From d83590dc35d1abb026e895bbf883bf97eee28b62 Mon Sep 17 00:00:00 2001 From: wataru Date: Thu, 22 Jun 2023 06:56:00 +0900 Subject: [PATCH] WIP: integrate vcs to new gui 3 --- server/voice_changer/MMVCv15/MMVCv15.py | 9 +- .../voice_changer/MMVCv15/models/commons.py | 27 ++ .../voice_changer/MMVCv15/models/features.py | 200 ++++++++ .../voice_changer/MMVCv15/models/generator.py | 312 +++++++++++++ server/voice_changer/MMVCv15/models/index.py | 82 ++++ server/voice_changer/MMVCv15/models/models.py | 438 ++++++++++++++++++ .../voice_changer/MMVCv15/models/modules.py | 186 ++++++++ .../voice_changer/MMVCv15/models/readme.txt | 1 + .../MMVCv15/models/residual_block.py | 257 ++++++++++ server/voice_changer/MMVCv15/models/snake.py | 47 ++ 10 files changed, 1552 insertions(+), 7 deletions(-) create mode 100644 server/voice_changer/MMVCv15/models/commons.py create mode 100644 server/voice_changer/MMVCv15/models/features.py create mode 100644 server/voice_changer/MMVCv15/models/generator.py create mode 100644 server/voice_changer/MMVCv15/models/index.py create mode 100644 server/voice_changer/MMVCv15/models/models.py create mode 100644 server/voice_changer/MMVCv15/models/modules.py create mode 100644 server/voice_changer/MMVCv15/models/readme.txt create mode 100644 server/voice_changer/MMVCv15/models/residual_block.py create mode 100644 server/voice_changer/MMVCv15/models/snake.py diff --git a/server/voice_changer/MMVCv15/MMVCv15.py b/server/voice_changer/MMVCv15/MMVCv15.py index 73275625..78d46705 100644 --- a/server/voice_changer/MMVCv15/MMVCv15.py +++ b/server/voice_changer/MMVCv15/MMVCv15.py @@ -20,7 +20,7 @@ import torch import onnxruntime import pyworld as pw -from models import SynthesizerTrn # type:ignore +from voice_changer.MMVCv15.models.models import SynthesizerTrn # type:ignore from voice_changer.MMVCv15.client_modules import ( convert_continuos_f0, spectrogram_torch, @@ -156,8 +156,7 @@ class MMVCv15: def get_info(self): data = asdict(self.settings) - data["onnxExecutionProviders"] = self.onnx_session.get_providers() if self.settings.onnxModelFile != "" and self.settings.onnxModelFile is not None else [] - + data["onnxExecutionProviders"] = self.onnx_session.get_providers() if self.onnx_session is not None else [] return data def get_processing_sampling_rate(self): @@ -231,10 +230,6 @@ class MMVCv15: return [spec, f0, sid] def _onnx_inference(self, data): - if self.settings.onnxModelFile == "" and self.settings.onnxModelFile is None: - print("[Voice Changer] No ONNX session.") - raise NoModeLoadedException("ONNX") - spec, f0, sid_src = data spec = spec.unsqueeze(0) spec_lengths = torch.tensor([spec.size(2)]) diff --git a/server/voice_changer/MMVCv15/models/commons.py b/server/voice_changer/MMVCv15/models/commons.py new file mode 100644 index 00000000..b0928f92 --- /dev/null +++ b/server/voice_changer/MMVCv15/models/commons.py @@ -0,0 +1,27 @@ +import torch + + +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 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 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) diff --git a/server/voice_changer/MMVCv15/models/features.py b/server/voice_changer/MMVCv15/models/features.py new file mode 100644 index 00000000..766831ac --- /dev/null +++ b/server/voice_changer/MMVCv15/models/features.py @@ -0,0 +1,200 @@ +# -*- coding: utf-8 -*- + +# Copyright 2022 Reo Yoneyama (Nagoya University) +# MIT License (https://opensource.org/licenses/MIT) + +"""Feature-related functions. + +References: + - https://github.com/bigpon/QPPWG + - https://github.com/nii-yamagishilab/project-NN-Pytorch-scripts + +""" + +import sys +from logging import getLogger + +import numpy as np +import torch +from torch.nn.functional import interpolate + +# A logger for this file +logger = getLogger(__name__) + + +def validate_length(xs, ys=None, hop_size=None): + """Validate length + + Args: + xs (ndarray): numpy array of features + ys (ndarray): numpy array of audios + hop_size (int): upsampling factor + + Returns: + (ndarray): length adjusted features + + """ + min_len_x = min([x.shape[0] for x in xs]) + if ys is not None: + min_len_y = min([y.shape[0] for y in ys]) + if min_len_y < min_len_x * hop_size: + min_len_x = min_len_y // hop_size + if min_len_y > min_len_x * hop_size: + min_len_y = min_len_x * hop_size + ys = [y[:min_len_y] for y in ys] + xs = [x[:min_len_x] for x in xs] + + return xs + ys if ys is not None else xs + + +def dilated_factor(batch_f0, fs, dense_factor): + """Pitch-dependent dilated factor + + Args: + batch_f0 (ndarray): the f0 sequence (T) + fs (int): sampling rate + dense_factor (int): the number of taps in one cycle + + Return: + dilated_factors(np array): + float array of the pitch-dependent dilated factors (T) + + """ + batch_f0[batch_f0 == 0] = fs / dense_factor + dilated_factors = torch.ones_like(batch_f0) * fs / dense_factor / batch_f0 + # assert np.all(dilated_factors > 0) + return dilated_factors + + +class SignalGenerator: + """Input signal generator module.""" + + def __init__( + self, + sample_rate=24000, + hop_size=120, + sine_amp=0.1, + noise_amp=0.003, + signal_types=["sine", "noise"], + ): + """Initialize WaveNetResidualBlock module. + + Args: + sample_rate (int): Sampling rate. + hop_size (int): Hop size of input F0. + sine_amp (float): Sine amplitude for NSF-based sine generation. + noise_amp (float): Noise amplitude for NSF-based sine generation. + signal_types (list): List of input signal types for generator. + + """ + self.sample_rate = sample_rate + self.hop_size = hop_size + self.signal_types = signal_types + self.sine_amp = sine_amp + self.noise_amp = noise_amp + + for signal_type in signal_types: + if signal_type not in ["noise", "sine", "sines", "uv"]: + logger.info(f"{signal_type} is not supported type for generator input.") + sys.exit(0) + # logger.info(f"Use {signal_types} for generator input signals.") + + @torch.no_grad() + def __call__(self, f0, f0_scale=1.0): + signals = [] + for typ in self.signal_types: + if "noise" == typ: + signals.append(self.random_noise(f0)) + if "sine" == typ: + signals.append(self.sinusoid(f0)) + if "sines" == typ: + signals.append(self.sinusoids(f0)) + if "uv" == typ: + signals.append(self.vuv_binary(f0)) + + input_batch = signals[0] + for signal in signals[1:]: + input_batch = torch.cat([input_batch, signal], axis=1) + + return input_batch * f0_scale + + @torch.no_grad() + def random_noise(self, f0): + """Calculate noise signals. + + Args: + f0 (Tensor): F0 tensor (B, 1, T // hop_size). + + Returns: + Tensor: Gaussian noise signals (B, 1, T). + + """ + B, _, T = f0.size() + noise = torch.randn((B, 1, T * self.hop_size), device=f0.device) + + return noise + + @torch.no_grad() + def sinusoid(self, f0): + """Calculate sine signals. + + Args: + f0 (Tensor): F0 tensor (B, 1, T // hop_size). + + Returns: + Tensor: Sines generated following NSF (B, 1, T). + + """ + B, _, T = f0.size() + vuv = interpolate((f0 > 0) * torch.ones_like(f0), T * self.hop_size) + radious = (interpolate(f0, T * self.hop_size) / self.sample_rate) % 1 + sine = vuv * torch.sin(torch.cumsum(radious, dim=2) * 2 * np.pi) * self.sine_amp + if self.noise_amp > 0: + noise_amp = vuv * self.noise_amp + (1.0 - vuv) * self.noise_amp / 3.0 + noise = torch.randn((B, 1, T * self.hop_size), device=f0.device) * noise_amp + sine = sine + noise + + return sine + + @torch.no_grad() + def sinusoids(self, f0): + """Calculate sines. + + Args: + f0 (Tensor): F0 tensor (B, 1, T // hop_size). + + Returns: + Tensor: Sines generated following NSF (B, 1, T). + + """ + B, _, T = f0.size() + vuv = interpolate((f0 > 0) * torch.ones_like(f0), T * self.hop_size) + f0 = interpolate(f0, T * self.hop_size) + sines = torch.zeros_like(f0, device=f0.device) + harmonics = 5 # currently only fixed number of harmonics is supported + for i in range(harmonics): + radious = (f0 * (i + 1) / self.sample_rate) % 1 + sines += torch.sin(torch.cumsum(radious, dim=2) * 2 * np.pi) + sines = self.sine_amp * sines * vuv / harmonics + if self.noise_amp > 0: + noise_amp = vuv * self.noise_amp + (1.0 - vuv) * self.noise_amp / 3.0 + noise = torch.randn((B, 1, T * self.hop_size), device=f0.device) * noise_amp + sines = sines + noise + + return sines + + @torch.no_grad() + def vuv_binary(self, f0): + """Calculate V/UV binary sequences. + + Args: + f0 (Tensor): F0 tensor (B, 1, T // hop_size). + + Returns: + Tensor: V/UV binary sequences (B, 1, T). + + """ + _, _, T = f0.size() + uv = interpolate((f0 > 0) * torch.ones_like(f0), T * self.hop_size) + + return uv diff --git a/server/voice_changer/MMVCv15/models/generator.py b/server/voice_changer/MMVCv15/models/generator.py new file mode 100644 index 00000000..35b46af9 --- /dev/null +++ b/server/voice_changer/MMVCv15/models/generator.py @@ -0,0 +1,312 @@ +# -*- coding: utf-8 -*- + +# Copyright 2022 Reo Yoneyama (Nagoya University) +# MIT License (https://opensource.org/licenses/MIT) + +"""HiFiGAN and SiFiGAN Generator modules. + +References: + - https://github.com/kan-bayashi/ParallelWaveGAN + - https://github.com/bigpon/QPPWG + - https://github.com/jik876/hifi-gan + +""" + +from logging import getLogger + +import torch.nn as nn +from .residual_block import AdaptiveResidualBlock, Conv1d, ResidualBlock + +# A logger for this file +logger = getLogger(__name__) + + +class SiFiGANGenerator(nn.Module): + """SiFiGAN generator module.""" + + def __init__( + self, + in_channels, + out_channels=1, + channels=512, + kernel_size=7, + upsample_scales=(5, 4, 3, 2), + upsample_kernel_sizes=(10, 8, 6, 4), + source_network_params={ + "resblock_kernel_size": 3, # currently only 3 is supported. + "resblock_dilations": [(1,), (1, 2), (1, 2, 4), (1, 2, 4, 8)], + "use_additional_convs": True, + }, + filter_network_params={ + "resblock_kernel_sizes": (3, 5, 7), + "resblock_dilations": [(1, 3, 5), (1, 3, 5), (1, 3, 5)], + "use_additional_convs": False, + }, + share_upsamples=False, + share_downsamples=False, + bias=True, + nonlinear_activation="LeakyReLU", + nonlinear_activation_params={"negative_slope": 0.1}, + use_weight_norm=True, + requires_grad=True, + ): + """Initialize SiFiGANGenerator module. + + Args: + in_channels (int): Number of input channels. + out_channels (int): Number of output channels. + channels (int): Number of hidden representation channels. + kernel_size (int): Kernel size of initial and final conv layer. + upsample_scales (list): List of upsampling scales. + upsample_kernel_sizes (list): List of kernel sizes for upsampling layers. + source_network_params (dict): Parameters for source-network. + filter_network_params (dict): Parameters for filter-network. + share_upsamples (bool): Whether to share up-sampling transposed CNNs. + share_downsamples (bool): Whether to share down-sampling CNNs. + bias (bool): Whether to add bias parameter in convolution layers. + nonlinear_activation (str): Activation function module name. + nonlinear_activation_params (dict): Hyperparameters for activation function. + use_weight_norm (bool): Whether to use weight norm. + If set to true, it will be applied to all of the conv layers. + + """ + super().