import math import torch import torch.nn as nn import torch.nn.functional as F from math import sqrt Linear = nn.Linear ConvTranspose2d = nn.ConvTranspose2d class Mish(nn.Module): def forward(self, x): return x * torch.tanh(F.softplus(x)) class SinusoidalPosEmb(nn.Module): def __init__(self, dim): super().__init__() self.dim = dim def forward(self, x): device = x.device half_dim = self.dim // 2 emb = math.log(10000) / (half_dim - 1) emb = torch.exp(torch.arange(half_dim, device=device) * -emb) emb = x[:, None] * emb[None, :] emb = torch.cat((emb.sin(), emb.cos()), dim=-1) return emb def Conv1d(*args, **kwargs): layer = nn.Conv1d(*args, **kwargs) nn.init.kaiming_normal_(layer.weight) return layer class ResidualBlock(nn.Module): def __init__(self, encoder_hidden, residual_channels, dilation): super().__init__() self.dilated_conv = Conv1d(residual_channels, 2 * residual_channels, 3, padding=dilation, dilation=dilation) self.diffusion_projection = Linear(residual_channels, residual_channels) self.conditioner_projection = Conv1d(encoder_hidden, 2 * residual_channels, 1) self.output_projection = Conv1d(residual_channels, 2 * residual_channels, 1) def forward(self, x, conditioner, diffusion_step): diffusion_step = self.diffusion_projection(diffusion_step).unsqueeze(-1) conditioner = self.conditioner_projection(conditioner) y = x + diffusion_step y = self.dilated_conv(y) + conditioner gate, filter = torch.chunk(y, 2, dim=1) y = torch.sigmoid(gate) * torch.tanh(filter) y = self.output_projection(y) residual, skip = torch.chunk(y, 2, dim=1) return (x + residual) / sqrt(2.0), skip class DiffNet(nn.Module): def __init__(self, hparams): super().__init__() in_dims = hparams['audio_num_mel_bins'] self.encoder_hidden = hparams['hidden_size'] self.residual_layers = hparams['residual_layers'] self.residual_channels = hparams['residual_channels'] self.dilation_cycle_length = hparams['dilation_cycle_length'] self.input_projection = Conv1d(in_dims, self.residual_channels, 1) self.diffusion_embedding = SinusoidalPosEmb(self.residual_channels) dim = self.residual_channels self.mlp = nn.Sequential( nn.Linear(dim, dim * 4), Mish(), nn.Linear(dim * 4, dim) ) self.residual_layers = nn.ModuleList([ ResidualBlock(self.encoder_hidden, self.residual_channels, 2 ** (i % self.dilation_cycle_length)) for i in range(self.residual_layers) ]) self.skip_projection = Conv1d(self.residual_channels, self.residual_channels, 1) self.output_projection = Conv1d(self.residual_channels, in_dims, 1) nn.init.zeros_(self.output_projection.weight) def forward(self, spec, diffusion_step, cond): """ :param spec: [B, 1, M, T] :param diffusion_step: [B, 1] :param cond: [B, M, T] :return: """ x = spec[:, 0] x = self.input_projection(x) # x [B, residual_channel, T] x = F.relu(x) diffusion_step = self.diffusion_embedding(diffusion_step) diffusion_step = self.mlp(diffusion_step) skip = [] for layer_id, layer in enumerate(self.residual_layers): x, skip_connection = layer(x, cond, diffusion_step) skip.append(skip_connection) x = torch.sum(torch.stack(skip), dim=0) / sqrt(len(self.residual_layers)) x = self.skip_projection(x) x = F.relu(x) x = self.output_projection(x) # [B, 80, T] return x[:, None, :, :]