sr33075_hl384",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json",
+                    "param_name": "1band_sr33075_hl384"
+                },
+                {
+                    "hash_name": "1band_sr44100_hl256",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl256.json",
+                    "param_name": "1band_sr44100_hl256"
+                },
+                {
+                    "hash_name": "1band_sr44100_hl512",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json",
+                    "param_name": "1band_sr44100_hl512"
+                },
+                {
+                    "hash_name": "1band_sr44100_hl1024",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl1024.json",
+                    "param_name": "1band_sr44100_hl1024"
+                }
+            ],
+            "2 Band": [
+                {
+                    "hash_name": "2band_44100_lofi",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/2band_44100_lofi.json",
+                    "param_name": "2band_44100_lofi"
+                },
+                {
+                    "hash_name": "2band_32000",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/2band_32000.json",
+                    "param_name": "2band_32000"
+                },
+                {
+                    "hash_name": "2band_48000",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/2band_48000.json",
+                    "param_name": "2band_48000"
+                }
+            ],
+            "3 Band": [
+                {
+                    "hash_name": "3band_44100",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100.json",
+                    "param_name": "3band_44100"
+                },
+                {
+                    "hash_name": "3band_44100_mid",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_mid.json",
+                    "param_name": "3band_44100_mid"
+                },
+                {
+                    "hash_name": "3band_44100_msb2",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_msb2.json",
+                    "param_name": "3band_44100_msb2"
+                }
+            ],
+            "4 Band": [
+                {
+                    "hash_name": "4band_44100",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json",
+                    "param_name": "4band_44100"
+                },
+                {
+                    "hash_name": "4band_44100_mid",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_mid.json",
+                    "param_name": "4band_44100_mid"
+                },
+                {
+                    "hash_name": "4band_44100_msb",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_msb.json",
+                    "param_name": "4band_44100_msb"
+                },
+                {
+                    "hash_name": "4band_44100_msb2",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_msb2.json",
+                    "param_name": "4band_44100_msb2"
+                },
+                {
+                    "hash_name": "4band_44100_reverse",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_reverse.json",
+                    "param_name": "4band_44100_reverse"
+                },
+                {
+                    "hash_name": "4band_44100_sw",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_sw.json",
+                    "param_name": "4band_44100_sw"
+                },
+                {
+                    "hash_name": "4band_v2",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2.json",
+                    "param_name": "4band_v2"
+                },
+                {
+                    "hash_name": "4band_v2_sn",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2_sn.json",
+                    "param_name": "4band_v2_sn"
+                },
+                {
+                    "hash_name": "tmodelparam",
+                    "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/tmodelparam.json",
+                    "param_name": "User Model Param Set"
+                }
+            ]
+        }
+    ]
+}

infer/lib/uvr5_pack/utils.py

@@ -0,0 +1,121 @@
+import json
+
+import numpy as np
+import torch
+from tqdm import tqdm
+
+
+def load_data(file_name: str = "./infer/lib/uvr5_pack/name_params.json") -> dict:
+    with open(file_name, "r") as f:
+        data = json.load(f)
+
+    return data
+
+
+def make_padding(width, cropsize, offset):
+    left = offset
+    roi_size = cropsize - left * 2
+    if roi_size == 0:
+        roi_size = cropsize
+    right = roi_size - (width % roi_size) + left
+
+    return left, right, roi_size
+
+
+def inference(X_spec, device, model, aggressiveness, data):
+    """
+    data ： dic configs
+    """
+
+    def _execute(
+        X_mag_pad, roi_size, n_window, device, model, aggressiveness, is_half=True
+    ):
+        model.eval()
+        with torch.no_grad():
+            preds = []
+
+            iterations = [n_window]
+
+            total_iterations = sum(iterations)
+            for i in tqdm(range(n_window)):
+                start = i * roi_size
+                X_mag_window = X_mag_pad[
+                    None, :, :, start : start + data["window_size"]
+                ]
+                X_mag_window = torch.from_numpy(X_mag_window)
+                if is_half:
+                    X_mag_window = X_mag_window.half()
+                X_mag_window = X_mag_window.to(device)
+
+                pred = model.predict(X_mag_window, aggressiveness)
+
+                pred = pred.detach().cpu().numpy()
+                preds.append(pred[0])
+
+            pred = np.concatenate(preds, axis=2)
+        return pred
+
+    def preprocess(X_spec):
+        X_mag = np.abs(X_spec)
+        X_phase = np.angle(X_spec)
+
+        return X_mag, X_phase
+
+    X_mag, X_phase = preprocess(X_spec)
+
+    coef = X_mag.max()
+    X_mag_pre = X_mag / coef
+
+    n_frame = X_mag_pre.shape[2]
+    pad_l, pad_r, roi_size = make_padding(n_frame, data["window_size"], model.offset)
+    n_window = int(np.ceil(n_frame / roi_size))
+
+    X_mag_pad = np.pad(X_mag_pre, ((0, 0), (0, 0), (pad_l, pad_r)), mode="constant")
+
+    if list(model.state_dict().values())[0].dtype == torch.float16:
+        is_half = True
+    else:
+        is_half = False
+    pred = _execute(
+        X_mag_pad, roi_size, n_window, device, model, aggressiveness, is_half
+    )
+    pred = pred[:, :, :n_frame]
+
+    if data["tta"]:
+        pad_l += roi_size // 2
+        pad_r += roi_size // 2
+        n_window += 1
+
+        X_mag_pad = np.pad(X_mag_pre, ((0, 0), (0, 0), (pad_l, pad_r)), mode="constant")
+
+        pred_tta = _execute(
+            X_mag_pad, roi_size, n_window, device, model, aggressiveness, is_half
+        )
+        pred_tta = pred_tta[:, :, roi_size // 2 :]
+        pred_tta = pred_tta[:, :, :n_frame]
+
+        return (pred + pred_tta) * 0.5 * coef, X_mag, np.exp(1.0j * X_phase)
+    else:
+        return pred * coef, X_mag, np.exp(1.0j * X_phase)
+
+
+def _get_name_params(model_path, model_hash):
+    data = load_data()
+    flag = False
+    ModelName = model_path
+    for type in list(data):
+        for model in list(data[type][0]):
+            for i in range(len(data[type][0][model])):
+                if str(data[type][0][model][i]["hash_name"]) == model_hash:
+                    flag = True
+                elif str(data[type][0][model][i]["hash_name"]) in ModelName:
+                    flag = True
+
+                if flag:
+                    model_params_auto = data[type][0][model][i]["model_params"]
+                    param_name_auto = data[type][0][model][i]["param_name"]
+                    if type == "equivalent":
+                        return param_name_auto, model_params_auto
+                    else:
+                        flag = False
+    return param_name_auto, model_params_auto

infer/modules/ipex/__init__.py

@@ -0,0 +1,190 @@
+import os
+import sys
+import contextlib
+import torch
+import intel_extension_for_pytorch as ipex  # pylint: disable=import-error, unused-import
+from .hijacks import ipex_hijacks
+from .attention import attention_init
+
+# pylint: disable=protected-access, missing-function-docstring, line-too-long
+
+
+def ipex_init():  # pylint: disable=too-many-statements
+    try:
+        # Replace cuda with xpu:
+        torch.cuda.current_device = torch.xpu.current_device
+        torch.cuda.current_stream = torch.xpu.current_stream
+        torch.cuda.device = torch.xpu.device
+        torch.cuda.device_count = torch.xpu.device_count
+        torch.cuda.device_of = torch.xpu.device_of
+        torch.cuda.get_device_name = torch.xpu.get_device_name
+        torch.cuda.get_device_properties = torch.xpu.get_device_properties
+        torch.cuda.init = torch.xpu.init
+        torch.cuda.is_available = torch.xpu.is_available
+        torch.cuda.is_initialized = torch.xpu.is_initialized
+        torch.cuda.is_current_stream_capturing = lambda: False
+        torch.cuda.set_device = torch.xpu.set_device
+        torch.cuda.stream = torch.xpu.stream
+        torch.cuda.synchronize = torch.xpu.synchronize
+        torch.cuda.Event = torch.xpu.Event
+        torch.cuda.Stream = torch.xpu.Stream
+        torch.cuda.FloatTensor = torch.xpu.FloatTensor
+        torch.Tensor.cuda = torch.Tensor.xpu
+        torch.Tensor.is_cuda = torch.Tensor.is_xpu
+        torch.cuda._initialization_lock = torch.xpu.lazy_init._initialization_lock
+        torch.cuda._initialized = torch.xpu.lazy_init._initialized
+        torch.cuda._lazy_seed_tracker = torch.xpu.lazy_init._lazy_seed_tracker
+        torch.cuda._queued_calls = torch.xpu.lazy_init._queued_calls
+        torch.cuda._tls = torch.xpu.lazy_init._tls
+        torch.cuda.threading = torch.xpu.lazy_init.threading
+        torch.cuda.traceback = torch.xpu.lazy_init.traceback
+        torch.cuda.Optional = torch.xpu.Optional
+        torch.cuda.__cached__ = torch.xpu.__cached__
+        torch.cuda.__loader__ = torch.xpu.__loader__
+        torch.cuda.ComplexFloatStorage = torch.xpu.ComplexFloatStorage
+        torch.cuda.Tuple = torch.xpu.Tuple
+        torch.cuda.streams = torch.xpu.streams
+        torch.cuda._lazy_new = torch.xpu._lazy_new
+        torch.cuda.FloatStorage = torch.xpu.FloatStorage
+        torch.cuda.Any = torch.xpu.Any
+        torch.cuda.__doc__ = torch.xpu.__doc__
+        torch.cuda.default_generators = torch.xpu.default_generators
+        torch.cuda.HalfTensor = torch.xpu.HalfTensor
+        torch.cuda._get_device_index = torch.xpu._get_device_index
+        torch.cuda.__path__ = torch.xpu.__path__
+        torch.cuda.Device = torch.xpu.Device
+        torch.cuda.IntTensor = torch.xpu.IntTensor
+        torch.cuda.ByteStorage = torch.xpu.ByteStorage
+        torch.cuda.set_stream = torch.xpu.set_stream
+        torch.cuda.BoolStorage = torch.xpu.BoolStorage
+        torch.cuda.os = torch.xpu.os
+        torch.cuda.torch = torch.xpu.torch
+        torch.cuda.BFloat16Storage = torch.xpu.BFloat16Storage
+        torch.cuda.Union = torch.xpu.Union
+        torch.cuda.DoubleTensor = torch.xpu.DoubleTensor
+        torch.cuda.ShortTensor = torch.xpu.ShortTensor
+        torch.cuda.LongTensor = torch.xpu.LongTensor
+        torch.cuda.IntStorage = torch.xpu.IntStorage
+        torch.cuda.LongStorage = torch.xpu.LongStorage
+        torch.cuda.__annotations__ = torch.xpu.__annotations__
+        torch.cuda.__package__ = torch.xpu.__package__
+        torch.cuda.__builtins__ = torch.xpu.__builtins__
+        torch.cuda.CharTensor = torch.xpu.CharTensor
+        torch.cuda.List = torch.xpu.List
+        torch.cuda._lazy_init = torch.xpu._lazy_init
+        torch.cuda.BFloat16Tensor = torch.xpu.BFloat16Tensor
+        torch.cuda.DoubleStorage = torch.xpu.DoubleStorage
+        torch.cuda.ByteTensor = torch.xpu.ByteTensor
+        torch.cuda.StreamContext = torch.xpu.StreamContext
+        torch.cuda.ComplexDoubleStorage = torch.xpu.ComplexDoubleStorage
+        torch.cuda.ShortStorage = torch.xpu.ShortStorage
+        torch.cuda._lazy_call = torch.xpu._lazy_call
+        torch.cuda.HalfStorage = torch.xpu.HalfStorage
+        torch.cuda.random = torch.xpu.random
+        torch.cuda._device = torch.xpu._device
+        torch.cuda.classproperty = torch.xpu.classproperty
+        torch.cuda.__name__ = torch.xpu.__name__
+        torch.cuda._device_t = torch.xpu._device_t
+        torch.cuda.warnings = torch.xpu.warnings
+        torch.cuda.__spec__ = torch.xpu.__spec__
+        torch.cuda.BoolTensor = torch.xpu.BoolTensor
+        torch.cuda.CharStorage = torch.xpu.CharStorage
+        torch.cuda.__file__ = torch.xpu.__file__
+        torch.cuda._is_in_bad_fork = torch.xpu.lazy_init._is_in_bad_fork
+        # torch.cuda.is_current_stream_capturing = torch.xpu.is_current_stream_capturing
+
+        # Memory:
+        torch.cuda.memory = torch.xpu.memory
+        if "linux" in sys.platform and "WSL2" in os.popen("uname -a").read():
+            torch.xpu.empty_cache = lambda: None
+        torch.cuda.empty_cache = torch.xpu.empty_cache
+        torch.cuda.memory_stats = torch.xpu.memory_stats
+        torch.cuda.memory_summary = torch.xpu.memory_summary
+        torch.cuda.memory_snapshot = torch.xpu.memory_snapshot
+        torch.cuda.memory_allocated = torch.xpu.memory_allocated
+        torch.cuda.max_memory_allocated = torch.xpu.max_memory_allocated
+        torch.cuda.memory_reserved = torch.xpu.memory_reserved
+        torch.cuda.memory_cached = torch.xpu.memory_reserved
+        torch.cuda.max_memory_reserved = torch.xpu.max_memory_reserved
+        torch.cuda.max_memory_cached = torch.xpu.max_memory_reserved
+        torch.cuda.reset_peak_memory_stats = torch.xpu.reset_peak_memory_stats
+        torch.cuda.reset_max_memory_cached = torch.xpu.reset_peak_memory_stats
+        torch.cuda.reset_max_memory_allocated = torch.xpu.reset_peak_memory_stats
+        torch.cuda.memory_stats_as_nested_dict = torch.xpu.memory_stats_as_nested_dict
+        torch.cuda.reset_accumulated_memory_stats = (
+            torch.xpu.reset_accumulated_memory_stats
+        )
+
+        # RNG:
+        torch.cuda.get_rng_state = torch.xpu.get_rng_state
+        torch.cuda.get_rng_state_all = torch.xpu.get_rng_state_all
+        torch.cuda.set_rng_state = torch.xpu.set_rng_state
+        torch.cuda.set_rng_state_all = torch.xpu.set_rng_state_all
+        torch.cuda.manual_seed = torch.xpu.manual_seed
+        torch.cuda.manual_seed_all = torch.xpu.manual_seed_all
+        torch.cuda.seed = torch.xpu.seed
+        torch.cuda.seed_all = torch.xpu.seed_all
+        torch.cuda.initial_seed = torch.xpu.initial_seed
+
+        # AMP:
+        torch.cuda.amp = torch.xpu.amp
+        if not hasattr(torch.cuda.amp, "common"):
+            torch.cuda.amp.common = contextlib.nullcontext()
+        torch.cuda.amp.common.amp_definitely_not_available = lambda: False
+        try:
+            torch.cuda.amp.GradScaler = torch.xpu.amp.GradScaler
+        except Exception:  # pylint: disable=broad-exception-caught
+            try:
+                from .gradscaler import (
+                    gradscaler_init,
+                )  # pylint: disable=import-outside-toplevel, import-error
+
+                gradscaler_init()
+                torch.cuda.amp.GradScaler = torch.xpu.amp.GradScaler
+            except Exception:  # pylint: disable=broad-exception-caught
+                torch.cuda.amp.GradScaler = ipex.cpu.autocast._grad_scaler.GradScaler
+
+        # C
+        torch._C._cuda_getCurrentRawStream = ipex._C._getCurrentStream
+        ipex._C._DeviceProperties.major = 2023
+        ipex._C._DeviceProperties.minor = 2
+
+        # Fix functions with ipex:
+        torch.cuda.mem_get_info = lambda device=None: [
+            (
+                torch.xpu.get_device_properties(device).total_memory
+                - torch.xpu.memory_allocated(device)
+            ),
+            torch.xpu.get_device_properties(device).total_memory,
+        ]
+        torch._utils._get_available_device_type = lambda: "xpu"
+        torch.has_cuda = True
+        torch.cuda.has_half = True
+        torch.cuda.is_bf16_supported = lambda *args, **kwargs: True
+        torch.cuda.is_fp16_supported = lambda *args, **kwargs: True
+        torch.version.cuda = "11.7"
+        torch.cuda.get_device_capability = lambda *args, **kwargs: [11, 7]
+        torch.cuda.get_device_properties.major = 11
+        torch.cuda.get_device_properties.minor = 7
+        torch.cuda.ipc_collect = lambda *args, **kwargs: None
+        torch.cuda.utilization = lambda *args, **kwargs: 0
+        if hasattr(torch.xpu, "getDeviceIdListForCard"):
+            torch.cuda.getDeviceIdListForCard = torch.xpu.getDeviceIdListForCard
+            torch.cuda.get_device_id_list_per_card = torch.xpu.getDeviceIdListForCard
+        else:
+            torch.cuda.getDeviceIdListForCard = torch.xpu.get_device_id_list_per_card
+            torch.cuda.get_device_id_list_per_card = (
+                torch.xpu.get_device_id_list_per_card
+            )
+
+        ipex_hijacks()
+        attention_init()
+        try:
+            from .diffusers import ipex_diffusers
+
+            ipex_diffusers()
+        except Exception:  # pylint: disable=broad-exception-caught
+            pass
+    except Exception as e:
+        return False, e
+    return True, None