__init__() + # check hyperparameters are valid + assert kernel_size % 2 == 1, "Kernel size must be odd number." + assert len(upsample_scales) == len(upsample_kernel_sizes) + + # define modules + self.num_upsamples = len(upsample_kernel_sizes) + self.source_network_params = source_network_params + self.filter_network_params = filter_network_params + self.share_upsamples = share_upsamples + self.share_downsamples = share_downsamples + self.sn = nn.ModuleDict() + self.fn = nn.ModuleDict() + self.input_conv = Conv1d( + in_channels, + channels, + kernel_size, + bias=bias, + padding=(kernel_size - 1) // 2, + ) + self.sn["upsamples"] = nn.ModuleList() + self.fn["upsamples"] = nn.ModuleList() + self.sn["blocks"] = nn.ModuleList() + self.fn["blocks"] = nn.ModuleList() + for i in range(len(upsample_kernel_sizes)): + assert upsample_kernel_sizes[i] == 2 * upsample_scales[i] + self.sn["upsamples"] += [ + nn.Sequential( + getattr(nn, nonlinear_activation)(**nonlinear_activation_params), + nn.ConvTranspose1d( + channels // (2**i), + channels // (2 ** (i + 1)), + upsample_kernel_sizes[i], + upsample_scales[i], + padding=upsample_scales[i] // 2 + upsample_scales[i] % 2, + output_padding=upsample_scales[i] % 2, + bias=bias, + ), + ) + ] + if not share_upsamples: + self.fn["upsamples"] += [ + nn.Sequential( + getattr(nn, nonlinear_activation)(**nonlinear_activation_params), + nn.ConvTranspose1d( + channels // (2**i), + channels // (2 ** (i + 1)), + upsample_kernel_sizes[i], + upsample_scales[i], + padding=upsample_scales[i] // 2 + upsample_scales[i] % 2, + output_padding=upsample_scales[i] % 2, + bias=bias, + ), + ) + ] + self.sn["blocks"] += [ + AdaptiveResidualBlock( + kernel_size=source_network_params["resblock_kernel_size"], + channels=channels // (2 ** (i + 1)), + dilations=source_network_params["resblock_dilations"][i], + bias=bias, + use_additional_convs=source_network_params["use_additional_convs"], + nonlinear_activation=nonlinear_activation, + nonlinear_activation_params=nonlinear_activation_params, + ) + ] + for j in range(len(filter_network_params["resblock_kernel_sizes"])): + self.fn["blocks"] += [ + ResidualBlock( + kernel_size=filter_network_params["resblock_kernel_sizes"][j], + channels=channels // (2 ** (i + 1)), + dilations=filter_network_params["resblock_dilations"][j], + bias=bias, + use_additional_convs=filter_network_params["use_additional_convs"], + nonlinear_activation=nonlinear_activation, + nonlinear_activation_params=nonlinear_activation_params, + ) + ] + self.sn["output_conv"] = nn.Sequential( + nn.LeakyReLU(), + nn.Conv1d( + channels // (2 ** (i + 1)), + out_channels, + kernel_size, + bias=bias, + padding=(kernel_size - 1) // 2, + ), + ) + self.fn["output_conv"] = nn.Sequential( + nn.LeakyReLU(), + nn.Conv1d( + channels // (2 ** (i + 1)), + out_channels, + kernel_size, + bias=bias, + padding=(kernel_size - 1) // 2, + ), + nn.Tanh(), + ) + + # sine embedding layers + self.sn["emb"] = Conv1d( + 1, + channels // (2 ** len(upsample_kernel_sizes)), + kernel_size, + bias=bias, + padding=(kernel_size - 1) // 2, + ) + # down-sampling CNNs + self.sn["downsamples"] = nn.ModuleList() + for i in reversed(range(1, len(upsample_kernel_sizes))): + self.sn["downsamples"] += [ + nn.Sequential( + nn.Conv1d( + channels // (2 ** (i + 1)), + channels // (2**i), + upsample_kernel_sizes[i], + upsample_scales[i], + padding=upsample_scales[i] - (upsample_kernel_sizes[i] % 2 == 0), + bias=bias, + ), + getattr(nn, nonlinear_activation)(**nonlinear_activation_params), + ) + ] + if not share_downsamples: + self.fn["downsamples"] = nn.ModuleList() + for i in reversed(range(1, len(upsample_kernel_sizes))): + self.fn["downsamples"] += [ + nn.Sequential( + nn.Conv1d( + channels // (2 ** (i + 1)), + channels // (2**i), + upsample_kernel_sizes[i], + upsample_scales[i], + padding=upsample_scales[i] - (upsample_kernel_sizes[i] % 2 == 0), + bias=bias, + ), + getattr(nn, nonlinear_activation)(**nonlinear_activation_params), + ) + ] + + # apply weight norm + if use_weight_norm: + self.apply_weight_norm() + + # reset parameters + self.reset_parameters() + + if requires_grad is False: + for param in self.parameters(): + param.requires_grad = False + + def forward(self, x, c, d, sid): + """Calculate forward propagation. + + Args: + x (Tensor): Input sine signal (B, 1, T). + c (Tensor): Input tensor (B, in_channels, T). + d (List): F0-dependent dilation factors [(B, 1, T) x num_upsamples]. + + Returns: + Tensor: Output tensor (B, out_channels, T). + + """ + + # currently, same input feature is input to each network + c = self.input_conv(c) + e = c + + # source-network forward + x = self.sn["emb"](x) + embs = [x] + for i in range(self.num_upsamples - 1): + x = self.sn["downsamples"][i](x) + embs += [x] + for i in range(self.num_upsamples): + # excitation generation network + e = self.sn["upsamples"][i](e) + embs[-i - 1] + e = self.sn["blocks"][i](e, d[i]) + e_ = self.sn["output_conv"](e) + + # filter-network forward + embs = [e] + for i in range(self.num_upsamples - 1): + if self.share_downsamples: + e = self.sn["downsamples"][i](e) + else: + e = self.fn["downsamples"][i](e) + embs += [e] + num_blocks = len(self.filter_network_params["resblock_kernel_sizes"]) + for i in range(self.num_upsamples): + # resonance filtering network + if self.share_upsamples: + c = self.sn["upsamples"][i](c) + embs[-i - 1] + else: + c = self.fn["upsamples"][i](c) + embs[-i - 1] + cs = 0.0 # initialize + for j in range(num_blocks): + cs += self.fn["blocks"][i * num_blocks + j](c) + c = cs / num_blocks + c = self.fn["output_conv"](c) + + return c, e_ + + def reset_parameters(self): + """Reset parameters. + + This initialization follows the official implementation manner. + https://github.com/jik876/hifi-gan/blob/master/models.py + + """ + + def _reset_parameters(m): + if isinstance(m, (nn.Conv1d, nn.ConvTranspose1d)): + m.weight.data.normal_(0.0, 0.01) + logger.debug(f"Reset parameters in {m}.") + + self.apply(_reset_parameters) + + def remove_weight_norm(self): + """Remove weight normalization module from all of the