Spaces:
Running
Running
| # Copyright (c) 2023 Amphion. | |
| # | |
| # This source code is licensed under the MIT license found in the | |
| # LICENSE file in the root directory of this source tree. | |
| import math | |
| import torch | |
| import numpy as np | |
| import torch.nn as nn | |
| from utils.util import convert_pad_shape | |
| class BaseModule(torch.nn.Module): | |
| def __init__(self): | |
| super(BaseModule, self).__init__() | |
| def nparams(self): | |
| """ | |
| Returns number of trainable parameters of the module. | |
| """ | |
| num_params = 0 | |
| for name, param in self.named_parameters(): | |
| if param.requires_grad: | |
| num_params += np.prod(param.detach().cpu().numpy().shape) | |
| return num_params | |
| def relocate_input(self, x: list): | |
| """ | |
| Relocates provided tensors to the same device set for the module. | |
| """ | |
| device = next(self.parameters()).device | |
| for i in range(len(x)): | |
| if isinstance(x[i], torch.Tensor) and x[i].device != device: | |
| x[i] = x[i].to(device) | |
| return x | |
| class LayerNorm(BaseModule): | |
| def __init__(self, channels, eps=1e-4): | |
| super(LayerNorm, self).__init__() | |
| self.channels = channels | |
| self.eps = eps | |
| self.gamma = torch.nn.Parameter(torch.ones(channels)) | |
| self.beta = torch.nn.Parameter(torch.zeros(channels)) | |
| def forward(self, x): | |
| n_dims = len(x.shape) | |
| mean = torch.mean(x, 1, keepdim=True) | |
| variance = torch.mean((x - mean) ** 2, 1, keepdim=True) | |
| x = (x - mean) * torch.rsqrt(variance + self.eps) | |
| shape = [1, -1] + [1] * (n_dims - 2) | |
| x = x * self.gamma.view(*shape) + self.beta.view(*shape) | |
| return x | |
| class ConvReluNorm(BaseModule): | |
| def __init__( | |
| self, | |
| in_channels, | |
| hidden_channels, | |
| out_channels, | |
| kernel_size, | |
| n_layers, | |
| p_dropout, | |
| eps=1e-5, | |
| ): | |
| 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 = p_dropout | |
| self.eps = eps | |
| self.conv_layers = torch.nn.ModuleList() | |
| self.conv_layers.append( | |
| torch.nn.Conv1d( | |
| in_channels, hidden_channels, kernel_size, padding=kernel_size // 2 | |
| ) | |
| ) | |
| self.relu_drop = torch.nn.Sequential( | |
| torch.nn.ReLU(), torch.nn.Dropout(p_dropout) | |
| ) | |
| for _ in range(n_layers - 1): | |
| self.conv_layers.append( | |
| torch.nn.Conv1d( | |
| hidden_channels, | |
| hidden_channels, | |
| kernel_size, | |
| padding=kernel_size // 2, | |
| ) | |
| ) | |
| self.proj = torch.nn.Conv1d(hidden_channels, out_channels, 1) | |
| self.proj.weight.data.zero_() | |
| self.proj.bias.data.zero_() | |
| def forward(self, x, x_mask): | |
| for i in range(self.n_layers): | |
| x = self.conv_layers[i](x * x_mask) | |
| x = self.instance_norm(x, x_mask) | |
| x = self.relu_drop(x) | |
| x = self.proj(x) | |
| return x * x_mask | |
| def instance_norm(self, x, mask, return_mean_std=False): | |
| mean, std = self.calc_mean_std(x, mask) | |
| x = (x - mean) / std | |
| if return_mean_std: | |
| return x, mean, std | |
| else: | |
| return x | |
| def calc_mean_std(self, x, mask=None): | |
| x = x * mask | |
| B, C = x.shape[:2] | |
| mn = x.view(B, C, -1).mean(-1) | |
| sd = (x.view(B, C, -1).var(-1) + self.eps).sqrt() | |
| mn = mn.view(B, C, *((len(x.shape) - 2) * [1])) | |
| sd = sd.view(B, C, *((len(x.shape) - 2) * [1])) | |
| return mn, sd | |
| class MultiHeadAttention(BaseModule): | |
| def __init__( | |
| self, | |
| channels, | |
| out_channels, | |
| n_heads, | |
| window_size=None, | |
| heads_share=True, | |
| p_dropout=0.0, | |
| 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.window_size = window_size | |
| self.heads_share = heads_share | |
| self.proximal_bias = proximal_bias | |
| self.p_dropout = p_dropout | |
| self.attn = None | |
| self.k_channels = channels // n_heads | |
| self.conv_q = torch.nn.Conv1d(channels, channels, 1) | |
| self.conv_k = torch.nn.Conv1d(channels, channels, 1) | |
| self.conv_v = torch.nn.Conv1d(channels, channels, 1) | |
| 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 = torch.nn.Parameter( | |
| torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) | |
| * rel_stddev | |
| ) | |
| self.emb_rel_v = torch.nn.Parameter( | |
| torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) | |
| * rel_stddev | |
| ) | |
| self.conv_o = torch.nn.Conv1d(channels, out_channels, 1) | |
| self.drop = torch.nn.Dropout(p_dropout) | |
| torch.nn.init.xavier_uniform_(self.conv_q.weight) | |
| torch.nn.init.xavier_uniform_(self.conv_k.weight) | |
| if proximal_init: | |
| self.conv_k.weight.data.copy_(self.conv_q.weight.data) | |
| self.conv_k.bias.data.copy_(self.conv_q.bias.data) | |
| torch.nn.init.xavier_uniform_(self.conv_v.weight) | |
| 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.attn = self.attention(q, k, v, mask=attn_mask) | |
| x = self.conv_o(x) | |
| return x | |