infer/modules/ipex/attention.py

@@ -0,0 +1,212 @@
+import torch
+import intel_extension_for_pytorch as ipex  # pylint: disable=import-error, unused-import
+
+# pylint: disable=protected-access, missing-function-docstring, line-too-long
+
+original_torch_bmm = torch.bmm
+
+
+def torch_bmm(input, mat2, *, out=None):
+    if input.dtype != mat2.dtype:
+        mat2 = mat2.to(input.dtype)
+
+    # ARC GPUs can't allocate more than 4GB to a single block, Slice it:
+    batch_size_attention, input_tokens, mat2_shape = (
+        input.shape[0],
+        input.shape[1],
+        mat2.shape[2],
+    )
+    block_multiply = input.element_size()
+    slice_block_size = input_tokens * mat2_shape / 1024 / 1024 * block_multiply
+    block_size = batch_size_attention * slice_block_size
+
+    split_slice_size = batch_size_attention
+    if block_size > 4:
+        do_split = True
+        # Find something divisible with the input_tokens
+        while (split_slice_size * slice_block_size) > 4:
+            split_slice_size = split_slice_size // 2
+            if split_slice_size <= 1:
+                split_slice_size = 1
+                break
+    else:
+        do_split = False
+
+    split_2_slice_size = input_tokens
+    if split_slice_size * slice_block_size > 4:
+        slice_block_size2 = split_slice_size * mat2_shape / 1024 / 1024 * block_multiply
+        do_split_2 = True
+        # Find something divisible with the input_tokens
+        while (split_2_slice_size * slice_block_size2) > 4:
+            split_2_slice_size = split_2_slice_size // 2
+            if split_2_slice_size <= 1:
+                split_2_slice_size = 1
+                break
+    else:
+        do_split_2 = False
+
+    if do_split:
+        hidden_states = torch.zeros(
+            input.shape[0],
+            input.shape[1],
+            mat2.shape[2],
+            device=input.device,
+            dtype=input.dtype,
+        )
+        for i in range(batch_size_attention // split_slice_size):
+            start_idx = i * split_slice_size
+            end_idx = (i + 1) * split_slice_size
+            if do_split_2:
+                for i2 in range(
+                    input_tokens // split_2_slice_size
+                ):  # pylint: disable=invalid-name
+                    start_idx_2 = i2 * split_2_slice_size
+                    end_idx_2 = (i2 + 1) * split_2_slice_size
+                    hidden_states[
+                        start_idx:end_idx, start_idx_2:end_idx_2
+                    ] = original_torch_bmm(
+                        input[start_idx:end_idx, start_idx_2:end_idx_2],
+                        mat2[start_idx:end_idx, start_idx_2:end_idx_2],
+                        out=out,
+                    )
+            else:
+                hidden_states[start_idx:end_idx] = original_torch_bmm(
+                    input[start_idx:end_idx], mat2[start_idx:end_idx], out=out
+                )
+    else:
+        return original_torch_bmm(input, mat2, out=out)
+    return hidden_states
+
+
+original_scaled_dot_product_attention = torch.nn.functional.scaled_dot_product_attention
+
+
+def scaled_dot_product_attention(
+    query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False
+):
+    # ARC GPUs can't allocate more than 4GB to a single block, Slice it:
+    if len(query.shape) == 3:
+        batch_size_attention, query_tokens, shape_four = query.shape
+        shape_one = 1
+        no_shape_one = True
+    else:
+        shape_one, batch_size_attention, query_tokens, shape_four = query.shape
+        no_shape_one = False
+
+    block_multiply = query.element_size()
+    slice_block_size = (
+        shape_one * query_tokens * shape_four / 1024 / 1024 * block_multiply
+    )
+    block_size = batch_size_attention * slice_block_size
+
+    split_slice_size = batch_size_attention
+    if block_size > 4:
+        do_split = True
+        # Find something divisible with the shape_one
+        while (split_slice_size * slice_block_size) > 4:
+            split_slice_size = split_slice_size // 2
+            if split_slice_size <= 1:
+                split_slice_size = 1
+                break
+    else:
+        do_split = False
+
+    split_2_slice_size = query_tokens
+    if split_slice_size * slice_block_size > 4:
+        slice_block_size2 = (
+            shape_one * split_slice_size * shape_four / 1024 / 1024 * block_multiply
+        )
+        do_split_2 = True
+        # Find something divisible with the batch_size_attention
+        while (split_2_slice_size * slice_block_size2) > 4:
+            split_2_slice_size = split_2_slice_size // 2
+            if split_2_slice_size <= 1:
+                split_2_slice_size = 1
+                break
+    else:
+        do_split_2 = False
+
+    if do_split:
+        hidden_states = torch.zeros(query.shape, device=query.device, dtype=query.dtype)
+        for i in range(batch_size_attention // split_slice_size):
+            start_idx = i * split_slice_size
+            end_idx = (i + 1) * split_slice_size
+            if do_split_2:
+                for i2 in range(
+                    query_tokens // split_2_slice_size
+                ):  # pylint: disable=invalid-name
+                    start_idx_2 = i2 * split_2_slice_size
+                    end_idx_2 = (i2 + 1) * split_2_slice_size
+                    if no_shape_one:
+                        hidden_states[
+                            start_idx:end_idx, start_idx_2:end_idx_2
+                        ] = original_scaled_dot_product_attention(
+                            query[start_idx:end_idx, start_idx_2:end_idx_2],
+                            key[start_idx:end_idx, start_idx_2:end_idx_2],
+                            value[start_idx:end_idx, start_idx_2:end_idx_2],
+                            attn_mask=attn_mask[
+                                start_idx:end_idx, start_idx_2:end_idx_2
+                            ]
+                            if attn_mask is not None
+                            else attn_mask,
+                            dropout_p=dropout_p,
+                            is_causal=is_causal,
+                        )
+                    else:
+                        hidden_states[
+                            :, start_idx:end_idx, start_idx_2:end_idx_2
+                        ] = original_scaled_dot_product_attention(
+                            query[:, start_idx:end_idx, start_idx_2:end_idx_2],
+                            key[:, start_idx:end_idx, start_idx_2:end_idx_2],
+                            value[:, start_idx:end_idx, start_idx_2:end_idx_2],
+                            attn_mask=attn_mask[
+                                :, start_idx:end_idx, start_idx_2:end_idx_2
+                            ]
+                            if attn_mask is not None
+                            else attn_mask,
+                            dropout_p=dropout_p,
+                            is_causal=is_causal,
+                        )
+            else:
+                if no_shape_one:
+                    hidden_states[
+                        start_idx:end_idx
+                    ] = original_scaled_dot_product_attention(
+                        query[start_idx:end_idx],
+                        key[start_idx:end_idx],
+                        value[start_idx:end_idx],
+                        attn_mask=attn_mask[start_idx:end_idx]
+                        if attn_mask is not None
+                        else attn_mask,
+                        dropout_p=dropout_p,
+                        is_causal=is_causal,
+                    )
+                else:
+                    hidden_states[
+                        :, start_idx:end_idx
+                    ] = original_scaled_dot_product_attention(
+                        query[:, start_idx:end_idx],
+                        key[:, start_idx:end_idx],
+                        value[:, start_idx:end_idx],
+                        attn_mask=attn_mask[:, start_idx:end_idx]
+                        if attn_mask is not None
+                        else attn_mask,
+                        dropout_p=dropout_p,
+                        is_causal=is_causal,
+                    )
+    else:
+        return original_scaled_dot_product_attention(
+            query,
+            key,
+            value,
+            attn_mask=attn_mask,
+            dropout_p=dropout_p,
+            is_causal=is_causal,
+        )
+    return hidden_states
+
+
+def attention_init():
+    # ARC GPUs can't allocate more than 4GB to a single block:
+    torch.bmm = torch_bmm
+    torch.nn.functional.scaled_dot_product_attention = scaled_dot_product_attention

infer/modules/ipex/gradscaler.py

@@ -0,0 +1,187 @@
+from collections import defaultdict
+import torch
+import intel_extension_for_pytorch as ipex  # pylint: disable=import-error, unused-import
+import intel_extension_for_pytorch._C as core  # pylint: disable=import-error, unused-import
+
+# pylint: disable=protected-access, missing-function-docstring, line-too-long
+
+OptState = ipex.cpu.autocast._grad_scaler.OptState
+_MultiDeviceReplicator = ipex.cpu.autocast._grad_scaler._MultiDeviceReplicator
+_refresh_per_optimizer_state = (
+    ipex.cpu.autocast._grad_scaler._refresh_per_optimizer_state
+)
+
+
+def _unscale_grads_(
+    self, optimizer, inv_scale, found_inf, allow_fp16
+):  # pylint: disable=unused-argument
+    per_device_inv_scale = _MultiDeviceReplicator(inv_scale)
+    per_device_found_inf = _MultiDeviceReplicator(found_inf)
+
+    # To set up _amp_foreach_non_finite_check_and_unscale_, split grads by device and dtype.
+    # There could be hundreds of grads, so we'd like to iterate through them just once.
+    # However, we don't know their devices or dtypes in advance.
+
+    # https://stackoverflow.com/questions/5029934/defaultdict-of-defaultdict
+    # Google says mypy struggles with defaultdicts type annotations.
+    per_device_and_dtype_grads = defaultdict(lambda: defaultdict(list))  # type: ignore[var-annotated]
+    # sync grad to master weight
+    if hasattr(optimizer, "sync_grad"):
+        optimizer.sync_grad()
+    with torch.no_grad():
+        for group in optimizer.param_groups:
+            for param in group["params"]:
+                if param.grad is None:
+                    continue
+                if (not allow_fp16) and param.grad.dtype == torch.float16:
+                    raise ValueError("Attempting to unscale FP16 gradients.")
+                if param.grad.is_sparse:
+                    # is_coalesced() == False means the sparse grad has values with duplicate indices.
+                    # coalesce() deduplicates indices and adds all values that have the same index.
+                    # For scaled fp16 values, there's a good chance coalescing will cause overflow,
+                    # so we should check the coalesced _values().
+                    if param.grad.dtype is torch.float16:
+                        param.grad = param.grad.coalesce()
+                    to_unscale = param.grad._values()
+                else:
+                    to_unscale = param.grad
+
+                # -: is there a way to split by device and dtype without appending in the inner loop?
+                to_unscale = to_unscale.to("cpu")
+                per_device_and_dtype_grads[to_unscale.device][to_unscale.dtype].append(
+                    to_unscale
+                )
+
+        for _, per_dtype_grads in per_device_and_dtype_grads.items():
+            for grads in per_dtype_grads.values():
+                core._amp_foreach_non_finite_check_and_unscale_(
+                    grads,
+                    per_device_found_inf.get("cpu"),
+                    per_device_inv_scale.get("cpu"),
+                )
+
+    return per_device_found_inf._per_device_tensors
+
+
+def unscale_(self, optimizer):
+    """
+    Divides ("unscales") the optimizer's gradient tensors by the scale factor.
+    :meth:`unscale_` is optional, serving cases where you need to
+    :ref:`modify or inspect gradients<working-with-unscaled-gradients>`
+    between the backward pass(es) and :meth:`step`.
+    If :meth:`unscale_` is not called explicitly,  gradients will be unscaled  automatically during :meth:`step`.
+    Simple example, using :meth:`unscale_` to enable clipping of unscaled gradients::
+        ...
+        scaler.scale(loss).backward()
+        scaler.unscale_(optimizer)
+        torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm)
+        scaler.step(optimizer)
+        scaler.update()
+    Args:
+        optimizer (torch.optim.Optimizer):  Optimizer that owns the gradients to be unscaled.
+    .. warning::
+        :meth:`unscale_` should only be called once per optimizer per :meth:`step` call,
+        and only after all gradients for that optimizer's assigned parameters have been accumulated.
+        Calling :meth:`unscale_` twice for a given optimizer between each :meth:`step` triggers a RuntimeError.
+    .. warning::
+        :meth:`unscale_` may unscale sparse gradients out of place, replacing the ``.grad`` attribute.
+    """
+    if not self._enabled:
+        return
+
+    self._check_scale_growth_tracker("unscale_")
+
+    optimizer_state = self._per_optimizer_states[id(optimizer)]
+
+    if optimizer_state["stage"] is OptState.UNSCALED:  # pylint: disable=no-else-raise
+        raise RuntimeError(
+            "unscale_() has already been called on this optimizer since the last update()."
+        )
+    elif optimizer_state["stage"] is OptState.STEPPED:
+        raise RuntimeError("unscale_() is being called after step().")
+
+    # FP32 division can be imprecise for certain compile options, so we carry out the reciprocal in FP64.
+    assert self._scale is not None
+    inv_scale = (
+        self._scale.to("cpu").double().reciprocal().float().to(self._scale.device)
+    )
+    found_inf = torch.full((1,), 0.0, dtype=torch.float32, device=self._scale.device)
+
+    optimizer_state["found_inf_per_device"] = self._unscale_grads_(
+        optimizer, inv_scale, found_inf, False
+    )
+    optimizer_state["stage"] = OptState.UNSCALED
+
+
+def update(self, new_scale=None):
+    """
+    Updates the scale factor.
+    If any optimizer steps were skipped the scale is multiplied by ``backoff_factor``
+    to reduce it. If ``growth_interval`` unskipped iterations occurred consecutively,
+    the scale is multiplied by ``growth_factor`` to increase it.
+    Passing ``new_scale`` sets the new scale value manually. (``new_scale`` is not
+    used directly, it's used to fill GradScaler's internal scale tensor. So if
+    ``new_scale`` was a tensor, later in-place changes to that tensor will not further
+    affect the scale GradScaler uses internally.)
+    Args:
+        new_scale (float or :class:`torch.FloatTensor`, optional, default=None):  New scale factor.
+    .. warning::
+        :meth:`update` should only be called at the end of the iteration, after ``scaler.step(optimizer)`` has
+        been invoked for all optimizers used this iteration.
+    """
+    if not self._enabled:
+        return
+
+    _scale, _growth_tracker = self._check_scale_growth_tracker("update")
+
+    if new_scale is not None:
+        # Accept a new user-defined scale.
+        if isinstance(new_scale, float):
+            self._scale.fill_(new_scale)  # type: ignore[union-attr]
+        else:
+            reason = "new_scale should be a float or a 1-element torch.FloatTensor with requires_grad=False."
+            assert isinstance(new_scale, torch.FloatTensor), reason  # type: ignore[attr-defined]
+            assert new_scale.numel() == 1, reason
+            assert new_scale.requires_grad is False, reason
+            self._scale.copy_(new_scale)  # type: ignore[union-attr]
+    else:
+        # Consume shared inf/nan data collected from optimizers to update the scale.
+        # If all found_inf tensors are on the same device as self._scale, this operation is asynchronous.
+        found_infs = [
+            found_inf.to(device="cpu", non_blocking=True)
+            for state in self._per_optimizer_states.values()
+            for found_inf in state["found_inf_per_device"].values()
+        ]
+
+        assert len(found_infs) > 0, "No inf checks were recorded prior to update."
+
+        found_inf_combined = found_infs[0]
+        if len(found_infs) > 1:
+            for i in range(1, len(found_infs)):
+                found_inf_combined += found_infs[i]
+
+        to_device = _scale.device
+        _scale = _scale.to("cpu")
+        _growth_tracker = _growth_tracker.to("cpu")
+
+        core._amp_update_scale_(
+            _scale,
+            _growth_tracker,
+            found_inf_combined,
+            self._growth_factor,
+            self._backoff_factor,
+            self._growth_interval,
+        )
+
+        _scale = _scale.to(to_device)
+        _growth_tracker = _growth_tracker.to(to_device)
+    # To prepare for next iteration, clear the data collected from optimizers this iteration.
+    self._per_optimizer_states = defaultdict(_refresh_per_optimizer_state)
+
+
+def gradscaler_init():
+    torch.xpu.amp.GradScaler = ipex.cpu.autocast._grad_scaler.GradScaler
+    torch.xpu.amp.GradScaler._unscale_grads_ = _unscale_grads_
+    torch.xpu.amp.GradScaler.unscale_ = unscale_
+    torch.xpu.amp.GradScaler.update = update
+    return torch.xpu.amp.GradScaler