layers.""" + + def _remove_weight_norm(m): + try: + logger.debug(f"Weight norm is removed from {m}.") + nn.utils.remove_weight_norm(m) + except ValueError: # this module didn't have weight norm + return + + self.apply(_remove_weight_norm) + + def apply_weight_norm(self): + """Apply weight normalization module from all of the layers.""" + + def _apply_weight_norm(m): + if isinstance(m, nn.Conv1d) or isinstance(m, nn.ConvTranspose1d): + nn.utils.weight_norm(m) + logger.debug(f"Weight norm is applied to {m}.") + + self.apply(_apply_weight_norm) diff --git a/server/voice_changer/MMVCv15/models/index.py b/server/voice_changer/MMVCv15/models/index.py new file mode 100644 index 00000000..e1288ba5 --- /dev/null +++ b/server/voice_changer/MMVCv15/models/index.py @@ -0,0 +1,82 @@ +# -*- coding: utf-8 -*- + +# Copyright 2020 Yi-Chiao Wu (Nagoya University) +# MIT License (https://opensource.org/licenses/MIT) + +"""Indexing-related functions.""" + +import torch +from torch.nn import ConstantPad1d as pad1d + + +def pd_indexing(x, d, dilation, batch_index, ch_index): + """Pitch-dependent indexing of past and future samples. + + Args: + x (Tensor): Input feature map (B, C, T). + d (Tensor): Input pitch-dependent dilated factors (B, 1, T). + dilation (Int): Dilation size. + batch_index (Tensor): Batch index + ch_index (Tensor): Channel index + + Returns: + Tensor: Past output tensor (B, out_channels, T) + Tensor: Future output tensor (B, out_channels, T) + + """ + (_, _, batch_length) = d.size() + dilations = d * dilation + + # get past index + idxP = torch.arange(-batch_length, 0).float() + idxP = idxP.to(x.device) + idxP = torch.add(-dilations, idxP) + idxP = idxP.round().long() + maxP = -((torch.min(idxP) + batch_length)) + assert maxP >= 0 + idxP = (batch_index, ch_index, idxP) + # padding past tensor + xP = pad1d((maxP, 0), 0)(x) + + # get future index + idxF = torch.arange(0, batch_length).float() + idxF = idxF.to(x.device) + idxF = torch.add(dilations, idxF) + idxF = idxF.round().long() + maxF = torch.max(idxF) - (batch_length - 1) + assert maxF >= 0 + idxF = (batch_index, ch_index, idxF) + # padding future tensor + xF = pad1d((0, maxF), 0)(x) + + return xP[idxP], xF[idxF] + + +def index_initial(n_batch, n_ch, tensor=True): + """Tensor batch and channel index initialization. + + Args: + n_batch (Int): Number of batch. + n_ch (Int): Number of channel. + tensor (bool): Return tensor or numpy array + + Returns: + Tensor: Batch index + Tensor: Channel index + + """ + batch_index = [] + for i in range(n_batch): + batch_index.append([[i]] * n_ch) + ch_index = [] + for i in range(n_ch): + ch_index += [[i]] + ch_index = [ch_index] * n_batch + + if tensor: + batch_index = torch.tensor(batch_index) + ch_index = torch.tensor(ch_index) + if torch.cuda.is_available(): + batch_index = batch_index.cuda() + ch_index = ch_index.cuda() + return batch_index, ch_index diff --git a/server/voice_changer/MMVCv15/models/models.py b/server/voice_changer/MMVCv15/models/models.py new file mode 100644 index 00000000..669e56be --- /dev/null +++ b/server/voice_changer/MMVCv15/models/models.py @@ -0,0 +1,438 @@ +import numpy as np +import torch +from torch import nn +from torch.nn import functional as F + +from .modules import ResidualCouplingLayer, Flip, WN, ResBlock1, ResBlock2, LRELU_SLOPE + +from torch.nn import Conv1d, ConvTranspose1d, Conv2d +from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm +from .commons import init_weights, get_padding, sequence_mask +from .generator import SiFiGANGenerator +from .features import SignalGenerator, dilated_factor + + +class TextEncoder(nn.Module): + def __init__(self, out_channels, hidden_channels, requires_grad=True): + super().__init__() + self.out_channels = out_channels + self.hidden_channels = hidden_channels + self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1) + # パラメータを学習しない + if requires_grad is False: + for param in self.parameters(): + param.requires_grad = False + + def forward(self, x, x_lengths): + x = torch.transpose(x.half(), 1, -1) # [b, h, t] + x_mask = torch.unsqueeze(sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype) + stats = self.proj(x) * x_mask + m, logs = torch.split(stats, self.out_channels, dim=1) + return x, m, logs, x_mask + + +class ResidualCouplingBlock(nn.Module): + def __init__(self, channels, hidden_channels, kernel_size, dilation_rate, n_layers, n_flows=4, gin_channels=0, requires_grad=True): + super().__init__() + self.channels = channels + self.hidden_channels = hidden_channels + self.kernel_size = kernel_size + self.dilation_rate = dilation_rate + self.n_layers = n_layers + self.n_flows = n_flows + self.gin_channels = gin_channels + + self.flows = nn.ModuleList() + for i in range(n_flows): + self.flows.append(ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True)) + self.flows.append(Flip()) + + # パラメータを学習しない + if requires_grad is False: + for param in self.parameters(): + param.requires_grad = False + + def forward(self, x, x_mask, g=None, reverse=False): + if not reverse: + for flow in self.flows: + x, _ = flow(x, x_mask, g=g, reverse=reverse) + else: + for flow in reversed(self.flows): + x = flow(x, x_mask, g=g, reverse=reverse) + return x + + +class PosteriorEncoder(nn.Module): + def __init__(self, in_channels, out_channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, requires_grad=True): + super().