Import RVC 20260716 Nvidia 50x0 v2bb

infer/audio.py

@@ -0,0 +1,51 @@
+import platform, os
+import ffmpeg
+import numpy as np
+import av
+from io import BytesIO
+
+
+def wav2(i, o, format):
+    inp = av.open(i, "r")
+    if format == "m4a":
+        format = "mp4"
+    out = av.open(o, "w", 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 load_audio(file, sr):
+    try:
+        # https://github.com/openai/whisper/blob/main/whisper/audio.py#L26
+        # This launches a subprocess to decode audio while down-mixing and resampling as necessary.
+        # Requires the ffmpeg CLI and `ffmpeg-python` package to be installed.
+        file = clean_path(file)  # 防止小白拷路径头尾带了空格和"和回车
+        out, _ = (
+            ffmpeg.input(file, threads=0)
+            .output("-", format="f32le", acodec="pcm_f32le", ac=1, ar=sr)
+            .run(cmd=["ffmpeg", "-nostdin"], capture_stdout=True, capture_stderr=True)
+        )
+    except Exception as e:
+        raise RuntimeError(f"Failed to load audio: {e}")
+
+    return np.frombuffer(out, np.float32).flatten()
+
+
+def clean_path(path_str):
+    if platform.system() == "Windows":
+        path_str = path_str.replace("/", "\\")
+    return path_str.strip(" ").strip('"').strip("\n").strip('"').strip(" ")

infer/fcpe.py

@@ -0,0 +1,97 @@
+import torch
+
+
+def _is_directml_device(device):
+    """Return whether *device* is the PrivateUse1 device registered by DirectML."""
+    return getattr(device, "type", None) == "privateuseone" or "privateuseone" in str(
+        device
+    ).lower()
+
+
+class FCPEInfer:
+    """Project-local FCPE inference adapter with a DirectML execution path.
+
+    DirectML does not support the complex tensor produced by ``torch.stft`` in
+    torchfcpe's wav2mel stage.  Keep preprocessing and the small indexed
+    decoder on CPU while the FCPE neural network runs on DirectML.  Other
+    devices retain torchfcpe's original end-to-end inference path.
+    """
+
+    def __init__(self, device):
+        from torchfcpe import spawn_bundled_infer_model
+
+        self.device = device
+        self.is_directml = _is_directml_device(device)
+        if self.is_directml:
+            # Loading a checkpoint directly with map_location=privateuseone is
+            # not supported consistently.  Load on CPU, leave wav2mel there,
+            # and move only the real-valued FCPE network to DirectML.
+            self.infer_model = spawn_bundled_infer_model("cpu")
+            self.infer_model.wav2mel.eval()
+            self.cent_table_cpu = (
+                self.infer_model.model.cent_table.detach().float().cpu().clone()
+            )
+            self.out_dims = int(self.infer_model.model.out_dims)
+            self.infer_model.model.to(device).eval()
+        else:
+            self.infer_model = spawn_bundled_infer_model(device)
+
+    def _decode_on_cpu(self, latent, decoder_mode, threshold):
+        """Decode DML network logits on CPU with torchfcpe's exact formulas.
+
+        The current DirectML backend's ``aten::gather`` returns incorrect FCPE
+        bin values even though its indices and the neural-network logits match
+        CPU.  Decoding is tiny compared with the model, so keep this
+        compatibility boundary on CPU as well as the complex STFT.
+        """
+        latent = latent.detach().float().cpu()
+        batch, frames, _ = latent.shape
+        cents = self.cent_table_cpu[None, None, :].expand(batch, frames, -1)
+
+        if decoder_mode == "argmax":
+            confidence = torch.max(latent, dim=-1, keepdim=True).values
+            decoded = torch.sum(cents * latent, dim=-1, keepdim=True) / torch.sum(
+                latent, dim=-1, keepdim=True
+            )
+        elif decoder_mode == "local_argmax":
+            confidence, max_index = torch.max(latent, dim=-1, keepdim=True)
+            local_index = torch.arange(9, dtype=torch.long) + (max_index - 4)
+            local_index.clamp_(0, self.out_dims - 1)
+            local_cents = torch.gather(cents, -1, local_index)
+            local_latent = torch.gather(latent, -1, local_index)
+            decoded = torch.sum(
+                local_cents * local_latent, dim=-1, keepdim=True
+            ) / torch.sum(local_latent, dim=-1, keepdim=True)
+        else:
+            raise ValueError(f"Unknown FCPE decoder mode: {decoder_mode}")
+
+        decoded = decoded.masked_fill(confidence <= threshold, float("-inf"))
+        return 10.0 * torch.pow(2.0, decoded / 1200.0)
+
+    # torch.no_grad is used instead of inference_mode because DirectML's
+    # PrivateUse1 backend still updates version counters in a few operators.
+    @torch.no_grad()
+    def infer(
+        self,
+        wav,
+        sr,
+        decoder_mode="local_argmax",
+        threshold=0.006,
+    ):
+        if not self.is_directml:
+            return self.infer_model.infer(
+                wav,
+                sr=sr,
+                decoder_mode=decoder_mode,
+                threshold=threshold,
+            )
+
+        wav_cpu = wav.detach().to(device="cpu", dtype=torch.float32)
+        mel_cpu = self.infer_model.wav2mel(wav_cpu, sr)
+        mel_dml = mel_cpu.to(device=self.device, dtype=torch.float32)
+        latent_dml = self.infer_model.model(mel_dml)
+        return self._decode_on_cpu(
+            latent_dml,
+            decoder_mode=decoder_mode,
+            threshold=threshold,
+        )