infer/modules/ipex/hijacks.py

@@ -0,0 +1,357 @@
+import contextlib
+import importlib
+import torch
+import intel_extension_for_pytorch as ipex  # pylint: disable=import-error, unused-import
+
+# pylint: disable=protected-access, missing-function-docstring, line-too-long, unnecessary-lambda, no-else-return
+
+
+class CondFunc:  # pylint: disable=missing-class-docstring
+    def __new__(cls, orig_func, sub_func, cond_func):
+        self = super(CondFunc, cls).__new__(cls)
+        if isinstance(orig_func, str):
+            func_path = orig_func.split(".")
+            for i in range(len(func_path) - 1, -1, -1):
+                try:
+                    resolved_obj = importlib.import_module(".".join(func_path[:i]))
+                    break
+                except ImportError:
+                    pass
+            for attr_name in func_path[i:-1]:
+                resolved_obj = getattr(resolved_obj, attr_name)
+            orig_func = getattr(resolved_obj, func_path[-1])
+            setattr(
+                resolved_obj,
+                func_path[-1],
+                lambda *args, **kwargs: self(*args, **kwargs),
+            )
+        self.__init__(orig_func, sub_func, cond_func)
+        return lambda *args, **kwargs: self(*args, **kwargs)
+
+    def __init__(self, orig_func, sub_func, cond_func):
+        self.__orig_func = orig_func
+        self.__sub_func = sub_func
+        self.__cond_func = cond_func
+
+    def __call__(self, *args, **kwargs):
+        if not self.__cond_func or self.__cond_func(self.__orig_func, *args, **kwargs):
+            return self.__sub_func(self.__orig_func, *args, **kwargs)
+        else:
+            return self.__orig_func(*args, **kwargs)
+
+
+_utils = torch.utils.data._utils
+
+
+def _shutdown_workers(self):
+    if (
+        torch.utils.data._utils is None
+        or torch.utils.data._utils.python_exit_status is True
+        or torch.utils.data._utils.python_exit_status is None
+    ):
+        return
+    if hasattr(self, "_shutdown") and not self._shutdown:
+        self._shutdown = True
+        try:
+            if hasattr(self, "_pin_memory_thread"):
+                self._pin_memory_thread_done_event.set()
+                self._worker_result_queue.put((None, None))
+                self._pin_memory_thread.join()
+                self._worker_result_queue.cancel_join_thread()
+                self._worker_result_queue.close()
+            self._workers_done_event.set()
+            for worker_id in range(len(self._workers)):
+                if self._persistent_workers or self._workers_status[worker_id]:
+                    self._mark_worker_as_unavailable(worker_id, shutdown=True)
+            for w in self._workers:  # pylint: disable=invalid-name
+                w.join(timeout=torch.utils.data._utils.MP_STATUS_CHECK_INTERVAL)
+            for q in self._index_queues:  # pylint: disable=invalid-name
+                q.cancel_join_thread()
+                q.close()
+        finally:
+            if self._worker_pids_set:
+                torch.utils.data._utils.signal_handling._remove_worker_pids(id(self))
+                self._worker_pids_set = False
+            for w in self._workers:  # pylint: disable=invalid-name
+                if w.is_alive():
+                    w.terminate()
+
+
+class DummyDataParallel(
+    torch.nn.Module
+):  # pylint: disable=missing-class-docstring, unused-argument, too-few-public-methods
+    def __new__(
+        cls, module, device_ids=None, output_device=None, dim=0
+    ):  # pylint: disable=unused-argument
+        if isinstance(device_ids, list) and len(device_ids) > 1:
+            print("IPEX backend doesn't support DataParallel on multiple XPU devices")
+        return module.to("xpu")
+
+
+def return_null_context(*args, **kwargs):  # pylint: disable=unused-argument
+    return contextlib.nullcontext()
+
+
+def check_device(device):
+    return bool(
+        (isinstance(device, torch.device) and device.type == "cuda")
+        or (isinstance(device, str) and "cuda" in device)
+        or isinstance(device, int)
+    )
+
+
+def return_xpu(device):
+    return (
+        f"xpu:{device[-1]}"
+        if isinstance(device, str) and ":" in device
+        else f"xpu:{device}"
+        if isinstance(device, int)
+        else torch.device("xpu")
+        if isinstance(device, torch.device)
+        else "xpu"
+    )
+
+
+def ipex_no_cuda(orig_func, *args, **kwargs):
+    torch.cuda.is_available = lambda: False
+    orig_func(*args, **kwargs)
+    torch.cuda.is_available = torch.xpu.is_available
+
+
+original_autocast = torch.autocast
+
+
+def ipex_autocast(*args, **kwargs):
+    if len(args) > 0 and args[0] == "cuda":
+        return original_autocast("xpu", *args[1:], **kwargs)
+    else:
+        return original_autocast(*args, **kwargs)
+
+
+original_torch_cat = torch.cat
+
+
+def torch_cat(tensor, *args, **kwargs):
+    if len(tensor) == 3 and (
+        tensor[0].dtype != tensor[1].dtype or tensor[2].dtype != tensor[1].dtype
+    ):
+        return original_torch_cat(
+            [tensor[0].to(tensor[1].dtype), tensor[1], tensor[2].to(tensor[1].dtype)],
+            *args,
+            **kwargs,
+        )
+    else:
+        return original_torch_cat(tensor, *args, **kwargs)
+
+
+original_interpolate = torch.nn.functional.interpolate
+
+
+def interpolate(
+    tensor,
+    size=None,
+    scale_factor=None,
+    mode="nearest",
+    align_corners=None,
+    recompute_scale_factor=None,
+    antialias=False,
+):  # pylint: disable=too-many-arguments
+    if antialias or align_corners is not None:
+        return_device = tensor.device
+        return_dtype = tensor.dtype
+        return original_interpolate(
+            tensor.to("cpu", dtype=torch.float32),
+            size=size,
+            scale_factor=scale_factor,
+            mode=mode,
+            align_corners=align_corners,
+            recompute_scale_factor=recompute_scale_factor,
+            antialias=antialias,
+        ).to(return_device, dtype=return_dtype)
+    else:
+        return original_interpolate(
+            tensor,
+            size=size,
+            scale_factor=scale_factor,
+            mode=mode,
+            align_corners=align_corners,
+            recompute_scale_factor=recompute_scale_factor,
+            antialias=antialias,
+        )
+
+
+original_linalg_solve = torch.linalg.solve
+
+
+def linalg_solve(A, B, *args, **kwargs):  # pylint: disable=invalid-name
+    if A.device != torch.device("cpu") or B.device != torch.device("cpu"):
+        return_device = A.device
+        return original_linalg_solve(A.to("cpu"), B.to("cpu"), *args, **kwargs).to(
+            return_device
+        )
+    else:
+        return original_linalg_solve(A, B, *args, **kwargs)
+
+
+def ipex_hijacks():
+    CondFunc(
+        "torch.Tensor.to",
+        lambda orig_func, self, device=None, *args, **kwargs: orig_func(
+            self, return_xpu(device), *args, **kwargs
+        ),
+        lambda orig_func, self, device=None, *args, **kwargs: check_device(device),
+    )
+    CondFunc(
+        "torch.Tensor.cuda",
+        lambda orig_func, self, device=None, *args, **kwargs: orig_func(
+            self, return_xpu(device), *args, **kwargs
+        ),
+        lambda orig_func, self, device=None, *args, **kwargs: check_device(device),
+    )
+    CondFunc(
+        "torch.empty",
+        lambda orig_func, *args, device=None, **kwargs: orig_func(
+            *args, device=return_xpu(device), **kwargs
+        ),
+        lambda orig_func, *args, device=None, **kwargs: check_device(device),
+    )
+    CondFunc(
+        "torch.load",
+        lambda orig_func, *args, map_location=None, **kwargs: orig_func(
+            *args, return_xpu(map_location), **kwargs
+        ),
+        lambda orig_func, *args, map_location=None, **kwargs: map_location is None
+        or check_device(map_location),
+    )
+    CondFunc(
+        "torch.randn",
+        lambda orig_func, *args, device=None, **kwargs: orig_func(
+            *args, device=return_xpu(device), **kwargs
+        ),
+        lambda orig_func, *args, device=None, **kwargs: check_device(device),
+    )
+    CondFunc(
+        "torch.ones",
+        lambda orig_func, *args, device=None, **kwargs: orig_func(
+            *args, device=return_xpu(device), **kwargs
+        ),
+        lambda orig_func, *args, device=None, **kwargs: check_device(device),
+    )
+    CondFunc(
+        "torch.zeros",
+        lambda orig_func, *args, device=None, **kwargs: orig_func(
+            *args, device=return_xpu(device), **kwargs
+        ),
+        lambda orig_func, *args, device=None, **kwargs: check_device(device),
+    )
+    CondFunc(
+        "torch.tensor",
+        lambda orig_func, *args, device=None, **kwargs: orig_func(
+            *args, device=return_xpu(device), **kwargs
+        ),
+        lambda orig_func, *args, device=None, **kwargs: check_device(device),
+    )
+    CondFunc(
+        "torch.linspace",
+        lambda orig_func, *args, device=None, **kwargs: orig_func(
+            *args, device=return_xpu(device), **kwargs
+        ),
+        lambda orig_func, *args, device=None, **kwargs: check_device(device),
+    )
+
+    CondFunc(
+        "torch.Generator",
+        lambda orig_func, device=None: torch.xpu.Generator(device),
+        lambda orig_func, device=None: device is not None
+        and device != torch.device("cpu")
+        and device != "cpu",
+    )
+
+    CondFunc(
+        "torch.batch_norm",
+        lambda orig_func, input, weight, bias, *args, **kwargs: orig_func(
+            input,
+            weight
+            if weight is not None
+            else torch.ones(input.size()[1], device=input.device),
+            bias
+            if bias is not None
+            else torch.zeros(input.size()[1], device=input.device),
+            *args,
+            **kwargs,
+        ),
+        lambda orig_func, input, *args, **kwargs: input.device != torch.device("cpu"),
+    )
+    CondFunc(
+        "torch.instance_norm",
+        lambda orig_func, input, weight, bias, *args, **kwargs: orig_func(
+            input,
+            weight
+            if weight is not None
+            else torch.ones(input.size()[1], device=input.device),
+            bias
+            if bias is not None
+            else torch.zeros(input.size()[1], device=input.device),
+            *args,
+            **kwargs,
+        ),
+        lambda orig_func, input, *args, **kwargs: input.device != torch.device("cpu"),
+    )
+
+    # Functions with dtype errors:
+    CondFunc(
+        "torch.nn.modules.GroupNorm.forward",
+        lambda orig_func, self, input: orig_func(
+            self, input.to(self.weight.data.dtype)
+        ),
+        lambda orig_func, self, input: input.dtype != self.weight.data.dtype,
+    )
+    CondFunc(
+        "torch.nn.modules.linear.Linear.forward",
+        lambda orig_func, self, input: orig_func(
+            self, input.to(self.weight.data.dtype)
+        ),
+        lambda orig_func, self, input: input.dtype != self.weight.data.dtype,
+    )
+    CondFunc(
+        "torch.nn.modules.conv.Conv2d.forward",
+        lambda orig_func, self, input: orig_func(
+            self, input.to(self.weight.data.dtype)
+        ),
+        lambda orig_func, self, input: input.dtype != self.weight.data.dtype,
+    )
+    CondFunc(
+        "torch.nn.functional.layer_norm",
+        lambda orig_func, input, normalized_shape=None, weight=None, *args, **kwargs: orig_func(
+            input.to(weight.data.dtype), normalized_shape, weight, *args, **kwargs
+        ),
+        lambda orig_func, input, normalized_shape=None, weight=None, *args, **kwargs: weight
+        is not None
+        and input.dtype != weight.data.dtype,
+    )
+
+    # Diffusers Float64 (ARC GPUs doesn't support double or Float64):
+    if not torch.xpu.has_fp64_dtype():
+        CondFunc(
+            "torch.from_numpy",
+            lambda orig_func, ndarray: orig_func(ndarray.astype("float32")),
+            lambda orig_func, ndarray: ndarray.dtype == float,
+        )
+
+    # Broken functions when torch.cuda.is_available is True:
+    CondFunc(
+        "torch.utils.data.dataloader._BaseDataLoaderIter.__init__",
+        lambda orig_func, *args, **kwargs: ipex_no_cuda(orig_func, *args, **kwargs),
+        lambda orig_func, *args, **kwargs: True,
+    )
+
+    # Functions that make compile mad with CondFunc:
+    torch.utils.data.dataloader._MultiProcessingDataLoaderIter._shutdown_workers = (
+        _shutdown_workers
+    )
+    torch.nn.DataParallel = DummyDataParallel
+    torch.autocast = ipex_autocast
+    torch.cat = torch_cat
+    torch.linalg.solve = linalg_solve
+    torch.nn.functional.interpolate = interpolate
+    torch.backends.cuda.sdp_kernel = return_null_context

infer/modules/onnx/export.py

@@ -0,0 +1,52 @@
+import torch
+
+from infer.lib.infer_pack.models_onnx import SynthesizerTrnMsNSFsidM
+
+
+def export_onnx(ModelPath, ExportedPath):
+    cpt = torch.load(ModelPath, map_location="cpu")
+    cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
+    vec_channels = 256 if cpt.get("version", "v1") == "v1" else 768
+
+    test_phone = torch.rand(1, 200, vec_channels)  # hidden unit
+    test_phone_lengths = torch.tensor([200]).long()  # hidden unit 长度（貌似没啥用）
+    test_pitch = torch.randint(size=(1, 200), low=5, high=255)  # 基频（单位赫兹）
+    test_pitchf = torch.rand(1, 200)  # nsf基频
+    test_ds = torch.LongTensor([0])  # 说话人ID
+    test_rnd = torch.rand(1, 192, 200)  # 噪声（加入随机因子）
+
+    device = "cpu"  # 导出时设备（不影响使用模型）
+
+    net_g = SynthesizerTrnMsNSFsidM(
+        *cpt["config"], is_half=False, version=cpt.get("version", "v1")
+    )  # fp32导出（C++要支持fp16必须手动将内存重新排列所以暂时不用fp16）
+    net_g.load_state_dict(cpt["weight"], strict=False)
+    input_names = ["phone", "phone_lengths", "pitch", "pitchf", "ds", "rnd"]
+    output_names = [
+        "audio",
+    ]
+    # net_g.construct_spkmixmap(n_speaker) 多角色混合轨道导出
+    torch.onnx.export(
+        net_g,
+        (
+            test_phone.to(device),
+            test_phone_lengths.to(device),
+            test_pitch.to(device),
+            test_pitchf.to(device),
+            test_ds.to(device),
+            test_rnd.to(device),
+        ),
+        ExportedPath,
+        dynamic_axes={
+            "phone": [1],
+            "pitch": [1],
+            "pitchf": [1],
+            "rnd": [2],
+        },
+        do_constant_folding=False,
+        opset_version=13,
+        verbose=False,
+        input_names=input_names,
+        output_names=output_names,
+    )
+    return "Finished"

infer/modules/train/extract/extract_f0_print.py

@@ -0,0 +1,175 @@
+import os
+import sys
+import traceback
+
+import parselmouth
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+import logging
+
+import numpy as np
+import pyworld
+
+from infer.lib.audio import load_audio
+
+logging.getLogger("numba").setLevel(logging.WARNING)
+from multiprocessing import Process
+
+exp_dir = sys.argv[1]
+f = open("%s/extract_f0_feature.log" % exp_dir, "a+")
+
+
+def printt(strr):
+    print(strr)
+    f.write("%s\n" % strr)
+    f.flush()
+
+
+n_p = int(sys.argv[2])
+f0method = sys.argv[3]
+
+
+class FeatureInput(object):
+    def __init__(self, samplerate=16000, hop_size=160):
+        self.fs = samplerate
+        self.hop = hop_size
+
+        self.f0_bin = 256
+        self.f0_max = 1100.0
+        self.f0_min = 50.0
+        self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
+        self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
+
+    def compute_f0(self, path, f0_method):
+        x = load_audio(path, self.fs)
+        p_len = x.shape[0] // self.hop
+        if f0_method == "pm":
+            time_step = 160 / 16000 * 1000
+            f0_min = 50
+            f0_max = 1100
+            f0 = (
+                parselmouth.Sound(x, self.fs)
+                .to_pitch_ac(
+                    time_step=time_step / 1000,
+                    voicing_threshold=0.6,
+                    pitch_floor=f0_min,
+                    pitch_ceiling=f0_max,
+                )
+                .selected_array["frequency"]
+            )
+            pad_size = (p_len - len(f0) + 1) // 2
+            if pad_size > 0 or p_len - len(f0) - pad_size > 0:
+                f0 = np.pad(
+                    f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant"
+                )
+        elif f0_method == "harvest":
+            f0, t = pyworld.harvest(
+                x.astype(np.double),
+                fs=self.fs,
+                f0_ceil=self.f0_max,
+                f0_floor=self.f0_min,
+                frame_period=1000 * self.hop / self.fs,
+            )
+            f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.fs)
+        elif f0_method == "dio":
+            f0, t = pyworld.dio(
+                x.astype(np.double),
+                fs=self.fs,
+                f0_ceil=self.f0_max,
+                f0_floor=self.f0_min,
+                frame_period=1000 * self.hop / self.fs,
+            )
+            f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.fs)
+        elif f0_method == "rmvpe":
+            if hasattr(self, "model_rmvpe") == False:
+                from infer.lib.rmvpe import RMVPE
+
+                print("Loading rmvpe model")
+                self.model_rmvpe = RMVPE(
+                    "assets/rmvpe/rmvpe.pt", is_half=False, device="cpu"
+                )
+            f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
+        return f0
+
+    def coarse_f0(self, f0):
+        f0_mel = 1127 * np.log(1 + f0 / 700)
+        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * (
+            self.f0_bin - 2
+        ) / (self.f0_mel_max - self.f0_mel_min) + 1
+
+        # use 0 or 1
+        f0_mel[f0_mel <= 1] = 1
+        f0_mel[f0_mel > self.f0_bin - 1] = self.f0_bin - 1
+        f0_coarse = np.rint(f0_mel).astype(int)
+        assert f0_coarse.max() <= 255 and f0_coarse.min() >= 1, (
+            f0_coarse.max(),
+            f0_coarse.min(),
+        )
+        return f0_coarse
+
+    def go(self, paths, f0_method):
+        if len(paths) == 0:
+            printt("no-f0-todo")
+        else:
+            printt("todo-f0-%s" % len(paths))
+            n = max(len(paths) // 5, 1)  # 每个进程最多打印5条
+            for idx, (inp_path, opt_path1, opt_path2) in enumerate(paths):
+                try:
+                    if idx % n == 0:
+                        printt("f0ing,now-%s,all-%s,-%s" % (idx, len(paths), inp_path))
+                    if (
+                        os.path.exists(opt_path1 + ".npy") == True
+                        and os.path.exists(opt_path2 + ".npy") == True
+                    ):
+                        continue
+                    featur_pit = self.compute_f0(inp_path, f0_method)
+                    np.save(
+                        opt_path2,
+                        featur_pit,
+                        allow_pickle=False,
+                    )  # nsf
+                    coarse_pit = self.coarse_f0(featur_pit)
+                    np.save(
+                        opt_path1,
+                        coarse_pit,
+                        allow_pickle=False,
+                    )  # ori
+                except:
+                    printt("f0fail-%s-%s-%s" % (idx, inp_path, traceback.format_exc()))
+
+
+if __name__ == "__main__":
+    # exp_dir=r"E:\codes\py39\dataset\mi-test"
+    # n_p=16
+    # f = open("%s/log_extract_f0.log"%exp_dir, "w")
+    printt(sys.argv)
+    featureInput = FeatureInput()
+    paths = []
+    inp_root = "%s/1_16k_wavs" % (exp_dir)
+    opt_root1 = "%s/2a_f0" % (exp_dir)
+    opt_root2 = "%s/2b-f0nsf" % (exp_dir)
+
+    os.makedirs(opt_root1, exist_ok=True)
+    os.makedirs(opt_root2, exist_ok=True)
+    for name in sorted(list(os.listdir(inp_root))):
+        inp_path = "%s/%s" % (inp_root, name)
+        if "spec" in inp_path:
+            continue
+        opt_path1 = "%s/%s" % (opt_root1, name)
+        opt_path2 = "%s/%s" % (opt_root2, name)
+        paths.append([inp_path, opt_path1, opt_path2])
+
+    ps = []
+    for i in range(n_p):
+        p = Process(
+            target=featureInput.go,
+            args=(
+                paths[i::n_p],
+                f0method,
+            ),
+        )
+        ps.append(p)
+        p.start()
+    for i in range(n_p):
+        ps[i].join()