__init__() + self.in_channels = in_channels + self.out_channels = out_channels + self.hidden_channels = hidden_channels + self.kernel_size = kernel_size + self.dilation_rate = dilation_rate + self.n_layers = n_layers + self.gin_channels = gin_channels + + self.pre = nn.Conv1d(in_channels, hidden_channels, 1) + self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels) + self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1) + + # パラメータを学習しない + if requires_grad is False: + for param in self.parameters(): + param.requires_grad = False + + def forward(self, x, x_lengths, g=None): + x_mask = torch.unsqueeze(sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype) + x = self.pre(x) * x_mask + x = self.enc(x, x_mask, g=g) + stats = self.proj(x) * x_mask + m, logs = torch.split(stats, self.out_channels, dim=1) + z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask + return z, m, logs, x_mask + + +class Generator(torch.nn.Module): + def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, requires_grad=True): + super(Generator, self).__init__() + self.num_kernels = len(resblock_kernel_sizes) + self.num_upsamples = len(upsample_rates) + self.conv_pre = Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3) + resblock = ResBlock1 if resblock == "1" else ResBlock2 + + self.ups = nn.ModuleList() + for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)): + self.ups.append(weight_norm(ConvTranspose1d(upsample_initial_channel // (2**i), upsample_initial_channel // (2 ** (i + 1)), k, u, padding=(k - u) // 2))) + + self.resblocks = nn.ModuleList() + for i in range(len(self.ups)): + ch = upsample_initial_channel // (2 ** (i + 1)) + for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)): + self.resblocks.append(resblock(ch, k, d)) + + self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False) + self.ups.apply(init_weights) + + if requires_grad is False: + for param in self.parameters(): + param.requires_grad = False + + def forward(self, x, g=None): + x = self.conv_pre(x) + + for i in range(self.num_upsamples): + x = F.leaky_relu(x, LRELU_SLOPE) + x = self.ups[i](x) + xs = None + for j in range(self.num_kernels): + if xs is None: + xs = self.resblocks[i * self.num_kernels + j](x) + else: + xs += self.resblocks[i * self.num_kernels + j](x) + x = xs / self.num_kernels + x = F.leaky_relu(x) + x = self.conv_post(x) + x = torch.tanh(x) + + return x + + def remove_weight_norm(self): + print("Removing weight norm...") + for l in self.ups: + remove_weight_norm(l) + for l in self.resblocks: + l.remove_weight_norm() + + +class DiscriminatorP(torch.nn.Module): + def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False): + super(DiscriminatorP, self).__init__() + self.period = period + self.use_spectral_norm = use_spectral_norm + norm_f = weight_norm if use_spectral_norm is False else spectral_norm + self.convs = nn.ModuleList( + [ + norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), + norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), + norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), + norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))), + norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))), + ] + ) + self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0))) + + def forward(self, x): + fmap = [] + + # 1d to 2d + b, c, t = x.shape + if t % self.period != 0: # pad first + n_pad = self.period - (t % self.period) + x = F.pad(x, (0, n_pad), "reflect") + t = t + n_pad + x = x.view(b, c, t // self.period, self.period) + + for l in self.convs: + x = l(x) + x = F.leaky_relu(x, LRELU_SLOPE) + fmap.append(x) + x = self.conv_post(x) + fmap.append(x) + x = torch.flatten(x, 1, -1) + + return x, fmap + + +class DiscriminatorS(torch.nn.Module): + def __init__(self, use_spectral_norm=False): + super(DiscriminatorS, self).__init__() + norm_f = weight_norm if use_spectral_norm is False else spectral_norm + self.convs = nn.ModuleList( + [ + norm_f(Conv1d(1, 16, 15, 1, padding=7)), + norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)), + norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)), + norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)), + norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)), + norm_f(Conv1d(1024, 1024, 5, 1, padding=2)), + ] + ) + self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1)) + + def forward(self, x): + fmap = [] + + for l in self.convs: + x = l(x) + x = F.leaky_relu(x, LRELU_SLOPE) + fmap.append(x) + x = self.conv_post(x) + fmap.append(x) + x = torch.flatten(x, 1, -1) + + return x, fmap + + +class MultiPeriodDiscriminator(torch.nn.Module): + def __init__(self, use_spectral_norm=False): + super(MultiPeriodDiscriminator, self).__init__() + # periods = [2,3,5,7,11] + periods = [3, 5, 7, 11, 13] + + discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)] + discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods] + self.discriminators = nn.ModuleList(discs) + + def forward(self, y, y_hat, flag=True): + if flag: + y_d_rs = [] + y_d_gs = [] + fmap_rs = [] + fmap_gs = [] + for i, d in enumerate(self.discriminators): + y_d_r, fmap_r = d(y) + y_d_g, fmap_g = d(y_hat) + y_d_rs.append(y_d_r) + y_d_gs.append(y_d_g) + fmap_rs.append(fmap_r) + fmap_gs.append(fmap_g) + + return y_d_rs, y_d_gs, fmap_rs, fmap_gs + else: + y_d_gs = [] + with torch.no_grad(): + for i, d in enumerate(self.discriminators): + y_d_g, _ = d(y_hat) + y_d_gs.append(y_d_g) + + return y_d_gs + + +class SynthesizerTrn(nn.Module): + """ + Synthesizer for Training + """ + + def __init__( + self, + spec_channels, + segment_size, + inter_channels, + hidden_channels, + upsample_rates, + upsample_initial_channel, + upsample_kernel_sizes, + n_flow, + dec_out_channels=1, + dec_kernel_size=7, + n_speakers=0, + gin_channels=0, + requires_grad_pe=True, + requires_grad_flow=True, + requires_grad_text_enc=True, + requires_grad_dec=True, + requires_grad_emb_g=True, + sample_rate=24000, + hop_size=128, + sine_amp=0.1, + noise_amp=0.003, + signal_types=["sine"], + dense_factors=[0.5, 1, 4, 8], + upsample_scales=[8, 4, 2, 2], + ): + super().