infer/hubert.py

@@ -0,0 +1,96 @@
+import logging
+from functools import lru_cache
+from pathlib import Path
+
+import torch
+from torch import nn
+from transformers import AutoFeatureExtractor, HubertModel
+
+
+logger = logging.getLogger(__name__)
+
+PROJECT_ROOT = Path(__file__).resolve().parent.parent
+
+
+class HubertModelWithFinalProj(HubertModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.final_proj = nn.Linear(config.hidden_size, config.classifier_proj_size)
+
+HUBERT_MODEL_PATH = (PROJECT_ROOT / "assets" / "hubert_base").resolve()
+
+
+def _device_type(device):
+    if isinstance(device, torch.device):
+        return device.type
+    return str(device).split(":", 1)[0]
+
+
+def load_hubert_model(device, is_half=False):
+    """Load the local Transformers HuBERT/ContentVec model for RVC."""
+    if not (HUBERT_MODEL_PATH / "config.json").is_file():
+        raise FileNotFoundError(
+            f"Transformers HuBERT model not found: {HUBERT_MODEL_PATH}"
+        )
+
+    dtype = torch.float16 if is_half else torch.float32
+    load_options = {
+        "local_files_only": True,
+        "torch_dtype": dtype,
+    }
+    # DirectML does not implement every SDPA kernel used by Transformers.
+    if _device_type(device) == "privateuseone":
+        load_options["attn_implementation"] = "eager"
+
+    logger.info(
+        "Loading Transformers HuBERT from %s (%s on %s)",
+        HUBERT_MODEL_PATH,
+        dtype,
+        device,
+    )
+    model = HubertModelWithFinalProj.from_pretrained(
+        str(HUBERT_MODEL_PATH), **load_options
+    )
+    model = model.to(device)
+    return model.eval()
+
+
+@lru_cache(maxsize=1)
+def hubert_audio_requires_normalization():
+    feature_extractor = AutoFeatureExtractor.from_pretrained(
+        str(HUBERT_MODEL_PATH), local_files_only=True
+    )
+    return bool(feature_extractor.do_normalize)
+
+
+def extract_hubert_features(model, source, version, padding_mask=None):
+    """Return the RVC v1 (256-D) or v2 (768-D) HuBERT representation.
+
+    Transformers hidden_states[N] is numerically equivalent to the source checkpoint's
+    output_layer=N for this converted checkpoint. RVC v1 uses layer 9 followed
+    by final_proj; RVC v2 uses the final (12th) encoder layer directly.
+    """
+    if version not in {"v1", "v2"}:
+        raise ValueError(f"Unsupported RVC feature version: {version!r}")
+
+    attention_mask = None
+    if padding_mask is not None and bool(torch.any(padding_mask).item()):
+        attention_mask = (~padding_mask.bool()).long()
+
+    if version == "v1":
+        outputs = model(
+            input_values=source,
+            attention_mask=attention_mask,
+            output_hidden_states=True,
+            return_dict=True,
+        )
+        features = outputs.hidden_states[9]
+        return model.final_proj(features)
+
+    outputs = model(
+        input_values=source,
+        attention_mask=attention_mask,
+        output_hidden_states=False,
+        return_dict=True,
+    )
+    return outputs.last_hidden_state