infer/modules/train/extract/extract_f0_rmvpe.py

@@ -0,0 +1,141 @@
+import os
+import sys
+import traceback
+
+import parselmouth
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+import logging
+
+import numpy as np
+import pyworld
+
+from infer.lib.audio import load_audio
+
+logging.getLogger("numba").setLevel(logging.WARNING)
+
+n_part = int(sys.argv[1])
+i_part = int(sys.argv[2])
+i_gpu = sys.argv[3]
+os.environ["CUDA_VISIBLE_DEVICES"] = str(i_gpu)
+exp_dir = sys.argv[4]
+is_half = sys.argv[5]
+f = open("%s/extract_f0_feature.log" % exp_dir, "a+")
+
+
+def printt(strr):
+    print(strr)
+    f.write("%s\n" % strr)
+    f.flush()
+
+
+class FeatureInput(object):
+    def __init__(self, samplerate=16000, hop_size=160):
+        self.fs = samplerate
+        self.hop = hop_size
+
+        self.f0_bin = 256
+        self.f0_max = 1100.0
+        self.f0_min = 50.0
+        self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
+        self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
+
+    def compute_f0(self, path, f0_method):
+        x = load_audio(path, self.fs)
+        # p_len = x.shape[0] // self.hop
+        if f0_method == "rmvpe":
+            if hasattr(self, "model_rmvpe") == False:
+                from infer.lib.rmvpe import RMVPE
+
+                print("Loading rmvpe model")
+                self.model_rmvpe = RMVPE(
+                    "assets/rmvpe/rmvpe.pt", is_half=is_half, device="cuda"
+                )
+            f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
+        return f0
+
+    def coarse_f0(self, f0):
+        f0_mel = 1127 * np.log(1 + f0 / 700)
+        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * (
+            self.f0_bin - 2
+        ) / (self.f0_mel_max - self.f0_mel_min) + 1
+
+        # use 0 or 1
+        f0_mel[f0_mel <= 1] = 1
+        f0_mel[f0_mel > self.f0_bin - 1] = self.f0_bin - 1
+        f0_coarse = np.rint(f0_mel).astype(int)
+        assert f0_coarse.max() <= 255 and f0_coarse.min() >= 1, (
+            f0_coarse.max(),
+            f0_coarse.min(),
+        )
+        return f0_coarse
+
+    def go(self, paths, f0_method):
+        if len(paths) == 0:
+            printt("no-f0-todo")
+        else:
+            printt("todo-f0-%s" % len(paths))
+            n = max(len(paths) // 5, 1)  # 每个进程最多打印5条
+            for idx, (inp_path, opt_path1, opt_path2) in enumerate(paths):
+                try:
+                    if idx % n == 0:
+                        printt("f0ing,now-%s,all-%s,-%s" % (idx, len(paths), inp_path))
+                    if (
+                        os.path.exists(opt_path1 + ".npy") == True
+                        and os.path.exists(opt_path2 + ".npy") == True
+                    ):
+                        continue
+                    featur_pit = self.compute_f0(inp_path, f0_method)
+                    np.save(
+                        opt_path2,
+                        featur_pit,
+                        allow_pickle=False,
+                    )  # nsf
+                    coarse_pit = self.coarse_f0(featur_pit)
+                    np.save(
+                        opt_path1,
+                        coarse_pit,
+                        allow_pickle=False,
+                    )  # ori
+                except:
+                    printt("f0fail-%s-%s-%s" % (idx, inp_path, traceback.format_exc()))
+
+
+if __name__ == "__main__":
+    # exp_dir=r"E:\codes\py39\dataset\mi-test"
+    # n_p=16
+    # f = open("%s/log_extract_f0.log"%exp_dir, "w")
+    printt(sys.argv)
+    featureInput = FeatureInput()
+    paths = []
+    inp_root = "%s/1_16k_wavs" % (exp_dir)
+    opt_root1 = "%s/2a_f0" % (exp_dir)
+    opt_root2 = "%s/2b-f0nsf" % (exp_dir)
+
+    os.makedirs(opt_root1, exist_ok=True)
+    os.makedirs(opt_root2, exist_ok=True)
+    for name in sorted(list(os.listdir(inp_root))):
+        inp_path = "%s/%s" % (inp_root, name)
+        if "spec" in inp_path:
+            continue
+        opt_path1 = "%s/%s" % (opt_root1, name)
+        opt_path2 = "%s/%s" % (opt_root2, name)
+        paths.append([inp_path, opt_path1, opt_path2])
+    try:
+        featureInput.go(paths[i_part::n_part], "rmvpe")
+    except:
+        printt("f0_all_fail-%s" % (traceback.format_exc()))
+    # ps = []
+    # for i in range(n_p):
+    #     p = Process(
+    #         target=featureInput.go,
+    #         args=(
+    #             paths[i::n_p],
+    #             f0method,
+    #         ),
+    #     )
+    #     ps.append(p)
+    #     p.start()
+    # for i in range(n_p):
+    #     ps[i].join()

infer/modules/train/extract/extract_f0_rmvpe_dml.py

@@ -0,0 +1,139 @@
+import os
+import sys
+import traceback
+
+import parselmouth
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+import logging
+
+import numpy as np
+import pyworld
+
+from infer.lib.audio import load_audio
+
+logging.getLogger("numba").setLevel(logging.WARNING)
+
+exp_dir = sys.argv[1]
+import torch_directml
+
+device = torch_directml.device(torch_directml.default_device())
+f = open("%s/extract_f0_feature.log" % exp_dir, "a+")
+
+
+def printt(strr):
+    print(strr)
+    f.write("%s\n" % strr)
+    f.flush()
+
+
+class FeatureInput(object):
+    def __init__(self, samplerate=16000, hop_size=160):
+        self.fs = samplerate
+        self.hop = hop_size
+
+        self.f0_bin = 256
+        self.f0_max = 1100.0
+        self.f0_min = 50.0
+        self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
+        self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
+
+    def compute_f0(self, path, f0_method):
+        x = load_audio(path, self.fs)
+        # p_len = x.shape[0] // self.hop
+        if f0_method == "rmvpe":
+            if hasattr(self, "model_rmvpe") == False:
+                from infer.lib.rmvpe import RMVPE
+
+                print("Loading rmvpe model")
+                self.model_rmvpe = RMVPE(
+                    "assets/rmvpe/rmvpe.pt", is_half=False, device=device
+                )
+            f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
+        return f0
+
+    def coarse_f0(self, f0):
+        f0_mel = 1127 * np.log(1 + f0 / 700)
+        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * (
+            self.f0_bin - 2
+        ) / (self.f0_mel_max - self.f0_mel_min) + 1
+
+        # use 0 or 1
+        f0_mel[f0_mel <= 1] = 1
+        f0_mel[f0_mel > self.f0_bin - 1] = self.f0_bin - 1
+        f0_coarse = np.rint(f0_mel).astype(int)
+        assert f0_coarse.max() <= 255 and f0_coarse.min() >= 1, (
+            f0_coarse.max(),
+            f0_coarse.min(),
+        )
+        return f0_coarse
+
+    def go(self, paths, f0_method):
+        if len(paths) == 0:
+            printt("no-f0-todo")
+        else:
+            printt("todo-f0-%s" % len(paths))
+            n = max(len(paths) // 5, 1)  # 每个进程最多打印5条
+            for idx, (inp_path, opt_path1, opt_path2) in enumerate(paths):
+                try:
+                    if idx % n == 0:
+                        printt("f0ing,now-%s,all-%s,-%s" % (idx, len(paths), inp_path))
+                    if (
+                        os.path.exists(opt_path1 + ".npy") == True
+                        and os.path.exists(opt_path2 + ".npy") == True
+                    ):
+                        continue
+                    featur_pit = self.compute_f0(inp_path, f0_method)
+                    np.save(
+                        opt_path2,
+                        featur_pit,
+                        allow_pickle=False,
+                    )  # nsf
+                    coarse_pit = self.coarse_f0(featur_pit)
+                    np.save(
+                        opt_path1,
+                        coarse_pit,
+                        allow_pickle=False,
+                    )  # ori
+                except:
+                    printt("f0fail-%s-%s-%s" % (idx, inp_path, traceback.format_exc()))
+
+
+if __name__ == "__main__":
+    # exp_dir=r"E:\codes\py39\dataset\mi-test"
+    # n_p=16
+    # f = open("%s/log_extract_f0.log"%exp_dir, "w")
+    printt(sys.argv)
+    featureInput = FeatureInput()
+    paths = []
+    inp_root = "%s/1_16k_wavs" % (exp_dir)
+    opt_root1 = "%s/2a_f0" % (exp_dir)
+    opt_root2 = "%s/2b-f0nsf" % (exp_dir)
+
+    os.makedirs(opt_root1, exist_ok=True)
+    os.makedirs(opt_root2, exist_ok=True)
+    for name in sorted(list(os.listdir(inp_root))):
+        inp_path = "%s/%s" % (inp_root, name)
+        if "spec" in inp_path:
+            continue
+        opt_path1 = "%s/%s" % (opt_root1, name)
+        opt_path2 = "%s/%s" % (opt_root2, name)
+        paths.append([inp_path, opt_path1, opt_path2])
+    try:
+        featureInput.go(paths, "rmvpe")
+    except:
+        printt("f0_all_fail-%s" % (traceback.format_exc()))
+    # ps = []
+    # for i in range(n_p):
+    #     p = Process(
+    #         target=featureInput.go,
+    #         args=(
+    #             paths[i::n_p],
+    #             f0method,
+    #         ),
+    #     )
+    #     ps.append(p)
+    #     p.start()
+    # for i in range(n_p):
+    #     ps[i].join()

infer/modules/train/extract_feature_print.py

@@ -0,0 +1,137 @@
+import os
+import sys
+import traceback
+
+os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
+os.environ["PYTORCH_MPS_HIGH_WATERMARK_RATIO"] = "0.0"
+
+device = sys.argv[1]
+n_part = int(sys.argv[2])
+i_part = int(sys.argv[3])
+if len(sys.argv) == 6:
+    exp_dir = sys.argv[4]
+    version = sys.argv[5]
+else:
+    i_gpu = sys.argv[4]
+    exp_dir = sys.argv[5]
+    os.environ["CUDA_VISIBLE_DEVICES"] = str(i_gpu)
+    version = sys.argv[6]
+import fairseq
+import numpy as np
+import soundfile as sf
+import torch
+import torch.nn.functional as F
+
+if "privateuseone" not in device:
+    device = "cpu"
+    if torch.cuda.is_available():
+        device = "cuda"
+    elif torch.backends.mps.is_available():
+        device = "mps"
+else:
+    import torch_directml
+
+    device = torch_directml.device(torch_directml.default_device())
+
+    def forward_dml(ctx, x, scale):
+        ctx.scale = scale
+        res = x.clone().detach()
+        return res
+
+    fairseq.modules.grad_multiply.GradMultiply.forward = forward_dml
+
+f = open("%s/extract_f0_feature.log" % exp_dir, "a+")
+
+
+def printt(strr):
+    print(strr)
+    f.write("%s\n" % strr)
+    f.flush()
+
+
+printt(sys.argv)
+model_path = "assets/hubert/hubert_base.pt"
+
+printt(exp_dir)
+wavPath = "%s/1_16k_wavs" % exp_dir
+outPath = (
+    "%s/3_feature256" % exp_dir if version == "v1" else "%s/3_feature768" % exp_dir
+)
+os.makedirs(outPath, exist_ok=True)
+
+
+# wave must be 16k, hop_size=320
+def readwave(wav_path, normalize=False):
+    wav, sr = sf.read(wav_path)
+    assert sr == 16000
+    feats = torch.from_numpy(wav).float()
+    if feats.dim() == 2:  # double channels
+        feats = feats.mean(-1)
+    assert feats.dim() == 1, feats.dim()
+    if normalize:
+        with torch.no_grad():
+            feats = F.layer_norm(feats, feats.shape)
+    feats = feats.view(1, -1)
+    return feats
+
+
+# HuBERT model
+printt("load model(s) from {}".format(model_path))
+# if hubert model is exist
+if os.access(model_path, os.F_OK) == False:
+    printt(
+        "Error: Extracting is shut down because %s does not exist, you may download it from https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main"
+        % model_path
+    )
+    exit(0)
+models, saved_cfg, task = fairseq.checkpoint_utils.load_model_ensemble_and_task(
+    [model_path],
+    suffix="",
+)
+model = models[0]
+model = model.to(device)
+printt("move model to %s" % device)
+if device not in ["mps", "cpu"]:
+    model = model.half()
+model.eval()
+
+todo = sorted(list(os.listdir(wavPath)))[i_part::n_part]
+n = max(1, len(todo) // 10)  # 最多打印十条
+if len(todo) == 0:
+    printt("no-feature-todo")
+else:
+    printt("all-feature-%s" % len(todo))
+    for idx, file in enumerate(todo):
+        try:
+            if file.endswith(".wav"):
+                wav_path = "%s/%s" % (wavPath, file)
+                out_path = "%s/%s" % (outPath, file.replace("wav", "npy"))
+
+                if os.path.exists(out_path):
+                    continue
+
+                feats = readwave(wav_path, normalize=saved_cfg.task.normalize)
+                padding_mask = torch.BoolTensor(feats.shape).fill_(False)
+                inputs = {
+                    "source": feats.half().to(device)
+                    if device not in ["mps", "cpu"]
+                    else feats.to(device),
+                    "padding_mask": padding_mask.to(device),
+                    "output_layer": 9 if version == "v1" else 12,  # layer 9
+                }
+                with torch.no_grad():
+                    logits = model.extract_features(**inputs)
+                    feats = (
+                        model.final_proj(logits[0]) if version == "v1" else logits[0]
+                    )
+
+                feats = feats.squeeze(0).float().cpu().numpy()
+                if np.isnan(feats).sum() == 0:
+                    np.save(out_path, feats, allow_pickle=False)
+                else:
+                    printt("%s-contains nan" % file)
+                if idx % n == 0:
+                    printt("now-%s,all-%s,%s,%s" % (len(todo), idx, file, feats.shape))
+        except:
+            printt(traceback.format_exc())
+    printt("all-feature-done")

infer/modules/train/preprocess.py

@@ -0,0 +1,147 @@
+import multiprocessing
+import os
+import sys
+
+from scipy import signal
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+print(sys.argv)
+inp_root = sys.argv[1]
+sr = int(sys.argv[2])
+n_p = int(sys.argv[3])
+exp_dir = sys.argv[4]
+noparallel = sys.argv[5] == "True"
+per = float(sys.argv[6])
+import multiprocessing
+import os
+import traceback
+
+import librosa
+import numpy as np
+from scipy.io import wavfile
+
+from infer.lib.audio import load_audio
+from infer.lib.slicer2 import Slicer
+
+mutex = multiprocessing.Lock()
+f = open("%s/preprocess.log" % exp_dir, "a+")
+
+
+def println(strr):
+    mutex.acquire()
+    print(strr)
+    f.write("%s\n" % strr)
+    f.flush()
+    mutex.release()
+
+
+class PreProcess:
+    def __init__(self, sr, exp_dir, per=3.7):
+        self.slicer = Slicer(
+            sr=sr,
+            threshold=-42,
+            min_length=1500,
+            min_interval=400,
+            hop_size=15,
+            max_sil_kept=500,
+        )
+        self.sr = sr
+        self.bh, self.ah = signal.butter(N=5, Wn=48, btype="high", fs=self.sr)
+        self.per = per
+        self.overlap = 0.3
+        self.tail = self.per + self.overlap
+        self.max = 0.9
+        self.alpha = 0.75
+        self.exp_dir = exp_dir
+        self.gt_wavs_dir = "%s/0_gt_wavs" % exp_dir
+        self.wavs16k_dir = "%s/1_16k_wavs" % exp_dir
+        os.makedirs(self.exp_dir, exist_ok=True)
+        os.makedirs(self.gt_wavs_dir, exist_ok=True)
+        os.makedirs(self.wavs16k_dir, exist_ok=True)
+
+    def norm_write(self, tmp_audio, idx0, idx1):
+        tmp_max = np.abs(tmp_audio).max()
+        if tmp_max > 2.5:
+            print("%s-%s-%s-filtered" % (idx0, idx1, tmp_max))
+            return
+        tmp_audio = (tmp_audio / tmp_max * (self.max * self.alpha)) + (
+            1 - self.alpha
+        ) * tmp_audio
+        wavfile.write(
+            "%s/%s_%s.wav" % (self.gt_wavs_dir, idx0, idx1),
+            self.sr,
+            tmp_audio.astype(np.float32),
+        )
+        tmp_audio = librosa.resample(
+            tmp_audio, orig_sr=self.sr, target_sr=16000
+        )  # , res_type="soxr_vhq"
+        wavfile.write(
+            "%s/%s_%s.wav" % (self.wavs16k_dir, idx0, idx1),
+            16000,
+            tmp_audio.astype(np.float32),
+        )
+
+    def pipeline(self, path, idx0):
+        try:
+            audio = load_audio(path, self.sr)
+            # zero phased digital filter cause pre-ringing noise...
+            # audio = signal.filtfilt(self.bh, self.ah, audio)
+            audio = signal.lfilter(self.bh, self.ah, audio)
+
+            idx1 = 0
+            for audio in self.slicer.slice(audio):
+                i = 0
+                while 1:
+                    start = int(self.sr * (self.per - self.overlap) * i)
+                    i += 1
+                    if len(audio[start:]) > self.tail * self.sr:
+                        tmp_audio = audio[start : start + int(self.per * self.sr)]
+                        self.norm_write(tmp_audio, idx0, idx1)
+                        idx1 += 1
+                    else:
+                        tmp_audio = audio[start:]
+                        idx1 += 1
+                        break
+                self.norm_write(tmp_audio, idx0, idx1)
+            println("%s->Suc." % path)
+        except:
+            println("%s->%s" % (path, traceback.format_exc()))
+
+    def pipeline_mp(self, infos):
+        for path, idx0 in infos:
+            self.pipeline(path, idx0)
+
+    def pipeline_mp_inp_dir(self, inp_root, n_p):
+        try:
+            infos = [
+                ("%s/%s" % (inp_root, name), idx)
+                for idx, name in enumerate(sorted(list(os.listdir(inp_root))))
+            ]
+            if noparallel:
+                for i in range(n_p):
+                    self.pipeline_mp(infos[i::n_p])
+            else:
+                ps = []
+                for i in range(n_p):
+                    p = multiprocessing.Process(
+                        target=self.pipeline_mp, args=(infos[i::n_p],)
+                    )
+                    ps.append(p)
+                    p.start()
+                for i in range(n_p):
+                    ps[i].join()
+        except:
+            println("Fail. %s" % traceback.format_exc())
+
+
+def preprocess_trainset(inp_root, sr, n_p, exp_dir, per):
+    pp = PreProcess(sr, exp_dir, per)
+    println("start preprocess")
+    println(sys.argv)
+    pp.pipeline_mp_inp_dir(inp_root, n_p)
+    println("end preprocess")
+
+
+if __name__ == "__main__":
+    preprocess_trainset(inp_root, sr, n_p, exp_dir, per)