__init__() + self.spec_channels = spec_channels + self.hidden_channels = hidden_channels + self.upsample_rates = upsample_rates + self.upsample_initial_channel = upsample_initial_channel + self.upsample_kernel_sizes = upsample_kernel_sizes + self.segment_size = segment_size + self.dec_out_channels = dec_out_channels + self.dec_kernel_size = dec_kernel_size + self.n_speakers = n_speakers + self.gin_channels = gin_channels + self.requires_grad_pe = requires_grad_pe + self.requires_grad_flow = requires_grad_flow + self.requires_grad_text_enc = requires_grad_text_enc + self.requires_grad_dec = requires_grad_dec + self.requires_grad_emb_g = requires_grad_emb_g + self.sample_rate = sample_rate + self.hop_size = hop_size + self.sine_amp = sine_amp + self.noise_amp = noise_amp + self.signal_types = signal_types + self.dense_factors = dense_factors + self.upsample_scales = upsample_scales + + self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels, requires_grad=requires_grad_pe) + self.enc_p = TextEncoder(inter_channels, hidden_channels, requires_grad=requires_grad_text_enc) + self.dec = SiFiGANGenerator(in_channels=inter_channels, out_channels=dec_out_channels, channels=upsample_initial_channel, kernel_size=dec_kernel_size, upsample_scales=upsample_rates, upsample_kernel_sizes=upsample_kernel_sizes, requires_grad=requires_grad_dec) + self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, n_flows=n_flow, gin_channels=gin_channels, requires_grad=requires_grad_flow) + self.signal_generator = SignalGenerator(sample_rate=sample_rate, hop_size=hop_size, noise_amp=noise_amp, signal_types=signal_types) + + if n_speakers > 1: + self.emb_g = nn.Embedding(n_speakers, gin_channels) + self.emb_g.requires_grad = requires_grad_emb_g + + def forward(self, x, x_lengths, y, y_lengths, f0, slice_id, sid=None, target_ids=None): + pass + # sin, d = self.make_sin_d(f0) + + # x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths) + # # target sid 作成 + # target_sids = self.make_random_target_sids(target_ids, sid) + + # if self.n_speakers > 0: + # g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1] + # tgt_g = self.emb_g(target_sids).unsqueeze(-1) # [b, h, 1] + # else: + # g = None + + # # PE + # z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g) + # # Flow + # z_p = self.flow(z, y_mask, g=g) + # # VC + # tgt_z = self.flow(z_p, y_mask, g=tgt_g, reverse=True) + # # アライメントの作成 + # liner_alignment = F.one_hot(torch.arange(0, x.shape[2] + 2)).cuda() + # liner_alignment = torch.stack([liner_alignment for _ in range(x.shape[0])], axis=0) + # liner_alignment = F.interpolate(liner_alignment.float(), size=(z.shape[2]), mode="linear", align_corners=True) + # liner_alignment = liner_alignment[:, 1:-1, :] + # # TextEncとPEのshape合わせ + # m_p = torch.matmul(m_p, liner_alignment) + # logs_p = torch.matmul(logs_p, liner_alignment) + + # # slice + # z_slice = slice_segments(z, slice_id, self.segment_size) + # # targetのslice + # tgt_z_slice = slice_segments(tgt_z, slice_id, self.segment_size) + # # Dec + # o = self.dec(sin, z_slice, d, sid=g) + # tgt_o = self.dec(sin, tgt_z_slice, d, sid=tgt_g) + + # return (o, tgt_o), slice_id, x_mask, y_mask, ((z, z_p, m_p), logs_p, m_q, logs_q) + + def make_sin_d(self, f0): + # f0 から sin と d を作成 + # f0 : [b, 1, t] + # sin : [b, 1, t] + # d : [4][b, 1, t] + prod_upsample_scales = np.cumprod(self.upsample_scales) + dfs_batch = [] + for df, us in zip(self.dense_factors, prod_upsample_scales): + dilated_tensor = dilated_factor(f0, self.sample_rate, df) + # result += [torch.repeat_interleave(dilated_tensor, us, dim=1)] + result = [torch.stack([dilated_tensor for _ in range(us)], -1).reshape(dilated_tensor.shape[0], -1)] + dfs_batch.append(torch.cat(result, dim=0).unsqueeze(1)) + in_batch = self.signal_generator(f0) + + return in_batch, dfs_batch + + def make_random_target_sids(self, target_ids, sid): + # target_sids は target_ids をランダムで埋める + target_sids = torch.zeros_like(sid) + for i in range(len(target_sids)): + source_id = sid[i] + deleted_target_ids = target_ids[target_ids != source_id] # source_id と target_id が同じにならないよう sid と同じものを削除 + if len(deleted_target_ids) >= 1: + target_sids[i] = deleted_target_ids[torch.randint(len(deleted_target_ids), (1,))] + else: + # target_id 候補が無いときは仕方ないので sid を使う + target_sids[i] = source_id + return target_sids + + def voice_conversion(self, y, y_lengths, f0, sid_src, sid_tgt): + assert self.n_speakers > 0, "n_speakers have to be larger than 0." + sin, d = self.make_sin_d(f0) + g_src = self.emb_g(sid_src).unsqueeze(-1) + g_tgt = self.emb_g(sid_tgt).unsqueeze(-1) + z, _, _, y_mask = self.enc_q(y, y_lengths, g=g_src) + z_p = self.flow(z, y_mask, g=g_src) + z_hat = self.flow(z_p, y_mask, g=g_tgt, reverse=True) + # print("VC", sin.device, d[0].device, g_tgt.device) + o_hat = self.dec(sin, z_hat * y_mask, d, sid=g_tgt) + return o_hat[0] + + def voice_ra_pa_db(self, y, y_lengths, sid_src, sid_tgt): + assert self.n_speakers > 0, "n_speakers have to be larger than 0." + g_src = self.emb_g(sid_src).unsqueeze(-1) + g_tgt = self.emb_g(sid_tgt).unsqueeze(-1) + z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g_src) + o_hat = self.dec(z * y_mask, g=g_tgt) + return o_hat, y_mask, (z) + + def voice_ra_pa_da(self, y, y_lengths, sid_src, sid_tgt): + assert self.n_speakers > 0, "n_speakers have to be larger than 0." + g_src = self.emb_g(sid_src).unsqueeze(-1) + # g_tgt = self.emb_g(sid_tgt).unsqueeze(-1) + z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g_src) + o_hat = self.dec(z * y_mask, g=g_src) + return o_hat, y_mask, (z) + + def voice_conversion_cycle(self, y, y_lengths, sid_src, sid_tgt): + assert self.n_speakers > 0, "n_speakers have to be larger than 0." + g_src = self.emb_g(sid_src).unsqueeze(-1) + g_tgt = self.emb_g(sid_tgt).unsqueeze(-1) + z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g_src) + z_p = self.flow(z, y_mask, g=g_src) + z_hat = self.flow(z_p, y_mask, g=g_tgt, reverse=True) + z_p_hat = self.flow(z_hat, y_mask, g=g_tgt) + z_hat_hat = self.flow(z_p_hat, y_mask, g=g_src, reverse=True) + o_hat = self.dec(z_hat_hat * y_mask, g=g_tgt) + return o_hat, y_mask, (z, z_p, z_hat) + + def save_synthesizer(self, path): + enc_q = self.enc_q.state_dict() + dec = self.dec.state_dict() + emb_g = self.emb_g.state_dict() + torch.save({"enc_q": enc_q, "dec": dec, "emb_g": emb_g}, path) + + def load_synthesizer(self, path): + dict = torch.load(path, map_location="cpu") + enc_q = dict["enc_q"] + dec = dict["dec"] + emb_g = dict["emb_g"] + self.enc_q.load_state_dict(enc_q) + self.dec.load_state_dict(dec) + self.emb_g.load_state_dict(emb_g) diff --git a/server/voice_changer/MMVCv15/models/modules.py b/server/voice_changer/MMVCv15/models/modules.py new file mode 100644 index 00000000..14889c8e --- /dev/null +++ b/server/voice_changer/MMVCv15/models/modules.py @@ -0,0 +1,186 @@ +import torch +from torch import nn +from torch.nn import functional as F + +from torch.nn import Conv1d +from torch.nn.utils import weight_norm, remove_weight_norm + +from .commons import init_weights, get_padding, fused_add_tanh_sigmoid_multiply + + +LRELU_SLOPE = 0.1 + + +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 = p_dropout + + self.in_layers = torch.nn.ModuleList() + self.res_skip_layers = torch.nn.ModuleList() + self.drop = nn.Dropout(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, **kwargs): + 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 range(self.n_layers): + x_in = self.in_layers[i](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 = fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor) + acts = self.drop(acts) + + res_skip_acts = self.res_skip_layers[i](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) + + def forward(self, x, x_mask=None): + for c1, c2 in zip(self.convs1, self.convs2): + xt = F.leaky_relu(x, LRELU_SLOPE) + if x_mask is not None: + xt = xt * x_mask + xt = c1(xt) + xt = F.leaky_relu(xt, 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) + + def forward(self, x, x_mask=None): + for c in self.convs: + xt = F.leaky_relu(x, 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 Flip(nn.Module): + def forward(self, x, *args, reverse=False, **kwargs): + 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 + + +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().__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=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 diff --git a/server/voice_changer/MMVCv15/models/readme.txt b/server/voice_changer/MMVCv15/models/readme.txt new file mode 100644 index 00000000..5b4639fd --- /dev/null +++ b/server/voice_changer/MMVCv15/models/readme.txt @@ -0,0 +1 @@ +modules in this folder from https://github.com/isletennos/MMVC_Client.git at 461cb231b57cbb17243110eaac8435d9cca24a26 \ No newline at end of file diff --git a/server/voice_changer/MMVCv15/models/residual_block.py b/server/voice_changer/MMVCv15/models/residual_block.py new file mode 100644 index 00000000..88bffe4a --- /dev/null +++ b/server/voice_changer/MMVCv15/models/residual_block.py @@ -0,0 +1,257 @@ +# -*- coding: utf-8 -*- + +# Copyright 2022 Reo Yoneyama (Nagoya University) +# MIT License (https://opensource.org/licenses/MIT) + +"""Residual block modules. + +References: + - https://github.com/kan-bayashi/ParallelWaveGAN + - https://github.com/bigpon/QPPWG + - https://github.com/r9y9/wavenet_vocoder + +""" + +from logging import getLogger + +import torch +import torch.nn as nn +from .snake import Snake +from .index import index_initial, pd_indexing + +# A logger for this file +logger = getLogger(__name__) + + +class Conv1d(nn.Conv1d): + """Conv1d module with customized initialization.""" + + def __init__(self, *args, **kwargs): + """Initialize Conv1d module.""" + super(Conv1d, self).__init__(*args, **kwargs) + + def reset_parameters(self): + """Reset parameters.""" + nn.init.kaiming_normal_(self.weight, nonlinearity="relu") + if self.bias is not None: + nn.init.constant_(self.bias, 0.0) + + +class Conv1d1x1(Conv1d): + """1x1 Conv1d with customized initialization.""" + + def __init__(self, in_channels, out_channels, bias=True): + """Initialize 1x1 Conv1d module.""" + super(Conv1d1x1, self).__init__(in_channels, out_channels, kernel_size=1, padding=0, dilation=1, bias=bias) + + +class Conv2d(nn.Conv2d): + """Conv2d module with customized initialization.""" + + def __init__(self, *args, **kwargs): + """Initialize Conv2d module.""" + super(Conv2d, self).__init__(*args, **kwargs) + + def reset_parameters(self): + """Reset parameters.""" + nn.init.kaiming_normal_(self.weight, mode="fan_out", nonlinearity="relu") + if self.bias is not None: + nn.init.constant_(self.bias, 0.0) + + +class Conv2d1x1(Conv2d): + """1x1 Conv2d with customized initialization.""" + + def __init__(self, in_channels, out_channels, bias=True): + """Initialize 1x1 Conv2d module.""" + super(Conv2d1x1, self).__init__(in_channels, out_channels, kernel_size=1, padding=0, dilation=1, bias=bias) + + +class ResidualBlock(nn.Module): + """Residual block module in HiFiGAN.""" + + def __init__( + self, + kernel_size=3, + channels=512, + dilations=(1, 3, 5), + bias=True, + use_additional_convs=True, + nonlinear_activation="LeakyReLU", + nonlinear_activation_params={"negative_slope": 0.1}, + ): + """Initialize ResidualBlock module. + + Args: + kernel_size (int): Kernel size of dilation convolution layer. + channels (int): Number of channels for convolution layer. + dilations (List[int]): List of dilation factors. + use_additional_convs (bool): Whether to use additional convolution layers. + bias (bool): Whether to add bias parameter in convolution layers. + nonlinear_activation (str): Activation function module name. + nonlinear_activation_params (dict): Hyperparameters for activation function. + + """ + super().