infer/module/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.module import commons, modules
+from infer.module.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, c, attn_mask = 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,
+        key,
+        value,
+        mask = 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):
+        max_relative_position = 2 * self.window_size + 1
+        # Pad first before slice to avoid using cond ops.
+        pad_length = 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):
+        """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 = 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, x_mask) :
+        if self.causal:
+            padding = self._causal_padding(x * x_mask)
+        else:
+            padding = self._same_padding(x * x_mask)
+        return padding
+
+    def forward(self, x, x_mask):
+        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 = self.kernel_size - 1
+        pad_r = 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 = (self.kernel_size - 1) // 2
+        pad_r = 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/module/commons.py

@@ -0,0 +1,171 @@
+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
+
+
+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) :
+    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, max_length = 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/module/modules.py

@@ -0,0 +1,548 @@
+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.module import commons
+from infer.module.commons import get_padding, init_weights
+from infer.module.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 = 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, x_mask, g = 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)
+
+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, x_mask = 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)
+
+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 = 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)
+
+class Log(nn.Module):
+    def forward(
+        self,
+        x,
+        x_mask,
+        g = None,
+        reverse = False,
+    ) :
+        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):
+    def forward(
+        self,
+        x,
+        x_mask,
+        g = None,
+        reverse = False,
+    ) :
+        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,
+        x_mask,
+        g = None,
+        reverse = 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()
+
+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,
+        x_mask,
+        g = 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/module/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/rmvpe.py

@@ -0,0 +1,640 @@
+import os
+from typing import List, Optional, Tuple
+import numpy as np
+import torch
+
+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):
+        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):
+        concat_tensors = []
+        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, concat_tensors):
+        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) :
+        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, is_half, device=None):
+        self.resample_kernel = {}
+        self.resample_kernel = {}
+        if isinstance(is_half, str):
+            is_half = is_half.lower() == "true"
+        if device is None:
+            from configs.config import infer_device, infer_dtype
+
+            device = str(infer_device)
+            is_half = infer_dtype == torch.float16
+        elif str(device).startswith("cuda"):
+            from configs.config import get_device_dtype_sm
+
+            parsed_device = torch.device(device)
+            device_index = parsed_device.index
+            if device_index is None:
+                device_index = torch.cuda.current_device()
+            selected_device, selected_dtype, _, _ = get_device_dtype_sm(device_index)
+            device = str(selected_device)
+            is_half = selected_dtype == torch.float16
+        else:
+            is_half = False
+        self.is_half = is_half
+        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(
+                os.path.splitext(model_path)[0] + ".onnx",
+                providers=["DmlExecutionProvider"],
+            )
+            self.model = ort_session
+        else:
+            if str(self.device) == "cuda":
+                self.device = torch.device("cuda:0")
+
+            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
+
+            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/rtrvc.py