infer/modules/train/train.py

@@ -0,0 +1,643 @@
+import os
+import sys
+import logging
+
+logger = logging.getLogger(__name__)
+
+now_dir = os.getcwd()
+sys.path.append(os.path.join(now_dir))
+
+import datetime
+
+from infer.lib.train import utils
+
+hps = utils.get_hparams()
+os.environ["CUDA_VISIBLE_DEVICES"] = hps.gpus.replace("-", ",")
+n_gpus = len(hps.gpus.split("-"))
+from random import randint, shuffle
+
+import torch
+
+try:
+    import intel_extension_for_pytorch as ipex  # pylint: disable=import-error, unused-import
+
+    if torch.xpu.is_available():
+        from infer.modules.ipex import ipex_init
+        from infer.modules.ipex.gradscaler import gradscaler_init
+        from torch.xpu.amp import autocast
+
+        GradScaler = gradscaler_init()
+        ipex_init()
+    else:
+        from torch.cuda.amp import GradScaler, autocast
+except Exception:
+    from torch.cuda.amp import GradScaler, autocast
+
+torch.backends.cudnn.deterministic = False
+torch.backends.cudnn.benchmark = False
+from time import sleep
+from time import time as ttime
+
+import torch.distributed as dist
+import torch.multiprocessing as mp
+from torch.nn import functional as F
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.data import DataLoader
+from torch.utils.tensorboard import SummaryWriter
+
+from infer.lib.infer_pack import commons
+from infer.lib.train.data_utils import (
+    DistributedBucketSampler,
+    TextAudioCollate,
+    TextAudioCollateMultiNSFsid,
+    TextAudioLoader,
+    TextAudioLoaderMultiNSFsid,
+)
+
+if hps.version == "v1":
+    from infer.lib.infer_pack.models import MultiPeriodDiscriminator
+    from infer.lib.infer_pack.models import SynthesizerTrnMs256NSFsid as RVC_Model_f0
+    from infer.lib.infer_pack.models import (
+        SynthesizerTrnMs256NSFsid_nono as RVC_Model_nof0,
+    )
+else:
+    from infer.lib.infer_pack.models import (
+        SynthesizerTrnMs768NSFsid as RVC_Model_f0,
+        SynthesizerTrnMs768NSFsid_nono as RVC_Model_nof0,
+        MultiPeriodDiscriminatorV2 as MultiPeriodDiscriminator,
+    )
+
+from infer.lib.train.losses import (
+    discriminator_loss,
+    feature_loss,
+    generator_loss,
+    kl_loss,
+)
+from infer.lib.train.mel_processing import mel_spectrogram_torch, spec_to_mel_torch
+from infer.lib.train.process_ckpt import savee
+
+global_step = 0
+
+
+class EpochRecorder:
+    def __init__(self):
+        self.last_time = ttime()
+
+    def record(self):
+        now_time = ttime()
+        elapsed_time = now_time - self.last_time
+        self.last_time = now_time
+        elapsed_time_str = str(datetime.timedelta(seconds=elapsed_time))
+        current_time = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+        return f"[{current_time}] | ({elapsed_time_str})"
+
+
+def main():
+    n_gpus = torch.cuda.device_count()
+
+    if torch.cuda.is_available() == False and torch.backends.mps.is_available() == True:
+        n_gpus = 1
+    if n_gpus < 1:
+        # patch to unblock people without gpus. there is probably a better way.
+        print("NO GPU DETECTED: falling back to CPU - this may take a while")
+        n_gpus = 1
+    os.environ["MASTER_ADDR"] = "localhost"
+    os.environ["MASTER_PORT"] = str(randint(20000, 55555))
+    children = []
+    for i in range(n_gpus):
+        subproc = mp.Process(
+            target=run,
+            args=(i, n_gpus, hps),
+        )
+        children.append(subproc)
+        subproc.start()
+
+    for i in range(n_gpus):
+        children[i].join()
+
+
+def run(
+    rank,
+    n_gpus,
+    hps,
+):
+    global global_step
+    if rank == 0:
+        logger = utils.get_logger(hps.model_dir)
+        logger.info(hps)
+        # utils.check_git_hash(hps.model_dir)
+        writer = SummaryWriter(log_dir=hps.model_dir)
+        writer_eval = SummaryWriter(log_dir=os.path.join(hps.model_dir, "eval"))
+
+    dist.init_process_group(
+        backend="gloo", init_method="env://", world_size=n_gpus, rank=rank
+    )
+    torch.manual_seed(hps.train.seed)
+    if torch.cuda.is_available():
+        torch.cuda.set_device(rank)
+
+    if hps.if_f0 == 1:
+        train_dataset = TextAudioLoaderMultiNSFsid(hps.data.training_files, hps.data)
+    else:
+        train_dataset = TextAudioLoader(hps.data.training_files, hps.data)
+    train_sampler = DistributedBucketSampler(
+        train_dataset,
+        hps.train.batch_size * n_gpus,
+        # [100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 1200,1400],  # 16s
+        [100, 200, 300, 400, 500, 600, 700, 800, 900],  # 16s
+        num_replicas=n_gpus,
+        rank=rank,
+        shuffle=True,
+    )
+    # It is possible that dataloader's workers are out of shared memory. Please try to raise your shared memory limit.
+    # num_workers=8 -> num_workers=4
+    if hps.if_f0 == 1:
+        collate_fn = TextAudioCollateMultiNSFsid()
+    else:
+        collate_fn = TextAudioCollate()
+    train_loader = DataLoader(
+        train_dataset,
+        num_workers=4,
+        shuffle=False,
+        pin_memory=True,
+        collate_fn=collate_fn,
+        batch_sampler=train_sampler,
+        persistent_workers=True,
+        prefetch_factor=8,
+    )
+    if hps.if_f0 == 1:
+        net_g = RVC_Model_f0(
+            hps.data.filter_length // 2 + 1,
+            hps.train.segment_size // hps.data.hop_length,
+            **hps.model,
+            is_half=hps.train.fp16_run,
+            sr=hps.sample_rate,
+        )
+    else:
+        net_g = RVC_Model_nof0(
+            hps.data.filter_length // 2 + 1,
+            hps.train.segment_size // hps.data.hop_length,
+            **hps.model,
+            is_half=hps.train.fp16_run,
+        )
+    if torch.cuda.is_available():
+        net_g = net_g.cuda(rank)
+    net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm)
+    if torch.cuda.is_available():
+        net_d = net_d.cuda(rank)
+    optim_g = torch.optim.AdamW(
+        net_g.parameters(),
+        hps.train.learning_rate,
+        betas=hps.train.betas,
+        eps=hps.train.eps,
+    )
+    optim_d = torch.optim.AdamW(
+        net_d.parameters(),
+        hps.train.learning_rate,
+        betas=hps.train.betas,
+        eps=hps.train.eps,
+    )
+    # net_g = DDP(net_g, device_ids=[rank], find_unused_parameters=True)
+    # net_d = DDP(net_d, device_ids=[rank], find_unused_parameters=True)
+    if hasattr(torch, "xpu") and torch.xpu.is_available():
+        pass
+    elif torch.cuda.is_available():
+        net_g = DDP(net_g, device_ids=[rank])
+        net_d = DDP(net_d, device_ids=[rank])
+    else:
+        net_g = DDP(net_g)
+        net_d = DDP(net_d)
+
+    try:  # 如果能加载自动resume
+        _, _, _, epoch_str = utils.load_checkpoint(
+            utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d, optim_d
+        )  # D多半加载没事
+        if rank == 0:
+            logger.info("loaded D")
+        # _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g,load_opt=0)
+        _, _, _, epoch_str = utils.load_checkpoint(
+            utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g
+        )
+        global_step = (epoch_str - 1) * len(train_loader)
+        # epoch_str = 1
+        # global_step = 0
+    except:  # 如果首次不能加载，加载pretrain
+        # traceback.print_exc()
+        epoch_str = 1
+        global_step = 0
+        if hps.pretrainG != "":
+            if rank == 0:
+                logger.info("loaded pretrained %s" % (hps.pretrainG))
+            if hasattr(net_g, "module"):
+                logger.info(
+                    net_g.module.load_state_dict(
+                        torch.load(hps.pretrainG, map_location="cpu")["model"]
+                    )
+                )  ##测试不加载优化器
+            else:
+                logger.info(
+                    net_g.load_state_dict(
+                        torch.load(hps.pretrainG, map_location="cpu")["model"]
+                    )
+                )  ##测试不加载优化器
+        if hps.pretrainD != "":
+            if rank == 0:
+                logger.info("loaded pretrained %s" % (hps.pretrainD))
+            if hasattr(net_d, "module"):
+                logger.info(
+                    net_d.module.load_state_dict(
+                        torch.load(hps.pretrainD, map_location="cpu")["model"]
+                    )
+                )
+            else:
+                logger.info(
+                    net_d.load_state_dict(
+                        torch.load(hps.pretrainD, map_location="cpu")["model"]
+                    )
+                )
+
+    scheduler_g = torch.optim.lr_scheduler.ExponentialLR(
+        optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
+    )
+    scheduler_d = torch.optim.lr_scheduler.ExponentialLR(
+        optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
+    )
+
+    scaler = GradScaler(enabled=hps.train.fp16_run)
+
+    cache = []
+    for epoch in range(epoch_str, hps.train.epochs + 1):
+        if rank == 0:
+            train_and_evaluate(
+                rank,
+                epoch,
+                hps,
+                [net_g, net_d],
+                [optim_g, optim_d],
+                [scheduler_g, scheduler_d],
+                scaler,
+                [train_loader, None],
+                logger,
+                [writer, writer_eval],
+                cache,
+            )
+        else:
+            train_and_evaluate(
+                rank,
+                epoch,
+                hps,
+                [net_g, net_d],
+                [optim_g, optim_d],
+                [scheduler_g, scheduler_d],
+                scaler,
+                [train_loader, None],
+                None,
+                None,
+                cache,
+            )
+        scheduler_g.step()
+        scheduler_d.step()
+
+
+def train_and_evaluate(
+    rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers, cache
+):
+    net_g, net_d = nets
+    optim_g, optim_d = optims
+    train_loader, eval_loader = loaders
+    if writers is not None:
+        writer, writer_eval = writers
+
+    train_loader.batch_sampler.set_epoch(epoch)
+    global global_step
+
+    net_g.train()
+    net_d.train()
+
+    # Prepare data iterator
+    if hps.if_cache_data_in_gpu == True:
+        # Use Cache
+        data_iterator = cache
+        if cache == []:
+            # Make new cache
+            for batch_idx, info in enumerate(train_loader):
+                # Unpack
+                if hps.if_f0 == 1:
+                    (
+                        phone,
+                        phone_lengths,
+                        pitch,
+                        pitchf,
+                        spec,
+                        spec_lengths,
+                        wave,
+                        wave_lengths,
+                        sid,
+                    ) = info
+                else:
+                    (
+                        phone,
+                        phone_lengths,
+                        spec,
+                        spec_lengths,
+                        wave,
+                        wave_lengths,
+                        sid,
+                    ) = info
+                # Load on CUDA
+                if torch.cuda.is_available():
+                    phone = phone.cuda(rank, non_blocking=True)
+                    phone_lengths = phone_lengths.cuda(rank, non_blocking=True)
+                    if hps.if_f0 == 1:
+                        pitch = pitch.cuda(rank, non_blocking=True)
+                        pitchf = pitchf.cuda(rank, non_blocking=True)
+                    sid = sid.cuda(rank, non_blocking=True)
+                    spec = spec.cuda(rank, non_blocking=True)
+                    spec_lengths = spec_lengths.cuda(rank, non_blocking=True)
+                    wave = wave.cuda(rank, non_blocking=True)
+                    wave_lengths = wave_lengths.cuda(rank, non_blocking=True)
+                # Cache on list
+                if hps.if_f0 == 1:
+                    cache.append(
+                        (
+                            batch_idx,
+                            (
+                                phone,
+                                phone_lengths,
+                                pitch,
+                                pitchf,
+                                spec,
+                                spec_lengths,
+                                wave,
+                                wave_lengths,
+                                sid,
+                            ),
+                        )
+                    )
+                else:
+                    cache.append(
+                        (
+                            batch_idx,
+                            (
+                                phone,
+                                phone_lengths,
+                                spec,
+                                spec_lengths,
+                                wave,
+                                wave_lengths,
+                                sid,
+                            ),
+                        )
+                    )
+        else:
+            # Load shuffled cache
+            shuffle(cache)
+    else:
+        # Loader
+        data_iterator = enumerate(train_loader)
+
+    # Run steps
+    epoch_recorder = EpochRecorder()
+    for batch_idx, info in data_iterator:
+        # Data
+        ## Unpack
+        if hps.if_f0 == 1:
+            (
+                phone,
+                phone_lengths,
+                pitch,
+                pitchf,
+                spec,
+                spec_lengths,
+                wave,
+                wave_lengths,
+                sid,
+            ) = info
+        else:
+            phone, phone_lengths, spec, spec_lengths, wave, wave_lengths, sid = info
+        ## Load on CUDA
+        if (hps.if_cache_data_in_gpu == False) and torch.cuda.is_available():
+            phone = phone.cuda(rank, non_blocking=True)
+            phone_lengths = phone_lengths.cuda(rank, non_blocking=True)
+            if hps.if_f0 == 1:
+                pitch = pitch.cuda(rank, non_blocking=True)
+                pitchf = pitchf.cuda(rank, non_blocking=True)
+            sid = sid.cuda(rank, non_blocking=True)
+            spec = spec.cuda(rank, non_blocking=True)
+            spec_lengths = spec_lengths.cuda(rank, non_blocking=True)
+            wave = wave.cuda(rank, non_blocking=True)
+            # wave_lengths = wave_lengths.cuda(rank, non_blocking=True)
+
+        # Calculate
+        with autocast(enabled=hps.train.fp16_run):
+            if hps.if_f0 == 1:
+                (
+                    y_hat,
+                    ids_slice,
+                    x_mask,
+                    z_mask,
+                    (z, z_p, m_p, logs_p, m_q, logs_q),
+                ) = net_g(phone, phone_lengths, pitch, pitchf, spec, spec_lengths, sid)
+            else:
+                (
+                    y_hat,
+                    ids_slice,
+                    x_mask,
+                    z_mask,
+                    (z, z_p, m_p, logs_p, m_q, logs_q),
+                ) = net_g(phone, phone_lengths, spec, spec_lengths, sid)
+            mel = spec_to_mel_torch(
+                spec,
+                hps.data.filter_length,
+                hps.data.n_mel_channels,
+                hps.data.sampling_rate,
+                hps.data.mel_fmin,
+                hps.data.mel_fmax,
+            )
+            y_mel = commons.slice_segments(
+                mel, ids_slice, hps.train.segment_size // hps.data.hop_length
+            )
+            with autocast(enabled=False):
+                y_hat_mel = mel_spectrogram_torch(
+                    y_hat.float().squeeze(1),
+                    hps.data.filter_length,
+                    hps.data.n_mel_channels,
+                    hps.data.sampling_rate,
+                    hps.data.hop_length,
+                    hps.data.win_length,
+                    hps.data.mel_fmin,
+                    hps.data.mel_fmax,
+                )
+            if hps.train.fp16_run == True:
+                y_hat_mel = y_hat_mel.half()
+            wave = commons.slice_segments(
+                wave, ids_slice * hps.data.hop_length, hps.train.segment_size
+            )  # slice
+
+            # Discriminator
+            y_d_hat_r, y_d_hat_g, _, _ = net_d(wave, y_hat.detach())
+            with autocast(enabled=False):
+                loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(
+                    y_d_hat_r, y_d_hat_g
+                )
+        optim_d.zero_grad()
+        scaler.scale(loss_disc).backward()
+        scaler.unscale_(optim_d)
+        grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None)
+        scaler.step(optim_d)
+
+        with autocast(enabled=hps.train.fp16_run):
+            # Generator
+            y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(wave, y_hat)
+            with autocast(enabled=False):
+                loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
+                loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl
+                loss_fm = feature_loss(fmap_r, fmap_g)
+                loss_gen, losses_gen = generator_loss(y_d_hat_g)
+                loss_gen_all = loss_gen + loss_fm + loss_mel + loss_kl
+        optim_g.zero_grad()
+        scaler.scale(loss_gen_all).backward()
+        scaler.unscale_(optim_g)
+        grad_norm_g = commons.clip_grad_value_(net_g.parameters(), None)
+        scaler.step(optim_g)
+        scaler.update()
+
+        if rank == 0:
+            if global_step % hps.train.log_interval == 0:
+                lr = optim_g.param_groups[0]["lr"]
+                logger.info(
+                    "Train Epoch: {} [{:.0f}%]".format(
+                        epoch, 100.0 * batch_idx / len(train_loader)
+                    )
+                )
+                # Amor For Tensorboard display
+                if loss_mel > 75:
+                    loss_mel = 75
+                if loss_kl > 9:
+                    loss_kl = 9
+
+                logger.info([global_step, lr])
+                logger.info(
+                    f"loss_disc={loss_disc:.3f}, loss_gen={loss_gen:.3f}, loss_fm={loss_fm:.3f},loss_mel={loss_mel:.3f}, loss_kl={loss_kl:.3f}"
+                )
+                scalar_dict = {
+                    "loss/g/total": loss_gen_all,
+                    "loss/d/total": loss_disc,
+                    "learning_rate": lr,
+                    "grad_norm_d": grad_norm_d,
+                    "grad_norm_g": grad_norm_g,
+                }
+                scalar_dict.update(
+                    {
+                        "loss/g/fm": loss_fm,
+                        "loss/g/mel": loss_mel,
+                        "loss/g/kl": loss_kl,
+                    }
+                )
+
+                scalar_dict.update(
+                    {"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)}
+                )
+                scalar_dict.update(
+                    {"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)}
+                )
+                scalar_dict.update(
+                    {"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)}
+                )
+                image_dict = {
+                    "slice/mel_org": utils.plot_spectrogram_to_numpy(
+                        y_mel[0].data.cpu().numpy()
+                    ),
+                    "slice/mel_gen": utils.plot_spectrogram_to_numpy(
+                        y_hat_mel[0].data.cpu().numpy()
+                    ),
+                    "all/mel": utils.plot_spectrogram_to_numpy(
+                        mel[0].data.cpu().numpy()
+                    ),
+                }
+                utils.summarize(
+                    writer=writer,
+                    global_step=global_step,
+                    images=image_dict,
+                    scalars=scalar_dict,
+                )
+        global_step += 1
+    # /Run steps
+
+    if epoch % hps.save_every_epoch == 0 and rank == 0:
+        if hps.if_latest == 0:
+            utils.save_checkpoint(
+                net_g,
+                optim_g,
+                hps.train.learning_rate,
+                epoch,
+                os.path.join(hps.model_dir, "G_{}.pth".format(global_step)),
+            )
+            utils.save_checkpoint(
+                net_d,
+                optim_d,
+                hps.train.learning_rate,
+                epoch,
+                os.path.join(hps.model_dir, "D_{}.pth".format(global_step)),
+            )
+        else:
+            utils.save_checkpoint(
+                net_g,
+                optim_g,
+                hps.train.learning_rate,
+                epoch,
+                os.path.join(hps.model_dir, "G_{}.pth".format(2333333)),
+            )
+            utils.save_checkpoint(
+                net_d,
+                optim_d,
+                hps.train.learning_rate,
+                epoch,
+                os.path.join(hps.model_dir, "D_{}.pth".format(2333333)),
+            )
+        if rank == 0 and hps.save_every_weights == "1":
+            if hasattr(net_g, "module"):
+                ckpt = net_g.module.state_dict()
+            else:
+                ckpt = net_g.state_dict()
+            logger.info(
+                "saving ckpt %s_e%s:%s"
+                % (
+                    hps.name,
+                    epoch,
+                    savee(
+                        ckpt,
+                        hps.sample_rate,
+                        hps.if_f0,
+                        hps.name + "_e%s_s%s" % (epoch, global_step),
+                        epoch,
+                        hps.version,
+                        hps,
+                    ),
+                )
+            )
+
+    if rank == 0:
+        logger.info("====> Epoch: {} {}".format(epoch, epoch_recorder.record()))
+    if epoch >= hps.total_epoch and rank == 0:
+        logger.info("Training is done. The program is closed.")
+
+        if hasattr(net_g, "module"):
+            ckpt = net_g.module.state_dict()
+        else:
+            ckpt = net_g.state_dict()
+        logger.info(
+            "saving final ckpt:%s"
+            % (
+                savee(
+                    ckpt, hps.sample_rate, hps.if_f0, hps.name, epoch, hps.version, hps
+                )
+            )
+        )
+        sleep(1)
+        os._exit(2333333)
+
+
+if __name__ == "__main__":
+    torch.multiprocessing.set_start_method("spawn")
+    main()