| def attention(self, query, key, value, mask=None): | |
| b, d, t_s, t_t = (*key.size(), 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, key.transpose(-2, -1)) / math.sqrt(self.k_channels) | |
| 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, key_relative_embeddings) | |
| rel_logits = self._relative_position_to_absolute_position(rel_logits) | |
| scores_local = rel_logits / math.sqrt(self.k_channels) | |
| 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) | |
| p_attn = torch.nn.functional.softmax(scores, dim=-1) | |
| 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) | |
| return output, p_attn | |
| def _matmul_with_relative_values(self, x, y): | |
| ret = torch.matmul(x, y.unsqueeze(0)) | |
| return ret | |
| def _matmul_with_relative_keys(self, x, y): | |
| ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1)) | |
| return ret | |
| def _get_relative_embeddings(self, relative_embeddings, length): | |
| 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 = torch.nn.functional.pad( | |
| relative_embeddings, | |
| convert_pad_shape([[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): | |
| batch, heads, length, _ = x.size() | |
| x = torch.nn.functional.pad( | |
| x, convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]]) | |
| ) | |
| x_flat = x.view([batch, heads, length * 2 * length]) | |
| x_flat = torch.nn.functional.pad( | |
| x_flat, convert_pad_shape([[0, 0], [0, 0], [0, length - 1]]) | |
| ) | |
| 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): | |
| batch, heads, length, _ = x.size() | |
| x = torch.nn.functional.pad( | |
| x, convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length - 1]]) | |
| ) | |
| x_flat = x.view([batch, heads, length**2 + length * (length - 1)]) | |
| x_flat = torch.nn.functional.pad( | |
| x_flat, convert_pad_shape([[0, 0], [0, 0], [length, 0]]) | |
| ) | |
| x_final = x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:] | |
| return x_final | |
| def _attention_bias_proximal(self, 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(BaseModule): | |
| def __init__( | |
| self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0.0 | |
| ): | |
| 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.conv_1 = torch.nn.Conv1d( | |
| in_channels, filter_channels, kernel_size, padding=kernel_size // 2 | |
| ) | |
| self.conv_2 = torch.nn.Conv1d( | |
| filter_channels, out_channels, kernel_size, padding=kernel_size // 2 | |
| ) | |
| self.drop = torch.nn.Dropout(p_dropout) | |
| def forward(self, x, x_mask): | |
| x = self.conv_1(x * x_mask) | |
| x = torch.relu(x) | |
| x = self.drop(x) | |
| x = self.conv_2(x * x_mask) | |
| return x * x_mask | |
| class Encoder(BaseModule): | |
| def __init__( | |
| self, | |
| hidden_channels, | |
| filter_channels, | |
| n_heads=2, | |
| n_layers=6, | |
| kernel_size=3, | |
| p_dropout=0.1, | |
| window_size=4, | |
| **kwargs | |
| ): | |
| super(Encoder, 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.window_size = window_size | |
| self.drop = torch.nn.Dropout(p_dropout) | |
| self.attn_layers = torch.nn.ModuleList() | |
| self.norm_layers_1 = torch.nn.ModuleList() | |
| self.ffn_layers = torch.nn.ModuleList() | |
| self.norm_layers_2 = torch.nn.ModuleList() | |
| for _ in range(self.n_layers): | |
| self.attn_layers.append( | |
| MultiHeadAttention( | |
| hidden_channels, | |
| hidden_channels, | |
| n_heads, | |
| window_size=window_size, | |
| 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, | |
| ) | |
| ) | |
| self.norm_layers_2.append(LayerNorm(hidden_channels)) | |
| def forward(self, x, x_mask): | |
| attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1) | |
| for i in range(self.n_layers): | |
| x = x * x_mask | |
| y = self.attn_layers[i](x, x, 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 Conformer(BaseModule): | |
| def __init__(self, cfg): | |
| super().__init__() | |
| self.cfg = cfg | |
| self.n_heads = self.cfg.n_heads | |
| self.n_layers = self.cfg.n_layers | |
| self.hidden_channels = self.cfg.input_dim | |
| self.filter_channels = self.cfg.filter_channels | |
| self.output_dim = self.cfg.output_dim | |
| self.dropout = self.cfg.dropout | |
| self.conformer_encoder = Encoder( | |
| self.hidden_channels, | |
| self.filter_channels, | |
| n_heads=self.n_heads, | |
| n_layers=self.n_layers, | |
| kernel_size=3, | |
| p_dropout=self.dropout, | |
| window_size=4, | |
| ) | |
| self.projection = nn.Conv1d(self.hidden_channels, self.output_dim, 1) | |
| def forward(self, x, x_mask): | |
| """ | |
| Args: | |
| x: (N, seq_len, input_dim) | |
| Returns: | |
| output: (N, seq_len, output_dim) | |
| """ | |
| # (N, seq_len, d_model) | |
| x = x.transpose(1, 2) | |
| x_mask = x_mask.transpose(1, 2) | |
| output = self.conformer_encoder(x, x_mask) | |
| # (N, seq_len, output_dim) | |
| output = self.projection(output) | |
| output = output.transpose(1, 2) | |
| return output | |