@@ -0,0 +1,342 @@
+import traceback
+from time import time as ttime
+import faiss
+import numpy as np
+import parselmouth
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchaudio.transforms import Resample
+
+from infer.hubert import extract_hubert_features, load_hubert_model
+from i18n.i18n import I18nAuto
+
+
+i18n = I18nAuto()
+
+
+def printt(strr, *args):
+    if len(args) == 0:
+        print(strr)
+    else:
+        print(strr % args)
+
+
+def get_synthesizer(pth_path, device=torch.device("cpu")):
+    from infer.module.models import (
+        SynthesizerTrnMs256NSFsid,
+        SynthesizerTrnMs256NSFsid_nono,
+        SynthesizerTrnMs768NSFsid,
+        SynthesizerTrnMs768NSFsid_nono,
+    )
+
+    cpt = torch.load(pth_path, map_location=torch.device("cpu"))
+    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.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
+
+
+# config.device=torch.device("cpu")########强制cpu测试
+# config.is_half=False########强制cpu测试
+class RVC:
+    def __init__(
+        self,
+        key,
+        formant,
+        pth_path,
+        index_path,
+        index_rate,
+        config,
+        last_rvc=None,
+    ) :
+        """
+        初始化
+        """
+        try:
+            # global config
+            self.config = config
+            # device="cpu"########强制cpu测试
+            self.device = config.device
+            self.f0_up_key = key
+            self.formant_shift = formant
+            self.f0_min = 50
+            self.f0_max = 1100
+            self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
+            self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
+            self.is_half = config.is_half
+            if index_rate != 0:
+                self.index = faiss.read_index(index_path)
+                self.big_npy = self.index.reconstruct_n(0, self.index.ntotal)
+                printt(i18n("已启用索引检索"))
+            self.pth_path = pth_path
+            self.index_path = index_path
+            self.index_rate = index_rate
+            self.cache_pitch = torch.zeros(
+                1024, device=self.device, dtype=torch.long
+            )
+            self.cache_pitchf = torch.zeros(
+                1024, device=self.device, dtype=torch.float32
+            )
+
+            self.resample_kernel = {}
+
+            if last_rvc is None:
+                self.model = load_hubert_model(self.device, self.is_half)
+            else:
+                self.model = last_rvc.model
+
+            self.net_g = None
+
+            def set_synthesizer():
+                self.net_g, cpt = get_synthesizer(self.pth_path, self.device)
+                self.tgt_sr = cpt["config"][-1]
+                cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
+                self.if_f0 = cpt.get("f0", 1)
+                self.version = cpt.get("version", "v1")
+                if self.is_half:
+                    self.net_g = self.net_g.half()
+                else:
+                    self.net_g = self.net_g.float()
+
+            if last_rvc is None or last_rvc.pth_path != self.pth_path:
+                set_synthesizer()
+            else:
+                self.tgt_sr = last_rvc.tgt_sr
+                self.if_f0 = last_rvc.if_f0
+                self.version = last_rvc.version
+                self.is_half = last_rvc.is_half
+                self.net_g = last_rvc.net_g
+
+            if last_rvc is not None and hasattr(last_rvc, "model_rmvpe"):
+                self.model_rmvpe = last_rvc.model_rmvpe
+            if last_rvc is not None and hasattr(last_rvc, "model_fcpe"):
+                self.model_fcpe = last_rvc.model_fcpe
+        except:
+            printt(traceback.format_exc())
+
+    def change_key(self, new_key):
+        self.f0_up_key = new_key
+
+    def change_formant(self, new_formant):
+        self.formant_shift = new_formant
+
+    def change_index_rate(self, new_index_rate):
+        if new_index_rate != 0 and self.index_rate == 0:
+            self.index = faiss.read_index(self.index_path)
+            self.big_npy = self.index.reconstruct_n(0, self.index.ntotal)
+            printt(i18n("已启用索引检索"))
+        self.index_rate = new_index_rate
+
+    def get_f0_post(self, f0):
+        if not torch.is_tensor(f0):
+            f0 = torch.from_numpy(f0)
+        f0 = f0.float().to(self.device).squeeze()
+        f0_mel = 1127 * torch.log(1 + f0 / 700)
+        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * 254 / (
+            self.f0_mel_max - self.f0_mel_min
+        ) + 1
+        f0_mel[f0_mel <= 1] = 1
+        f0_mel[f0_mel > 255] = 255
+        f0_coarse = torch.round(f0_mel).long()
+        return f0_coarse, f0
+
+    def get_f0(self, x, f0_up_key, method="rmvpe"):
+        if method == "rmvpe":
+            return self.get_f0_rmvpe(x, f0_up_key)
+        if method == "fcpe":
+            return self.get_f0_fcpe(x, f0_up_key)
+        if method != "pm":
+            raise ValueError(f"Unsupported F0 method: {method}")
+        x = x.cpu().numpy()
+        p_len = x.shape[0] // 160 + 1
+        f0_min = 65
+        l_pad = int(np.ceil(1.5 / f0_min * 16000))
+        r_pad = l_pad + 1
+        s = parselmouth.Sound(np.pad(x, (l_pad, r_pad)), 16000).to_pitch_ac(
+            time_step=0.01,
+            voicing_threshold=0.6,
+            pitch_floor=f0_min,
+            pitch_ceiling=1100,
+        )
+        assert np.abs(s.t1 - 1.5 / f0_min) < 0.001
+        f0 = s.selected_array["frequency"]
+        if len(f0) < p_len:
+            f0 = np.pad(f0, (0, p_len - len(f0)))
+        f0 = f0[:p_len]
+        try:
+            uv = f0 == 0
+            f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
+        except Exception:
+            traceback.print_exc()
+        f0 *= pow(2, f0_up_key / 12)
+        return self.get_f0_post(f0)
+
+    def get_f0_rmvpe(self, x, f0_up_key):
+        if hasattr(self, "model_rmvpe") == False:
+            from infer.rmvpe import RMVPE
+
+            printt(i18n("正在加载RMVPE模型"))
+            self.model_rmvpe = RMVPE(
+                "assets/rmvpe/rmvpe.pt",
+                is_half=self.is_half,
+                device=self.device,
+            )
+        f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
+        try:
+            uv = f0 == 0
+            f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
+        except Exception:
+            traceback.print_exc()
+        f0 *= pow(2, f0_up_key / 12)
+        return self.get_f0_post(f0)
+
+    def get_f0_fcpe(self, x, f0_up_key):
+        if hasattr(self, "model_fcpe") == False:
+            from infer.fcpe import FCPEInfer
+
+            printt("Loading fcpe model")
+            self.model_fcpe = FCPEInfer(self.device)
+        f0 = self.model_fcpe.infer(
+            x.unsqueeze(0).float(),
+            sr=16000,
+            decoder_mode="local_argmax",
+            threshold=0.006,
+        ).squeeze().detach().cpu().numpy()
+        try:
+            uv = f0 == 0
+            f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
+        except Exception:
+            traceback.print_exc()
+        f0 *= pow(2, f0_up_key / 12)
+        return self.get_f0_post(f0)
+
+    def infer(
+        self,
+        input_wav,
+        block_frame_16k,
+        skip_head,
+        return_length,
+        f0method,
+    ) :
+        t1 = ttime()
+        with torch.no_grad():
+            if self.config.is_half:
+                feats = input_wav.half().view(1, -1)
+            else:
+                feats = input_wav.float().view(1, -1)
+            padding_mask = torch.BoolTensor(feats.shape).to(self.device).fill_(False)
+            feats = extract_hubert_features(
+                self.model,
+                feats,
+                self.version,
+                padding_mask=padding_mask,
+            )
+            feats = torch.cat((feats, feats[:, -1:, :]), 1)
+        t2 = ttime()
+        try:
+            if hasattr(self, "index") and self.index_rate != 0:
+                npy = feats[0][skip_head // 2 :].cpu().numpy().astype("float32")
+                score, ix = self.index.search(npy, k=8)
+                if (ix >= 0).all():
+                    weight = np.square(1 / score)
+                    weight /= weight.sum(axis=1, keepdims=True)
+                    npy = np.sum(
+                        self.big_npy[ix] * np.expand_dims(weight, axis=2), axis=1
+                    )
+                    if self.config.is_half:
+                        npy = npy.astype("float16")
+                    feats[0][skip_head // 2 :] = (
+                        torch.from_numpy(npy).unsqueeze(0).to(self.device)
+                        * self.index_rate
+                        + (1 - self.index_rate) * feats[0][skip_head // 2 :]
+                    )
+                else:
+                    printt(
+                        i18n("索引无效：必须使用added_xxxx.index，不能使用trained_xxxx.index")
+                    )
+            else:
+                printt(i18n("索引检索失败或未启用"))
+        except Exception:
+            traceback.print_exc()
+            printt(i18n("索引检索失败"))
+        t3 = ttime()
+        p_len = input_wav.shape[0] // 160
+        factor = pow(2, self.formant_shift / 12)
+        return_length2 = int(np.ceil(return_length * factor))
+        if self.if_f0 == 1:
+            f0_extractor_frame = block_frame_16k + 800
+            if f0method == "rmvpe":
+                f0_extractor_frame = 5120 * ((f0_extractor_frame - 1) // 5120 + 1) - 160
+            pitch, pitchf = self.get_f0(
+                input_wav[-f0_extractor_frame:],
+                self.f0_up_key - self.formant_shift,
+                f0method,
+            )
+            shift = block_frame_16k // 160
+            self.cache_pitch[:-shift] = self.cache_pitch[shift:].clone()
+            self.cache_pitchf[:-shift] = self.cache_pitchf[shift:].clone()
+            self.cache_pitch[4 - pitch.shape[0] :] = pitch[3:-1]
+            self.cache_pitchf[4 - pitch.shape[0] :] = pitchf[3:-1]
+            cache_pitch = self.cache_pitch[None, -p_len:]
+            cache_pitchf = self.cache_pitchf[None, -p_len:] * return_length2 / return_length
+        t4 = ttime()
+        feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
+        feats = feats[:, :p_len, :]
+        p_len = torch.LongTensor([p_len]).to(self.device)
+        sid = torch.LongTensor([0]).to(self.device)
+        skip_head = torch.LongTensor([skip_head])
+        return_length2 = torch.LongTensor([return_length2])
+        return_length = torch.LongTensor([return_length])
+        with torch.no_grad():
+            if self.if_f0 == 1:
+                infered_audio, _, _ = self.net_g.infer(
+                    feats,
+                    p_len,
+                    cache_pitch,
+                    cache_pitchf,
+                    sid,
+                    skip_head,
+                    return_length,
+                    return_length2,
+                )
+            else:
+                infered_audio, _, _ = self.net_g.infer(
+                    feats, p_len, sid, skip_head, return_length, return_length2
+                )
+        infered_audio = infered_audio.squeeze(1).float()
+        upp_res = int(np.floor(factor * self.tgt_sr // 100))
+        if upp_res != self.tgt_sr // 100:
+            if upp_res not in self.resample_kernel:
+                self.resample_kernel[upp_res] = Resample(
+                    orig_freq=upp_res,
+                    new_freq=self.tgt_sr // 100,
+                    dtype=torch.float32,
+                ).to(self.device)
+            infered_audio = self.resample_kernel[upp_res](
+                infered_audio[:, : return_length * upp_res]
+            )
+        t5 = ttime()
+        printt(
+            i18n("耗时：特征=%.3f秒，索引=%.3f秒，音高=%.3f秒，模型=%.3f秒"),
+            t2 - t1,
+            t3 - t2,
+            t4 - t3,
+            t5 - t4,
+        )
+        return infered_audio.squeeze()