__init__() + self.use_additional_convs = use_additional_convs + self.convs1 = nn.ModuleList() + if use_additional_convs: + self.convs2 = nn.ModuleList() + assert kernel_size % 2 == 1, "Kernel size must be odd number." + for dilation in dilations: + if nonlinear_activation == "Snake": + nonlinear = Snake(channels, **nonlinear_activation_params) + else: + nonlinear = getattr(nn, nonlinear_activation)(**nonlinear_activation_params) + self.convs1 += [ + nn.Sequential( + nonlinear, + nn.Conv1d( + channels, + channels, + kernel_size, + dilation=dilation, + bias=bias, + padding=(kernel_size - 1) // 2 * dilation, + ), + ) + ] + if use_additional_convs: + if nonlinear_activation == "Snake": + nonlinear = Snake(channels, **nonlinear_activation_params) + else: + nonlinear = getattr(nn, nonlinear_activation)(**nonlinear_activation_params) + self.convs2 += [ + nn.Sequential( + nonlinear, + nn.Conv1d( + channels, + channels, + kernel_size, + dilation=1, + bias=bias, + padding=(kernel_size - 1) // 2, + ), + ) + ] + + def forward(self, x): + """Calculate forward propagation. + + Args: + x (Tensor): Input tensor (B, channels, T). + + Returns: + Tensor: Output tensor (B, channels, T). + + """ + for idx in range(len(self.convs1)): + xt = self.convs1[idx](x) + if self.use_additional_convs: + xt = self.convs2[idx](xt) + x = xt + x + return x + + +class AdaptiveResidualBlock(nn.Module): + """Residual block module in HiFiGAN.""" + + def __init__( + self, + kernel_size=3, + channels=512, + dilations=(1, 2, 4), + bias=True, + use_additional_convs=True, + nonlinear_activation="LeakyReLU", + nonlinear_activation_params={"negative_slope": 0.1}, + ): + """Initialize ResidualBlock module. + + Args: + kernel_size (int): Kernel size of dilation convolution layer. + channels (int): Number of channels for convolution layer. + bias (bool): Whether to add bias parameter in convolution layers. + nonlinear_activation (str): Activation function module name. + nonlinear_activation_params (dict): Hyperparameters for activation function. + + """ + super().__init__() + self.use_additional_convs = use_additional_convs + assert kernel_size == 3, "Currently only kernel_size = 3 is supported." + self.channels = channels + self.dilations = dilations + self.nonlinears = nn.ModuleList() + self.convsC = nn.ModuleList() + self.convsP = nn.ModuleList() + self.convsF = nn.ModuleList() + if use_additional_convs: + self.convsA = nn.ModuleList() + for _ in dilations: + if nonlinear_activation == "Snake": + self.nonlinears += [Snake(channels, **nonlinear_activation_params)] + else: + self.nonlinears += [getattr(nn, nonlinear_activation)(**nonlinear_activation_params)] + self.convsC += [ + Conv1d1x1( + channels, + channels, + bias=bias, + ), + ] + self.convsP += [ + Conv1d1x1( + channels, + channels, + bias=bias, + ), + ] + self.convsF += [ + Conv1d1x1( + channels, + channels, + bias=bias, + ), + ] + if use_additional_convs: + if nonlinear_activation == "Snake": + nonlinear = Snake(channels, **nonlinear_activation_params) + else: + nonlinear = getattr(nn, nonlinear_activation)(**nonlinear_activation_params) + self.convsA += [ + nn.Sequential( + nonlinear, + nn.Conv1d( + channels, + channels, + kernel_size, + dilation=1, + bias=bias, + padding=(kernel_size - 1) // 2, + ), + ) + ] + + def forward(self, x, d): + """Calculate forward propagation. + + Args: + x (Tensor): Input tensor (B, channels, T). + d (Tensor): Input pitch-dependent dilated factors (B, 1, T). + + Returns: + Tensor: Output tensor (B, channels, T). + + """ + batch_index, ch_index = index_initial(x.size(0), self.channels, tensor=False) + batch_index = torch.tensor(batch_index).to(x.device) + ch_index = torch.tensor(ch_index).to(x.device) + + for i, dilation in enumerate(self.dilations): + xt = self.nonlinears[i](x) + xP, xF = pd_indexing(xt, d, dilation, batch_index, ch_index) + xt = self.convsC[i](xt) + self.convsP[i](xP) + self.convsF[i](xF) + if self.use_additional_convs: + xt = self.convsA[i](xt) + x = xt + x + return x diff --git a/server/voice_changer/MMVCv15/models/snake.py b/server/voice_changer/MMVCv15/models/snake.py new file mode 100644 index 00000000..651736b9 --- /dev/null +++ b/server/voice_changer/MMVCv15/models/snake.py @@ -0,0 +1,47 @@ +# -*- coding: utf-8 -*- + +# Copyright 2022 Reo Yoneyama (Nagoya University) +# MIT License (https://opensource.org/licenses/MIT) + +"""Snake Activation Function Module. + +References: + - Neural Networks Fail to Learn Periodic Functions and How to Fix It + https://arxiv.org/pdf/2006.08195.pdf + - BigVGAN: A Universal Neural Vocoder with Large-Scale Training + https://arxiv.org/pdf/2206.04658.pdf + +""" + +import torch +import torch.nn as nn + + +class Snake(nn.Module): + """Snake activation function module.""" + + def __init__(self, channels, init=50): + """Initialize Snake module. + + Args: + channels (int): Number of feature channels. + init (float): Initial value of the learnable parameter alpha. + According to the original paper, 5 ~ 50 would be + suitable for periodic data (i.e. voices). + + """ + super(Snake, self).__init__() + alpha = init * torch.ones(1, channels, 1) + self.alpha = nn.Parameter(alpha) + + def forward(self, x): + """Calculate forward propagation. + + Args: + x (Tensor): Input noise signal (B, channels, T). + + Returns: + Tensor: Output tensor (B, channels, T). + + """ + return x + torch.sin(self.alpha * x) ** 2 / self.alpha