infer/modules/uvr5/.ipynb_checkpoints/vr-checkpoint.py

@@ -0,0 +1,195 @@
+import os
+import logging
+
+logger = logging.getLogger(__name__)
+
+import librosa
+import numpy as np
+import soundfile as sf
+import torch
+
+from infer.lib.uvr5_pack.lib_v5 import nets_61968KB as Nets
+from infer.lib.uvr5_pack.lib_v5 import spec_utils
+from infer.lib.uvr5_pack.lib_v5.model_param_init import ModelParameters
+from infer.lib.uvr5_pack.lib_v5.nets_new import CascadedNet
+from infer.lib.uvr5_pack.utils import inference
+
+
+class AudioPre:
+    def __init__(self, agg, model_path, device, is_half, tta=False):
+        self.model_path = model_path
+        self.device = device
+        self.data = {
+            # Processing Options
+            "postprocess": False,
+            "tta": tta,
+            # Constants
+            "window_size": 512,
+            "agg": agg,
+            "high_end_process": "mirroring",
+        }
+        mp = ModelParameters("infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2.json")
+        model = Nets.CascadedASPPNet(mp.param["bins"] * 2)
+        cpk = torch.load(model_path, map_location="cpu")
+        model.load_state_dict(cpk)
+        model.eval()
+        if is_half:
+            model = model.half().to(device)
+        else:
+            model = model.to(device)
+
+        self.mp = mp
+        self.model = model
+
+    def _path_audio_(
+        self, music_file, ins_root=None, vocal_root=None, format="flac", is_hp3=False
+    ):
+        if ins_root is None and vocal_root is None:
+            return "No save root."
+        name = os.path.basename(music_file)
+        if ins_root is not None:
+            os.makedirs(ins_root, exist_ok=True)
+        if vocal_root is not None:
+            os.makedirs(vocal_root, exist_ok=True)
+        X_wave, y_wave, X_spec_s, y_spec_s = {}, {}, {}, {}
+        bands_n = len(self.mp.param["band"])
+        # print(bands_n)
+        for d in range(bands_n, 0, -1):
+            bp = self.mp.param["band"][d]
+            if d == bands_n:  # high-end band
+                (
+                    X_wave[d],
+                    _,
+                ) = librosa.core.load(  # 理论上librosa读取可能对某些音频有bug，应该上ffmpeg读取，但是太麻烦了弃坑
+                    music_file,
+                    bp["sr"],
+                    False,
+                    dtype=np.float32,
+                    res_type=bp["res_type"],
+                )
+                if X_wave[d].ndim == 1:
+                    X_wave[d] = np.asfortranarray([X_wave[d], X_wave[d]])
+            else:  # lower bands
+                X_wave[d] = librosa.core.resample(
+                    X_wave[d + 1],
+                    self.mp.param["band"][d + 1]["sr"],
+                    bp["sr"],
+                    res_type=bp["res_type"],
+                )
+            # Stft of wave source
+            X_spec_s[d] = spec_utils.wave_to_spectrogram_mt(
+                X_wave[d],
+                bp["hl"],
+                bp["n_fft"],
+                self.mp.param["mid_side"],
+                self.mp.param["mid_side_b2"],
+                self.mp.param["reverse"],
+            )
+            # pdb.set_trace()
+            if d == bands_n and self.data["high_end_process"] != "none":
+                input_high_end_h = (bp["n_fft"] // 2 - bp["crop_stop"]) + (
+                    self.mp.param["pre_filter_stop"] - self.mp.param["pre_filter_start"]
+                )
+                input_high_end = X_spec_s[d][
+                    :, bp["n_fft"] // 2 - input_high_end_h : bp["n_fft"] // 2, :
+                ]
+
+        X_spec_m = spec_utils.combine_spectrograms(X_spec_s, self.mp)
+        aggresive_set = float(self.data["agg"] / 100)
+        aggressiveness = {
+            "value": aggresive_set,
+            "split_bin": self.mp.param["band"][1]["crop_stop"],
+        }
+        with torch.no_grad():
+            pred, X_mag, X_phase = inference(
+                X_spec_m, self.device, self.model, aggressiveness, self.data
+            )
+        # Postprocess
+        if self.data["postprocess"]:
+            pred_inv = np.clip(X_mag - pred, 0, np.inf)
+            pred = spec_utils.mask_silence(pred, pred_inv)
+        y_spec_m = pred * X_phase
+        v_spec_m = X_spec_m - y_spec_m
+
+        if ins_root is not None:
+            if self.data["high_end_process"].startswith("mirroring"):
+                input_high_end_ = spec_utils.mirroring(
+                    self.data["high_end_process"], y_spec_m, input_high_end, self.mp
+                )
+                wav_instrument = spec_utils.cmb_spectrogram_to_wave(
+                    y_spec_m, self.mp, input_high_end_h, input_high_end_
+                )
+            else:
+                wav_instrument = spec_utils.cmb_spectrogram_to_wave(y_spec_m, self.mp)
+            logger.info("%s instruments done" % name)
+            if is_hp3 == True:
+                head = "vocal_"
+            else:
+                head = "instrument_"
+            if format in ["wav", "flac"]:
+                sf.write(
+                    os.path.join(
+                        ins_root,
+                        head + "{}_{}.{}".format(name, self.data["agg"], format),
+                    ),
+                    (np.array(wav_instrument) * 32768).astype("int16"),
+                    self.mp.param["sr"],
+                )  #
+            else:
+                path = os.path.join(
+                    ins_root, head + "{}_{}.wav".format(name, self.data["agg"])
+                )
+                sf.write(
+                    path,
+                    (np.array(wav_instrument) * 32768).astype("int16"),
+                    self.mp.param["sr"],
+                )
+                if os.path.exists(path):
+                    opt_format_path = path[:-4] + ".%s" % format
+                    os.system("ffmpeg -i %s -vn %s -q:a 2 -y" % (path, opt_format_path))
+                    if os.path.exists(opt_format_path):
+                        try:
+                            os.remove(path)
+                        except:
+                            pass
+        if vocal_root is not None:
+            if is_hp3 == True:
+                head = "instrument_"
+            else:
+                head = "vocal_"
+            if self.data["high_end_process"].startswith("mirroring"):
+                input_high_end_ = spec_utils.mirroring(
+                    self.data["high_end_process"], v_spec_m, input_high_end, self.mp
+                )
+                wav_vocals = spec_utils.cmb_spectrogram_to_wave(
+                    v_spec_m, self.mp, input_high_end_h, input_high_end_
+                )
+            else:
+                wav_vocals = spec_utils.cmb_spectrogram_to_wave(v_spec_m, self.mp)
+            logger.info("%s vocals done" % name)
+            if format in ["wav", "flac"]:
+                sf.write(
+                    os.path.join(
+                        vocal_root,
+                        head + "{}_{}.{}".format(name, self.data["agg"], format),
+                    ),
+                    (np.array(wav_vocals) * 32768).astype("int16"),
+                    self.mp.param["sr"],
+                )
+            else:
+                path = os.path.join(
+                    vocal_root, head + "{}_{}.wav".format(name, self.data["agg"])
+                )
+                sf.write(
+                    path,
+                    (np.array(wav_vocals) * 32768).astype("int16"),
+                    self.mp.param["sr"],
+                )
+                if os.path.exists(path):
+                    opt_format_path = path[:-4] + ".%s" % format
+                    os.system("ffmpeg -i %s -vn %s -q:a 2 -y" % (path, opt_format_path))
+                    if os.path.exists(opt_format_path):
+                        try:
+                            os.remove(path)
+                        except:
+                            pass

infer/modules/uvr5/mdxnet.py

@@ -0,0 +1,256 @@
+import os
+import logging
+
+logger = logging.getLogger(__name__)
+
+import librosa
+import numpy as np
+import soundfile as sf
+import torch
+from tqdm import tqdm
+
+cpu = torch.device("cpu")
+
+
+class ConvTDFNetTrim:
+    def __init__(
+        self, device, model_name, target_name, L, dim_f, dim_t, n_fft, hop=1024
+    ):
+        super(ConvTDFNetTrim, self).__init__()
+
+        self.dim_f = dim_f
+        self.dim_t = 2**dim_t
+        self.n_fft = n_fft
+        self.hop = hop
+        self.n_bins = self.n_fft // 2 + 1
+        self.chunk_size = hop * (self.dim_t - 1)
+        self.window = torch.hann_window(window_length=self.n_fft, periodic=True).to(
+            device
+        )
+        self.target_name = target_name
+        self.blender = "blender" in model_name
+
+        self.dim_c = 4
+        out_c = self.dim_c * 4 if target_name == "*" else self.dim_c
+        self.freq_pad = torch.zeros(
+            [1, out_c, self.n_bins - self.dim_f, self.dim_t]
+        ).to(device)
+
+        self.n = L // 2
+
+    def stft(self, x):
+        x = x.reshape([-1, self.chunk_size])
+        x = torch.stft(
+            x,
+            n_fft=self.n_fft,
+            hop_length=self.hop,
+            window=self.window,
+            center=True,
+            return_complex=True,
+        )
+        x = torch.view_as_real(x)
+        x = x.permute([0, 3, 1, 2])
+        x = x.reshape([-1, 2, 2, self.n_bins, self.dim_t]).reshape(
+            [-1, self.dim_c, self.n_bins, self.dim_t]
+        )
+        return x[:, :, : self.dim_f]
+
+    def istft(self, x, freq_pad=None):
+        freq_pad = (
+            self.freq_pad.repeat([x.shape[0], 1, 1, 1])
+            if freq_pad is None
+            else freq_pad
+        )
+        x = torch.cat([x, freq_pad], -2)
+        c = 4 * 2 if self.target_name == "*" else 2
+        x = x.reshape([-1, c, 2, self.n_bins, self.dim_t]).reshape(
+            [-1, 2, self.n_bins, self.dim_t]
+        )
+        x = x.permute([0, 2, 3, 1])
+        x = x.contiguous()
+        x = torch.view_as_complex(x)
+        x = torch.istft(
+            x, n_fft=self.n_fft, hop_length=self.hop, window=self.window, center=True
+        )
+        return x.reshape([-1, c, self.chunk_size])
+
+
+def get_models(device, dim_f, dim_t, n_fft):
+    return ConvTDFNetTrim(
+        device=device,
+        model_name="Conv-TDF",
+        target_name="vocals",
+        L=11,
+        dim_f=dim_f,
+        dim_t=dim_t,
+        n_fft=n_fft,
+    )
+
+
+class Predictor:
+    def __init__(self, args):
+        import onnxruntime as ort
+
+        logger.info(ort.get_available_providers())
+        self.args = args
+        self.model_ = get_models(
+            device=cpu, dim_f=args.dim_f, dim_t=args.dim_t, n_fft=args.n_fft
+        )
+        self.model = ort.InferenceSession(
+            os.path.join(args.onnx, self.model_.target_name + ".onnx"),
+            providers=[
+                "CUDAExecutionProvider",
+                "DmlExecutionProvider",
+                "CPUExecutionProvider",
+            ],
+        )
+        logger.info("ONNX load done")
+
+    def demix(self, mix):
+        samples = mix.shape[-1]
+        margin = self.args.margin
+        chunk_size = self.args.chunks * 44100
+        assert not margin == 0, "margin cannot be zero!"
+        if margin > chunk_size:
+            margin = chunk_size
+
+        segmented_mix = {}
+
+        if self.args.chunks == 0 or samples < chunk_size:
+            chunk_size = samples
+
+        counter = -1
+        for skip in range(0, samples, chunk_size):
+            counter += 1
+
+            s_margin = 0 if counter == 0 else margin
+            end = min(skip + chunk_size + margin, samples)
+
+            start = skip - s_margin
+
+            segmented_mix[skip] = mix[:, start:end].copy()
+            if end == samples:
+                break
+
+        sources = self.demix_base(segmented_mix, margin_size=margin)
+        """
+        mix:(2,big_sample)
+        segmented_mix:offset->(2,small_sample)
+        sources:(1,2,big_sample)
+        """
+        return sources
+
+    def demix_base(self, mixes, margin_size):
+        chunked_sources = []
+        progress_bar = tqdm(total=len(mixes))
+        progress_bar.set_description("Processing")
+        for mix in mixes:
+            cmix = mixes[mix]
+            sources = []
+            n_sample = cmix.shape[1]
+            model = self.model_
+            trim = model.n_fft // 2
+            gen_size = model.chunk_size - 2 * trim
+            pad = gen_size - n_sample % gen_size
+            mix_p = np.concatenate(
+                (np.zeros((2, trim)), cmix, np.zeros((2, pad)), np.zeros((2, trim))), 1
+            )
+            mix_waves = []
+            i = 0
+            while i < n_sample + pad:
+                waves = np.array(mix_p[:, i : i + model.chunk_size])
+                mix_waves.append(waves)
+                i += gen_size
+            mix_waves = torch.tensor(mix_waves, dtype=torch.float32).to(cpu)
+            with torch.no_grad():
+                _ort = self.model
+                spek = model.stft(mix_waves)
+                if self.args.denoise:
+                    spec_pred = (
+                        -_ort.run(None, {"input": -spek.cpu().numpy()})[0] * 0.5
+                        + _ort.run(None, {"input": spek.cpu().numpy()})[0] * 0.5
+                    )
+                    tar_waves = model.istft(torch.tensor(spec_pred))
+                else:
+                    tar_waves = model.istft(
+                        torch.tensor(_ort.run(None, {"input": spek.cpu().numpy()})[0])
+                    )
+                tar_signal = (
+                    tar_waves[:, :, trim:-trim]
+                    .transpose(0, 1)
+                    .reshape(2, -1)
+                    .numpy()[:, :-pad]
+                )
+
+                start = 0 if mix == 0 else margin_size
+                end = None if mix == list(mixes.keys())[::-1][0] else -margin_size
+                if margin_size == 0:
+                    end = None
+                sources.append(tar_signal[:, start:end])
+
+                progress_bar.update(1)
+
+            chunked_sources.append(sources)
+        _sources = np.concatenate(chunked_sources, axis=-1)
+        # del self.model
+        progress_bar.close()
+        return _sources
+
+    def prediction(self, m, vocal_root, others_root, format):
+        os.makedirs(vocal_root, exist_ok=True)
+        os.makedirs(others_root, exist_ok=True)
+        basename = os.path.basename(m)
+        mix, rate = librosa.load(m, mono=False, sr=44100)
+        if mix.ndim == 1:
+            mix = np.asfortranarray([mix, mix])
+        mix = mix.T
+        sources = self.demix(mix.T)
+        opt = sources[0].T
+        if format in ["wav", "flac"]:
+            sf.write(
+                "%s/%s_main_vocal.%s" % (vocal_root, basename, format), mix - opt, rate
+            )
+            sf.write("%s/%s_others.%s" % (others_root, basename, format), opt, rate)
+        else:
+            path_vocal = "%s/%s_main_vocal.wav" % (vocal_root, basename)
+            path_other = "%s/%s_others.wav" % (others_root, basename)
+            sf.write(path_vocal, mix - opt, rate)
+            sf.write(path_other, opt, rate)
+            opt_path_vocal = path_vocal[:-4] + ".%s" % format
+            opt_path_other = path_other[:-4] + ".%s" % format
+            if os.path.exists(path_vocal):
+                os.system(
+                    "ffmpeg -i %s -vn %s -q:a 2 -y" % (path_vocal, opt_path_vocal)
+                )
+                if os.path.exists(opt_path_vocal):
+                    try:
+                        os.remove(path_vocal)
+                    except:
+                        pass
+            if os.path.exists(path_other):
+                os.system(
+                    "ffmpeg -i %s -vn %s -q:a 2 -y" % (path_other, opt_path_other)
+                )
+                if os.path.exists(opt_path_other):
+                    try:
+                        os.remove(path_other)
+                    except:
+                        pass
+
+
+class MDXNetDereverb:
+    def __init__(self, chunks, device):
+        self.onnx = "assets/uvr5_weights/onnx_dereverb_By_FoxJoy"
+        self.shifts = 10  # 'Predict with randomised equivariant stabilisation'
+        self.mixing = "min_mag"  # ['default','min_mag','max_mag']
+        self.chunks = chunks
+        self.margin = 44100
+        self.dim_t = 9
+        self.dim_f = 3072
+        self.n_fft = 6144
+        self.denoise = True
+        self.pred = Predictor(self)
+        self.device = device
+
+    def _path_audio_(self, input, vocal_root, others_root, format, is_hp3=False):
+        self.pred.prediction(input, vocal_root, others_root, format)