infer/vc/modules.py

@@ -0,0 +1,359 @@
+import traceback
+import logging
+
+logger = logging.getLogger(__name__)
+
+import numpy as np
+import soundfile as sf
+import torch
+from io import BytesIO
+
+from infer.audio import load_audio, wav2
+from infer.module.models import (
+    SynthesizerTrnMs256NSFsid,
+    SynthesizerTrnMs256NSFsid_nono,
+    SynthesizerTrnMs768NSFsid,
+    SynthesizerTrnMs768NSFsid_nono,
+)
+from infer.vc.pipeline import Pipeline
+from infer.vc.utils import *
+from i18n.i18n import I18nAuto
+from tools.progress import batch_status, should_report
+
+
+i18n = I18nAuto()
+
+
+def inference_status(title, state, detail=""):
+    lines = ["【%s】" % i18n(title), "%s：%s" % (i18n("状态"), i18n(state))]
+    if detail:
+        lines.extend(["", str(detail).strip()])
+    return "\n".join(lines)
+
+
+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("%s: %s", i18n("选择模型"), 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(i18n("清理模型缓存"))
+                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("%s: %s", i18n("正在加载模型"), 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("%s: %s", i18n("选择索引"), 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_method,
+        file_index,
+        index_rate,
+        resample_sr,
+        rms_mix_rate,
+        protect,
+    ):
+        if input_audio_path is None:
+            return inference_status("单次推理", "等待输入", i18n("请上传音频文件")), 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")
+                )
+            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,
+                times,
+                f0_up_key,
+                f0_method,
+                file_index,
+                index_rate,
+                self.if_f0,
+                self.tgt_sr,
+                resample_sr,
+                rms_mix_rate,
+                self.version,
+                protect,
+            )
+            if self.tgt_sr != resample_sr >= 16000:
+                tgt_sr = resample_sr
+            else:
+                tgt_sr = self.tgt_sr
+            index_info = (
+                "%s：%s" % (i18n("索引"), file_index)
+                if os.path.exists(file_index)
+                else "%s：%s" % (i18n("索引"), i18n("未使用"))
+            )
+            return (
+                inference_status(
+                    "单次推理",
+                    "成功",
+                    "%s\n%s：%s %.2fs | F0 %.2fs | %s %.2fs"
+                    % (
+                        index_info,
+                        i18n("耗时"),
+                        i18n("特征"),
+                        times[0],
+                        times[1],
+                        i18n("合成"),
+                        times[2],
+                    ),
+                ),
+                (tgt_sr, audio_opt),
+            )
+        except Exception:
+            info = traceback.format_exc()
+            logger.warning(info)
+            return inference_status("单次推理", "失败", info), (None, None)
+
+    def vc_multi(
+        self,
+        sid,
+        dir_path,
+        opt_root,
+        paths,
+        f0_up_key,
+        f0_method,
+        file_index,
+        index_rate,
+        resample_sr,
+        rms_mix_rate,
+        protect,
+        format1,
+    ):
+        try:
+            dir_path = (
+                (dir_path or "")
+                .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 or "")
+                .strip(" ")
+                .strip('"')
+                .strip("\n")
+                .strip('"')
+                .strip(" ")
+            )
+            if not opt_root:
+                yield inference_status(
+                    "批量推理", "等待输入", i18n("请填写输出文件夹路径")
+                )
+                return
+            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 if isinstance(path, str) else path.name for path in (paths or [])]
+            except Exception:
+                traceback.print_exc()
+                paths = [
+                    path if isinstance(path, str) else path.name for path in (paths or [])
+                ]
+            total = len(paths)
+            if total == 0:
+                yield batch_status(i18n("批量推理"), 0, 0, 0, 0)
+                return
+            success = 0
+            failed = 0
+            failures = []
+            for idx, path in enumerate(paths):
+                item_failed = False
+                info, opt = self.vc_single(
+                    sid,
+                    path,
+                    f0_up_key,
+                    f0_method,
+                    file_index,
+                    index_rate,
+                    resample_sr,
+                    rms_mix_rate,
+                    protect,
+                )
+                if opt and opt[0] is not None and opt[1] is not None:
+                    try:
+                        tgt_sr, audio_opt = opt
+                        if format1 in ["wav", "flac"]:
+                            sf.write(
+                                "%s/%s.%s"
+                                % (
+                                    opt_root,
+                                    os.path.splitext(os.path.basename(path))[0],
+                                    format1,
+                                ),
+                                audio_opt,
+                                tgt_sr,
+                            )
+                        else:
+                            path = "%s/%s.%s" % (
+                                opt_root,
+                                os.path.splitext(os.path.basename(path))[0],
+                                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)
+                        success += 1
+                    except Exception:
+                        info = "%s\n%s" % (info, traceback.format_exc())
+                        failed += 1
+                        item_failed = True
+                        failures.append("%s：%s" % (os.path.basename(path), info))
+                else:
+                    failed += 1
+                    item_failed = True
+                    failures.append("%s：%s" % (os.path.basename(path), info))
+                if should_report(idx, total) or item_failed:
+                    yield batch_status(
+                        i18n("批量推理"),
+                        idx + 1,
+                        total,
+                        success,
+                        failed,
+                        os.path.basename(path),
+                        failures,
+                    )
+        except Exception:
+            yield inference_status("批量推理", "失败", traceback.format_exc())