infer/modules/uvr5/modules.py

@@ -0,0 +1,108 @@
+import os
+import traceback
+import logging
+
+logger = logging.getLogger(__name__)
+
+import ffmpeg
+import torch
+
+from configs.config import Config
+from infer.modules.uvr5.mdxnet import MDXNetDereverb
+from infer.modules.uvr5.vr import AudioPre, AudioPreDeEcho
+
+config = Config()
+
+
+def uvr(model_name, inp_root, save_root_vocal, paths, save_root_ins, agg, format0):
+    infos = []
+    try:
+        inp_root = inp_root.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+        save_root_vocal = (
+            save_root_vocal.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+        )
+        save_root_ins = (
+            save_root_ins.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+        )
+        if model_name == "onnx_dereverb_By_FoxJoy":
+            pre_fun = MDXNetDereverb(15, config.device)
+        else:
+            func = AudioPre if "DeEcho" not in model_name else AudioPreDeEcho
+            pre_fun = func(
+                agg=int(agg),
+                model_path=os.path.join(
+                    os.getenv("weight_uvr5_root"), model_name + ".pth"
+                ),
+                device=config.device,
+                is_half=config.is_half,
+            )
+        is_hp3 = "HP3" in model_name
+        if inp_root != "":
+            paths = [os.path.join(inp_root, name) for name in os.listdir(inp_root)]
+        else:
+            paths = [path.name for path in paths]
+        for path in paths:
+            inp_path = os.path.join(inp_root, path)
+            need_reformat = 1
+            done = 0
+            try:
+                info = ffmpeg.probe(inp_path, cmd="ffprobe")
+                if (
+                    info["streams"][0]["channels"] == 2
+                    and info["streams"][0]["sample_rate"] == "44100"
+                ):
+                    need_reformat = 0
+                    pre_fun._path_audio_(
+                        inp_path, save_root_ins, save_root_vocal, format0, is_hp3=is_hp3
+                    )
+                    done = 1
+            except:
+                need_reformat = 1
+                traceback.print_exc()
+            if need_reformat == 1:
+                tmp_path = "%s/%s.reformatted.wav" % (
+                    os.path.join(os.environ["TEMP"]),
+                    os.path.basename(inp_path),
+                )
+                os.system(
+                    "ffmpeg -i %s -vn -acodec pcm_s16le -ac 2 -ar 44100 %s -y"
+                    % (inp_path, tmp_path)
+                )
+                inp_path = tmp_path
+            try:
+                if done == 0:
+                    pre_fun._path_audio_(
+                        inp_path, save_root_ins, save_root_vocal, format0
+                    )
+                infos.append("%s->Success" % (os.path.basename(inp_path)))
+                yield "\n".join(infos)
+            except:
+                try:
+                    if done == 0:
+                        pre_fun._path_audio_(
+                            inp_path, save_root_ins, save_root_vocal, format0
+                        )
+                    infos.append("%s->Success" % (os.path.basename(inp_path)))
+                    yield "\n".join(infos)
+                except:
+                    infos.append(
+                        "%s->%s" % (os.path.basename(inp_path), traceback.format_exc())
+                    )
+                    yield "\n".join(infos)
+    except:
+        infos.append(traceback.format_exc())
+        yield "\n".join(infos)
+    finally:
+        try:
+            if model_name == "onnx_dereverb_By_FoxJoy":
+                del pre_fun.pred.model
+                del pre_fun.pred.model_
+            else:
+                del pre_fun.model
+                del pre_fun
+        except:
+            traceback.print_exc()
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            logger.info("Executed torch.cuda.empty_cache()")
+    yield "\n".join(infos)

infer/modules/uvr5/vr.py

@@ -0,0 +1,195 @@
+import os
+import logging
+
+logger = logging.getLogger(__name__)
+
+import librosa
+import numpy as np
+import soundfile as sf
+import torch
+
+from infer.lib.uvr5_pack.lib_v5 import nets_61968KB as Nets
+from infer.lib.uvr5_pack.lib_v5 import spec_utils
+from infer.lib.uvr5_pack.lib_v5.model_param_init import ModelParameters
+from infer.lib.uvr5_pack.lib_v5.nets_new import CascadedNet
+from infer.lib.uvr5_pack.utils import inference
+
+
+class AudioPre:
+    def __init__(self, agg, model_path, device, is_half, tta=False):
+        self.model_path = model_path
+        self.device = device
+        self.data = {
+            # Processing Options
+            "postprocess": False,
+            "tta": tta,
+            # Constants
+            "window_size": 512,
+            "agg": agg,
+            "high_end_process": "mirroring",
+        }
+        mp = ModelParameters("infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2.json")
+        model = Nets.CascadedASPPNet(mp.param["bins"] * 2)
+        cpk = torch.load(model_path, map_location="cpu")
+        model.load_state_dict(cpk)
+        model.eval()
+        if is_half:
+            model = model.half().to(device)
+        else:
+            model = model.to(device)
+
+        self.mp = mp
+        self.model = model
+
+    def _path_audio_(
+        self, music_file, ins_root=None, vocal_root=None, format="flac", is_hp3=False
+    ):
+        if ins_root is None and vocal_root is None:
+            return "No save root."
+        name = os.path.basename(music_file)
+        if ins_root is not None:
+            os.makedirs(ins_root, exist_ok=True)
+        if vocal_root is not None:
+            os.makedirs(vocal_root, exist_ok=True)
+        X_wave, y_wave, X_spec_s, y_spec_s = {}, {}, {}, {}
+        bands_n = len(self.mp.param["band"])
+        # print(bands_n)
+        for d in range(bands_n, 0, -1):
+            bp = self.mp.param["band"][d]
+            if d == bands_n:  # high-end band
+                (
+                    X_wave[d],
+                    _,
+                ) = librosa.core.load(  # 理论上librosa读取可能对某些音频有bug，应该上ffmpeg读取，但是太麻烦了弃坑
+                    music_file,
+                    bp["sr"],
+                    False,
+                    dtype=np.float32,
+                    res_type=bp["res_type"],
+                )
+                if X_wave[d].ndim == 1:
+                    X_wave[d] = np.asfortranarray([X_wave[d], X_wave[d]])
+            else:  # lower bands
+                X_wave[d] = librosa.core.resample(
+                    X_wave[d + 1],
+                    self.mp.param["band"][d + 1]["sr"],
+                    bp["sr"],
+                    res_type=bp["res_type"],
+                )
+            # Stft of wave source
+            X_spec_s[d] = spec_utils.wave_to_spectrogram_mt(
+                X_wave[d],
+                bp["hl"],
+                bp["n_fft"],
+                self.mp.param["mid_side"],
+                self.mp.param["mid_side_b2"],
+                self.mp.param["reverse"],
+            )
+            # pdb.set_trace()
+            if d == bands_n and self.data["high_end_process"] != "none":
+                input_high_end_h = (bp["n_fft"] // 2 - bp["crop_stop"]) + (
+                    self.mp.param["pre_filter_stop"] - self.mp.param["pre_filter_start"]
+                )
+                input_high_end = X_spec_s[d][
+                    :, bp["n_fft"] // 2 - input_high_end_h : bp["n_fft"] // 2, :
+                ]
+
+        X_spec_m = spec_utils.combine_spectrograms(X_spec_s, self.mp)
+        aggresive_set = float(self.data["agg"] / 100)
+        aggressiveness = {
+            "value": aggresive_set,
+            "split_bin": self.mp.param["band"][1]["crop_stop"],
+        }
+        with torch.no_grad():
+            pred, X_mag, X_phase = inference(
+                X_spec_m, self.device, self.model, aggressiveness, self.data
+            )
+        # Postprocess
+        if self.data["postprocess"]:
+            pred_inv = np.clip(X_mag - pred, 0, np.inf)
+            pred = spec_utils.mask_silence(pred, pred_inv)
+        y_spec_m = pred * X_phase
+        v_spec_m = X_spec_m - y_spec_m
+
+        if ins_root is not None:
+            if self.data["high_end_process"].startswith("mirroring"):
+                input_high_end_ = spec_utils.mirroring(
+                    self.data["high_end_process"], y_spec_m, input_high_end, self.mp
+                )
+                wav_instrument = spec_utils.cmb_spectrogram_to_wave(
+                    y_spec_m, self.mp, input_high_end_h, input_high_end_
+                )
+            else:
+                wav_instrument = spec_utils.cmb_spectrogram_to_wave(y_spec_m, self.mp)
+            logger.info("%s instruments done" % name)
+            if is_hp3 == True:
+                head = "vocal_"
+            else:
+                head = "instrument_"
+            if format in ["wav", "flac"]:
+                sf.write(
+                    os.path.join(
+                        ins_root,
+                        head + "{}_{}.{}".format(name, self.data["agg"], format),
+                    ),
+                    (np.array(wav_instrument) * 32768).astype("int16"),
+                    self.mp.param["sr"],
+                )  #
+            else:
+                path = os.path.join(
+                    ins_root, head + "{}_{}.wav".format(name, self.data["agg"])
+                )
+                sf.write(
+                    path,
+                    (np.array(wav_instrument) * 32768).astype("int16"),
+                    self.mp.param["sr"],
+                )
+                if os.path.exists(path):
+                    opt_format_path = path[:-4] + ".%s" % format
+                    os.system("ffmpeg -i %s -vn %s -q:a 2 -y" % (path, opt_format_path))
+                    if os.path.exists(opt_format_path):
+                        try:
+                            os.remove(path)
+                        except:
+                            pass
+        if vocal_root is not None:
+            if is_hp3 == True:
+                head = "instrument_"
+            else:
+                head = "vocal_"
+            if self.data["high_end_process"].startswith("mirroring"):
+                input_high_end_ = spec_utils.mirroring(
+                    self.data["high_end_process"], v_spec_m, input_high_end, self.mp
+                )
+                wav_vocals = spec_utils.cmb_spectrogram_to_wave(
+                    v_spec_m, self.mp, input_high_end_h, input_high_end_
+                )
+            else:
+                wav_vocals = spec_utils.cmb_spectrogram_to_wave(v_spec_m, self.mp)
+            logger.info("%s vocals done" % name)
+            if format in ["wav", "flac"]:
+                sf.write(
+                    os.path.join(
+                        vocal_root,
+                        head + "{}_{}.{}".format(name, self.data["agg"], format),
+                    ),
+                    (np.array(wav_vocals) * 32768).astype("int16"),
+                    self.mp.param["sr"],
+                )
+            else:
+                path = os.path.join(
+                    vocal_root, head + "{}_{}.wav".format(name, self.data["agg"])
+                )
+                sf.write(
+                    path,
+                    (np.array(wav_vocals) * 32768).astype("int16"),
+                    self.mp.param["sr"],
+                )
+                if os.path.exists(path):
+                    opt_format_path = path[:-4] + ".%s" % format
+                    os.system("ffmpeg -i %s -vn %s -q:a 2 -y" % (path, opt_format_path))
+                    if os.path.exists(opt_format_path):
+                        try:
+                            os.remove(path)
+                        except:
+                            pass

infer/modules/vc/modules.py

@@ -0,0 +1,302 @@
+import traceback
+import logging
+
+logger = logging.getLogger(__name__)
+
+import numpy as np
+import soundfile as sf
+import torch
+from io import BytesIO
+
+from infer.lib.audio import load_audio, wav2
+from infer.lib.infer_pack.models import (
+    SynthesizerTrnMs256NSFsid,
+    SynthesizerTrnMs256NSFsid_nono,
+    SynthesizerTrnMs768NSFsid,
+    SynthesizerTrnMs768NSFsid_nono,
+)
+from infer.modules.vc.pipeline import Pipeline
+from infer.modules.vc.utils import *
+
+
+class VC:
+    def __init__(self, config):
+        self.n_spk = None
+        self.tgt_sr = None
+        self.net_g = None
+        self.pipeline = None
+        self.cpt = None
+        self.version = None
+        self.if_f0 = None
+        self.version = None
+        self.hubert_model = None
+
+        self.config = config
+
+    def get_vc(self, sid, *to_return_protect):
+        logger.info("Get sid: " + sid)
+
+        to_return_protect0 = {
+            "visible": self.if_f0 != 0,
+            "value": to_return_protect[0]
+            if self.if_f0 != 0 and to_return_protect
+            else 0.5,
+            "__type__": "update",
+        }
+        to_return_protect1 = {
+            "visible": self.if_f0 != 0,
+            "value": to_return_protect[1]
+            if self.if_f0 != 0 and to_return_protect
+            else 0.33,
+            "__type__": "update",
+        }
+
+        if sid == "" or sid == []:
+            if self.hubert_model is not None:  # 考虑到轮询, 需要加个判断看是否 sid 是由有模型切换到无模型的
+                logger.info("Clean model cache")
+                del (self.net_g, self.n_spk, self.hubert_model, self.tgt_sr)  # ,cpt
+                self.hubert_model = (
+                    self.net_g
+                ) = self.n_spk = self.hubert_model = self.tgt_sr = None
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+                ###楼下不这么折腾清理不干净
+                self.if_f0 = self.cpt.get("f0", 1)
+                self.version = self.cpt.get("version", "v1")
+                if self.version == "v1":
+                    if self.if_f0 == 1:
+                        self.net_g = SynthesizerTrnMs256NSFsid(
+                            *self.cpt["config"], is_half=self.config.is_half
+                        )
+                    else:
+                        self.net_g = SynthesizerTrnMs256NSFsid_nono(*self.cpt["config"])
+                elif self.version == "v2":
+                    if self.if_f0 == 1:
+                        self.net_g = SynthesizerTrnMs768NSFsid(
+                            *self.cpt["config"], is_half=self.config.is_half
+                        )
+                    else:
+                        self.net_g = SynthesizerTrnMs768NSFsid_nono(*self.cpt["config"])
+                del self.net_g, self.cpt
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+            return (
+                {"visible": False, "__type__": "update"},
+                {
+                    "visible": True,
+                    "value": to_return_protect0,
+                    "__type__": "update",
+                },
+                {
+                    "visible": True,
+                    "value": to_return_protect1,
+                    "__type__": "update",
+                },
+                "",
+                "",
+            )
+        person = f'{os.getenv("weight_root")}/{sid}'
+        logger.info(f"Loading: {person}")
+
+        self.cpt = torch.load(person, map_location="cpu")
+        self.tgt_sr = self.cpt["config"][-1]
+        self.cpt["config"][-3] = self.cpt["weight"]["emb_g.weight"].shape[0]  # n_spk
+        self.if_f0 = self.cpt.get("f0", 1)
+        self.version = self.cpt.get("version", "v1")
+
+        synthesizer_class = {
+            ("v1", 1): SynthesizerTrnMs256NSFsid,
+            ("v1", 0): SynthesizerTrnMs256NSFsid_nono,
+            ("v2", 1): SynthesizerTrnMs768NSFsid,
+            ("v2", 0): SynthesizerTrnMs768NSFsid_nono,
+        }
+
+        self.net_g = synthesizer_class.get(
+            (self.version, self.if_f0), SynthesizerTrnMs256NSFsid
+        )(*self.cpt["config"], is_half=self.config.is_half)
+
+        del self.net_g.enc_q
+
+        self.net_g.load_state_dict(self.cpt["weight"], strict=False)
+        self.net_g.eval().to(self.config.device)
+        if self.config.is_half:
+            self.net_g = self.net_g.half()
+        else:
+            self.net_g = self.net_g.float()
+
+        self.pipeline = Pipeline(self.tgt_sr, self.config)
+        n_spk = self.cpt["config"][-3]
+        index = {"value": get_index_path_from_model(sid), "__type__": "update"}
+        logger.info("Select index: " + index["value"])
+
+        return (
+            (
+                {"visible": True, "maximum": n_spk, "__type__": "update"},
+                to_return_protect0,
+                to_return_protect1,
+                index,
+                index,
+            )
+            if to_return_protect
+            else {"visible": True, "maximum": n_spk, "__type__": "update"}
+        )
+
+    def vc_single(
+        self,
+        sid,
+        input_audio_path,
+        f0_up_key,
+        f0_file,
+        f0_method,
+        file_index,
+        file_index2,
+        index_rate,
+        filter_radius,
+        resample_sr,
+        rms_mix_rate,
+        protect,
+    ):
+        if input_audio_path is None:
+            return "You need to upload an audio", None
+        f0_up_key = int(f0_up_key)
+        try:
+            audio = load_audio(input_audio_path, 16000)
+            audio_max = np.abs(audio).max() / 0.95
+            if audio_max > 1:
+                audio /= audio_max
+            times = [0, 0, 0]
+
+            if self.hubert_model is None:
+                self.hubert_model = load_hubert(self.config)
+
+            if file_index:
+                file_index = (
+                    file_index.strip(" ")
+                    .strip('"')
+                    .strip("\n")
+                    .strip('"')
+                    .strip(" ")
+                    .replace("trained", "added")
+                )
+            elif file_index2:
+                file_index = file_index2
+            else:
+                file_index = ""  # 防止小白写错，自动帮他替换掉
+
+            audio_opt = self.pipeline.pipeline(
+                self.hubert_model,
+                self.net_g,
+                sid,
+                audio,
+                input_audio_path,
+                times,
+                f0_up_key,
+                f0_method,
+                file_index,
+                index_rate,
+                self.if_f0,
+                filter_radius,
+                self.tgt_sr,
+                resample_sr,
+                rms_mix_rate,
+                self.version,
+                protect,
+                f0_file,
+            )
+            if self.tgt_sr != resample_sr >= 16000:
+                tgt_sr = resample_sr
+            else:
+                tgt_sr = self.tgt_sr
+            index_info = (
+                "Index:\n%s." % file_index
+                if os.path.exists(file_index)
+                else "Index not used."
+            )
+            return (
+                "Success.\n%s\nTime:\nnpy: %.2fs, f0: %.2fs, infer: %.2fs."
+                % (index_info, *times),
+                (tgt_sr, audio_opt),
+            )
+        except:
+            info = traceback.format_exc()
+            logger.warning(info)
+            return info, (None, None)
+
+    def vc_multi(
+        self,
+        sid,
+        dir_path,
+        opt_root,
+        paths,
+        f0_up_key,
+        f0_method,
+        file_index,
+        file_index2,
+        index_rate,
+        filter_radius,
+        resample_sr,
+        rms_mix_rate,
+        protect,
+        format1,
+    ):
+        try:
+            dir_path = (
+                dir_path.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+            )  # 防止小白拷路径头尾带了空格和"和回车
+            opt_root = opt_root.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+            os.makedirs(opt_root, exist_ok=True)
+            try:
+                if dir_path != "":
+                    paths = [
+                        os.path.join(dir_path, name) for name in os.listdir(dir_path)
+                    ]
+                else:
+                    paths = [path.name for path in paths]
+            except:
+                traceback.print_exc()
+                paths = [path.name for path in paths]
+            infos = []
+            for path in paths:
+                info, opt = self.vc_single(
+                    sid,
+                    path,
+                    f0_up_key,
+                    None,
+                    f0_method,
+                    file_index,
+                    file_index2,
+                    # file_big_npy,
+                    index_rate,
+                    filter_radius,
+                    resample_sr,
+                    rms_mix_rate,
+                    protect,
+                )
+                if "Success" in info:
+                    try:
+                        tgt_sr, audio_opt = opt
+                        if format1 in ["wav", "flac"]:
+                            sf.write(
+                                "%s/%s.%s"
+                                % (opt_root, os.path.basename(path), format1),
+                                audio_opt,
+                                tgt_sr,
+                            )
+                        else:
+                            path = "%s/%s.%s" % (
+                                opt_root,
+                                os.path.basename(path),
+                                format1,
+                            )
+                            with BytesIO() as wavf:
+                                sf.write(wavf, audio_opt, tgt_sr, format="wav")
+                                wavf.seek(0, 0)
+                                with open(path, "wb") as outf:
+                                    wav2(wavf, outf, format1)
+                    except:
+                        info += traceback.format_exc()
+                infos.append("%s->%s" % (os.path.basename(path), info))
+                yield "\n".join(infos)
+            yield "\n".join(infos)
+        except:
+            yield traceback.format_exc()

infer/modules/vc/pipeline.py

@@ -0,0 +1,457 @@
+import os
+import sys
+import traceback
+import logging
+
+logger = logging.getLogger(__name__)
+
+from functools import lru_cache
+from time import time as ttime
+
+import faiss
+import librosa
+import numpy as np
+import parselmouth
+import pyworld
+import torch
+import torch.nn.functional as F
+import torchcrepe
+from scipy import signal
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+bh, ah = signal.butter(N=5, Wn=48, btype="high", fs=16000)
+
+input_audio_path2wav = {}
+
+
+@lru_cache
+def cache_harvest_f0(input_audio_path, fs, f0max, f0min, frame_period):
+    audio = input_audio_path2wav[input_audio_path]
+    f0, t = pyworld.harvest(
+        audio,
+        fs=fs,
+        f0_ceil=f0max,
+        f0_floor=f0min,
+        frame_period=frame_period,
+    )
+    f0 = pyworld.stonemask(audio, f0, t, fs)
+    return f0
+
+
+def change_rms(data1, sr1, data2, sr2, rate):  # 1是输入音频，2是输出音频,rate是2的占比
+    # print(data1.max(),data2.max())
+    rms1 = librosa.feature.rms(
+        y=data1, frame_length=sr1 // 2 * 2, hop_length=sr1 // 2
+    )  # 每半秒一个点
+    rms2 = librosa.feature.rms(y=data2, frame_length=sr2 // 2 * 2, hop_length=sr2 // 2)
+    rms1 = torch.from_numpy(rms1)
+    rms1 = F.interpolate(
+        rms1.unsqueeze(0), size=data2.shape[0], mode="linear"
+    ).squeeze()
+    rms2 = torch.from_numpy(rms2)
+    rms2 = F.interpolate(
+        rms2.unsqueeze(0), size=data2.shape[0], mode="linear"
+    ).squeeze()
+    rms2 = torch.max(rms2, torch.zeros_like(rms2) + 1e-6)
+    data2 *= (
+        torch.pow(rms1, torch.tensor(1 - rate))
+        * torch.pow(rms2, torch.tensor(rate - 1))
+    ).numpy()
+    return data2
+
+
+class Pipeline(object):
+    def __init__(self, tgt_sr, config):
+        self.x_pad, self.x_query, self.x_center, self.x_max, self.is_half = (
+            config.x_pad,
+            config.x_query,
+            config.x_center,
+            config.x_max,
+            config.is_half,
+        )
+        self.sr = 16000  # hubert输入采样率
+        self.window = 160  # 每帧点数
+        self.t_pad = self.sr * self.x_pad  # 每条前后pad时间
+        self.t_pad_tgt = tgt_sr * self.x_pad
+        self.t_pad2 = self.t_pad * 2
+        self.t_query = self.sr * self.x_query  # 查询切点前后查询时间
+        self.t_center = self.sr * self.x_center  # 查询切点位置
+        self.t_max = self.sr * self.x_max  # 免查询时长阈值
+        self.device = config.device
+
+    def get_f0(
+        self,
+        input_audio_path,
+        x,
+        p_len,
+        f0_up_key,
+        f0_method,
+        filter_radius,
+        inp_f0=None,
+    ):
+        global input_audio_path2wav
+        time_step = self.window / self.sr * 1000
+        f0_min = 50
+        f0_max = 1100
+        f0_mel_min = 1127 * np.log(1 + f0_min / 700)
+        f0_mel_max = 1127 * np.log(1 + f0_max / 700)
+        if f0_method == "pm":
+            f0 = (
+                parselmouth.Sound(x, self.sr)
+                .to_pitch_ac(
+                    time_step=time_step / 1000,
+                    voicing_threshold=0.6,
+                    pitch_floor=f0_min,
+                    pitch_ceiling=f0_max,
+                )
+                .selected_array["frequency"]
+            )
+            pad_size = (p_len - len(f0) + 1) // 2
+            if pad_size > 0 or p_len - len(f0) - pad_size > 0:
+                f0 = np.pad(
+                    f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant"
+                )
+        elif f0_method == "harvest":
+            input_audio_path2wav[input_audio_path] = x.astype(np.double)
+            f0 = cache_harvest_f0(input_audio_path, self.sr, f0_max, f0_min, 10)
+            if filter_radius > 2:
+                f0 = signal.medfilt(f0, 3)
+        elif f0_method == "crepe":
+            model = "full"
+            # Pick a batch size that doesn't cause memory errors on your gpu
+            batch_size = 512
+            # Compute pitch using first gpu
+            audio = torch.tensor(np.copy(x))[None].float()
+            f0, pd = torchcrepe.predict(
+                audio,
+                self.sr,
+                self.window,
+                f0_min,
+                f0_max,
+                model,
+                batch_size=batch_size,
+                device=self.device,
+                return_periodicity=True,
+            )
+            pd = torchcrepe.filter.median(pd, 3)
+            f0 = torchcrepe.filter.mean(f0, 3)
+            f0[pd < 0.1] = 0
+            f0 = f0[0].cpu().numpy()
+        elif f0_method == "rmvpe":
+            if not hasattr(self, "model_rmvpe"):
+                from infer.lib.rmvpe import RMVPE
+
+                logger.info(
+                    "Loading rmvpe model,%s" % "%s/rmvpe.pt" % os.environ["rmvpe_root"]
+                )
+                self.model_rmvpe = RMVPE(
+                    "%s/rmvpe.pt" % os.environ["rmvpe_root"],
+                    is_half=self.is_half,
+                    device=self.device,
+                )
+            f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
+
+            if "privateuseone" in str(self.device):  # clean ortruntime memory
+                del self.model_rmvpe.model
+                del self.model_rmvpe
+                logger.info("Cleaning ortruntime memory")
+
+        f0 *= pow(2, f0_up_key / 12)
+        # with open("test.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
+        tf0 = self.sr // self.window  # 每秒f0点数
+        if inp_f0 is not None:
+            delta_t = np.round(
+                (inp_f0[:, 0].max() - inp_f0[:, 0].min()) * tf0 + 1
+            ).astype("int16")
+            replace_f0 = np.interp(
+                list(range(delta_t)), inp_f0[:, 0] * 100, inp_f0[:, 1]
+            )
+            shape = f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)].shape[0]
+            f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)] = replace_f0[
+                :shape
+            ]
+        # with open("test_opt.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
+        f0bak = f0.copy()
+        f0_mel = 1127 * np.log(1 + f0 / 700)
+        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (
+            f0_mel_max - f0_mel_min
+        ) + 1
+        f0_mel[f0_mel <= 1] = 1
+        f0_mel[f0_mel > 255] = 255
+        f0_coarse = np.rint(f0_mel).astype(np.int32)
+        return f0_coarse, f0bak  # 1-0
+
+    def vc(
+        self,
+        model,
+        net_g,
+        sid,
+        audio0,
+        pitch,
+        pitchf,
+        times,
+        index,
+        big_npy,
+        index_rate,
+        version,
+        protect,
+    ):  # ,file_index,file_big_npy
+        feats = torch.from_numpy(audio0)
+        if self.is_half:
+            feats = feats.half()
+        else:
+            feats = feats.float()
+        if feats.dim() == 2:  # double channels
+            feats = feats.mean(-1)
+        assert feats.dim() == 1, feats.dim()
+        feats = feats.view(1, -1)
+        padding_mask = torch.BoolTensor(feats.shape).to(self.device).fill_(False)
+
+        inputs = {
+            "source": feats.to(self.device),
+            "padding_mask": padding_mask,
+            "output_layer": 9 if version == "v1" else 12,
+        }
+        t0 = ttime()
+        with torch.no_grad():
+            logits = model.extract_features(**inputs)
+            feats = model.final_proj(logits[0]) if version == "v1" else logits[0]
+        if protect < 0.5 and pitch is not None and pitchf is not None:
+            feats0 = feats.clone()
+        if (
+            not isinstance(index, type(None))
+            and not isinstance(big_npy, type(None))
+            and index_rate != 0
+        ):
+            npy = feats[0].cpu().numpy()
+            if self.is_half:
+                npy = npy.astype("float32")
+
+            # _, I = index.search(npy, 1)
+            # npy = big_npy[I.squeeze()]
+
+            score, ix = index.search(npy, k=8)
+            weight = np.square(1 / score)
+            weight /= weight.sum(axis=1, keepdims=True)
+            npy = np.sum(big_npy[ix] * np.expand_dims(weight, axis=2), axis=1)
+
+            if self.is_half:
+                npy = npy.astype("float16")
+            feats = (
+                torch.from_numpy(npy).unsqueeze(0).to(self.device) * index_rate
+                + (1 - index_rate) * feats
+            )
+
+        feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
+        if protect < 0.5 and pitch is not None and pitchf is not None:
+            feats0 = F.interpolate(feats0.permute(0, 2, 1), scale_factor=2).permute(
+                0, 2, 1
+            )
+        t1 = ttime()
+        p_len = audio0.shape[0] // self.window
+        if feats.shape[1] < p_len:
+            p_len = feats.shape[1]
+            if pitch is not None and pitchf is not None:
+                pitch = pitch[:, :p_len]
+                pitchf = pitchf[:, :p_len]
+
+        if protect < 0.5 and pitch is not None and pitchf is not None:
+            pitchff = pitchf.clone()
+            pitchff[pitchf > 0] = 1
+            pitchff[pitchf < 1] = protect
+            pitchff = pitchff.unsqueeze(-1)
+            feats = feats * pitchff + feats0 * (1 - pitchff)
+            feats = feats.to(feats0.dtype)
+        p_len = torch.tensor([p_len], device=self.device).long()
+        with torch.no_grad():
+            hasp = pitch is not None and pitchf is not None
+            arg = (feats, p_len, pitch, pitchf, sid) if hasp else (feats, p_len, sid)
+            audio1 = (net_g.infer(*arg)[0][0, 0]).data.cpu().float().numpy()
+            del hasp, arg
+        del feats, p_len, padding_mask
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        t2 = ttime()
+        times[0] += t1 - t0
+        times[2] += t2 - t1
+        return audio1
+
+    def pipeline(
+        self,
+        model,
+        net_g,
+        sid,
+        audio,
+        input_audio_path,
+        times,
+        f0_up_key,
+        f0_method,
+        file_index,
+        index_rate,
+        if_f0,
+        filter_radius,
+        tgt_sr,
+        resample_sr,
+        rms_mix_rate,
+        version,
+        protect,
+        f0_file=None,
+    ):
+        if (
+            file_index != ""
+            # and file_big_npy != ""
+            # and os.path.exists(file_big_npy) == True
+            and os.path.exists(file_index)
+            and index_rate != 0
+        ):
+            try:
+                index = faiss.read_index(file_index)
+                # big_npy = np.load(file_big_npy)
+                big_npy = index.reconstruct_n(0, index.ntotal)
+            except:
+                traceback.print_exc()
+                index = big_npy = None
+        else:
+            index = big_npy = None
+        audio = signal.filtfilt(bh, ah, audio)
+        audio_pad = np.pad(audio, (self.window // 2, self.window // 2), mode="reflect")
+        opt_ts = []
+        if audio_pad.shape[0] > self.t_max:
+            audio_sum = np.zeros_like(audio)
+            for i in range(self.window):
+                audio_sum += np.abs(audio_pad[i : i - self.window])
+            for t in range(self.t_center, audio.shape[0], self.t_center):
+                opt_ts.append(
+                    t
+                    - self.t_query
+                    + np.where(
+                        audio_sum[t - self.t_query : t + self.t_query]
+                        == audio_sum[t - self.t_query : t + self.t_query].min()
+                    )[0][0]
+                )
+        s = 0
+        audio_opt = []
+        t = None
+        t1 = ttime()
+        audio_pad = np.pad(audio, (self.t_pad, self.t_pad), mode="reflect")
+        p_len = audio_pad.shape[0] // self.window
+        inp_f0 = None
+        if hasattr(f0_file, "name"):
+            try:
+                with open(f0_file.name, "r") as f:
+                    lines = f.read().strip("\n").split("\n")
+                inp_f0 = []
+                for line in lines:
+                    inp_f0.append([float(i) for i in line.split(",")])
+                inp_f0 = np.array(inp_f0, dtype="float32")
+            except:
+                traceback.print_exc()
+        sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
+        pitch, pitchf = None, None
+        if if_f0 == 1:
+            pitch, pitchf = self.get_f0(
+                input_audio_path,
+                audio_pad,
+                p_len,
+                f0_up_key,
+                f0_method,
+                filter_radius,
+                inp_f0,
+            )
+            pitch = pitch[:p_len]
+            pitchf = pitchf[:p_len]
+            if "mps" not in str(self.device) or "xpu" not in str(self.device):
+                pitchf = pitchf.astype(np.float32)
+            pitch = torch.tensor(pitch, device=self.device).unsqueeze(0).long()
+            pitchf = torch.tensor(pitchf, device=self.device).unsqueeze(0).float()
+        t2 = ttime()
+        times[1] += t2 - t1
+        for t in opt_ts:
+            t = t // self.window * self.window
+            if if_f0 == 1:
+                audio_opt.append(
+                    self.vc(
+                        model,
+                        net_g,
+                        sid,
+                        audio_pad[s : t + self.t_pad2 + self.window],
+                        pitch[:, s // self.window : (t + self.t_pad2) // self.window],
+                        pitchf[:, s // self.window : (t + self.t_pad2) // self.window],
+                        times,
+                        index,
+                        big_npy,
+                        index_rate,
+                        version,
+                        protect,
+                    )[self.t_pad_tgt : -self.t_pad_tgt]
+                )
+            else:
+                audio_opt.append(
+                    self.vc(
+                        model,
+                        net_g,
+                        sid,
+                        audio_pad[s : t + self.t_pad2 + self.window],
+                        None,
+                        None,
+                        times,
+                        index,
+                        big_npy,
+                        index_rate,
+                        version,
+                        protect,
+                    )[self.t_pad_tgt : -self.t_pad_tgt]
+                )
+            s = t
+        if if_f0 == 1:
+            audio_opt.append(
+                self.vc(
+                    model,
+                    net_g,
+                    sid,
+                    audio_pad[t:],
+                    pitch[:, t // self.window :] if t is not None else pitch,
+                    pitchf[:, t // self.window :] if t is not None else pitchf,
+                    times,
+                    index,
+                    big_npy,
+                    index_rate,
+                    version,
+                    protect,
+                )[self.t_pad_tgt : -self.t_pad_tgt]
+            )
+        else:
+            audio_opt.append(
+                self.vc(
+                    model,
+                    net_g,
+                    sid,
+                    audio_pad[t:],
+                    None,
+                    None,
+                    times,
+                    index,
+                    big_npy,
+                    index_rate,
+                    version,
+                    protect,
+                )[self.t_pad_tgt : -self.t_pad_tgt]
+            )
+        audio_opt = np.concatenate(audio_opt)
+        if rms_mix_rate != 1:
+            audio_opt = change_rms(audio, 16000, audio_opt, tgt_sr, rms_mix_rate)
+        if tgt_sr != resample_sr >= 16000:
+            audio_opt = librosa.resample(
+                audio_opt, orig_sr=tgt_sr, target_sr=resample_sr
+            )
+        audio_max = np.abs(audio_opt).max() / 0.99
+        max_int16 = 32768
+        if audio_max > 1:
+            max_int16 /= audio_max
+        audio_opt = (audio_opt * max_int16).astype(np.int16)
+        del pitch, pitchf, sid
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        return audio_opt

infer/modules/vc/utils.py

@@ -0,0 +1,33 @@
+import os
+
+from fairseq import checkpoint_utils
+
+
+def get_index_path_from_model(sid):
+    return next(
+        (
+            f
+            for f in [
+                os.path.join(root, name)
+                for root, _, files in os.walk(os.getenv("index_root"), topdown=False)
+                for name in files
+                if name.endswith(".index") and "trained" not in name
+            ]
+            if sid.split(".")[0] in f
+        ),
+        "",
+    )
+
+
+def load_hubert(config):
+    models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
+        ["assets/hubert/hubert_base.pt"],
+        suffix="",
+    )
+    hubert_model = models[0]
+    hubert_model = hubert_model.to(config.device)
+    if config.is_half:
+        hubert_model = hubert_model.half()
+    else:
+        hubert_model = hubert_model.float()
+    return hubert_model.eval()