infer/vc/pipeline.py

@@ -0,0 +1,386 @@
+import os
+import traceback
+import logging
+
+logger = logging.getLogger(__name__)
+
+from time import time as ttime
+
+import faiss
+import librosa
+import numpy as np
+import parselmouth
+import torch
+import torch.nn.functional as F
+from scipy import signal
+
+from infer.hubert import extract_hubert_features
+
+bh, ah = signal.butter(N=5, Wn=48, btype="high", fs=16000)
+
+
+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,
+        x,
+        p_len,
+        f0_up_key,
+        f0_method,
+    ):
+        if f0_method not in ("pm", "rmvpe", "fcpe"):
+            raise ValueError(f"Unsupported F0 method: {f0_method}")
+        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 == "rmvpe":
+            if not hasattr(self, "model_rmvpe"):
+                from infer.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")
+        elif f0_method == "fcpe":
+            if not hasattr(self, "model_fcpe"):
+                from infer.fcpe import FCPEInfer
+
+                logger.info("Loading fcpe model")
+                self.model_fcpe = FCPEInfer(self.device)
+            f0 = self.model_fcpe.infer(
+                torch.from_numpy(x).unsqueeze(0).float(),
+                sr=self.sr,
+                decoder_mode="local_argmax",
+                threshold=0.006,
+            ).squeeze().detach().cpu().numpy()
+
+        try:
+            uv = f0 == 0
+            f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
+        except Exception:
+            traceback.print_exc()
+        f0 *= pow(2, f0_up_key / 12)
+        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,
+        index_vectors,
+        index_rate,
+        version,
+        protect,
+    ):
+        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)
+
+        t0 = ttime()
+        with torch.no_grad():
+            feats = extract_hubert_features(
+                model,
+                feats.to(self.device),
+                version,
+                padding_mask=padding_mask,
+            )
+        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(index_vectors, type(None))
+            and index_rate != 0
+        ):
+            npy = feats[0].cpu().numpy()
+            if self.is_half:
+                npy = npy.astype("float32")
+
+            score, ix = index.search(npy, k=8)
+            weight = np.square(1 / score)
+            weight /= weight.sum(axis=1, keepdims=True)
+            npy = np.sum(index_vectors[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,
+        times,
+        f0_up_key,
+        f0_method,
+        file_index,
+        index_rate,
+        if_f0,
+        tgt_sr,
+        resample_sr,
+        rms_mix_rate,
+        version,
+        protect,
+    ):
+        if (
+            file_index != ""
+            and os.path.exists(file_index)
+            and index_rate != 0
+        ):
+            try:
+                index = faiss.read_index(file_index)
+                index_vectors = index.reconstruct_n(0, index.ntotal)
+            except:
+                traceback.print_exc()
+                index = index_vectors = None
+        else:
+            index = index_vectors = 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
+        sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
+        pitch, pitchf = None, None
+        if if_f0 == 1:
+            pitch, pitchf = self.get_f0(
+                audio_pad,
+                p_len,
+                f0_up_key,
+                f0_method,
+            )
+            pitch = pitch[:p_len]
+            pitchf = pitchf[:p_len]
+            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,
+                        index_vectors,
+                        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,
+                        index_vectors,
+                        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,
+                    index_vectors,
+                    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,
+                    index_vectors,
+                    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/vc/utils.py

@@ -0,0 +1,44 @@
+import os
+import re
+
+from infer.hubert import load_hubert_model
+
+
+def get_index_path_from_model(sid):
+    model_stem = os.path.splitext(os.path.basename(str(sid or "")))[0]
+    experiment_name = re.sub(r"_e\d+_s\d+$", "", model_stem, flags=re.IGNORECASE)
+    if not experiment_name:
+        return ""
+
+    candidates = []
+    roots = [os.getenv("outside_index_root"), os.getenv("index_root")]
+    for index_root in roots:
+        if not index_root or not os.path.isdir(index_root):
+            continue
+        for root, _, files in os.walk(index_root, topdown=False):
+            for name in files:
+                if not name.lower().endswith(".index") or "trained" in name.lower():
+                    continue
+                index_stem = os.path.splitext(name)[0]
+                lower_index = index_stem.lower()
+                lower_experiment = experiment_name.lower()
+                standard_match = (
+                    lower_index.startswith(lower_experiment + "_added_")
+                    or ("_" + lower_experiment + "_v1") in lower_index
+                    or ("_" + lower_experiment + "_v2") in lower_index
+                )
+                exact_model_match = model_stem.lower() in lower_index
+                if standard_match or exact_model_match:
+                    path = os.path.abspath(os.path.join(root, name))
+                    score = (
+                        0 if standard_match else 1,
+                        0 if os.path.abspath(index_root) == os.path.abspath(roots[0]) else 1,
+                        -os.path.getmtime(path),
+                        path.lower(),
+                    )
+                    candidates.append((score, path))
+    return min(candidates, default=(None, ""), key=lambda item: item[0])[1]
+
+
+def load_hubert(config):
+    return load_hubert_model(config.device, config.is_half)