6Simple9/ChatTTS-OpenVoice
9
1import math2import torch3from torch import nn4from torch.nn import functional as F5 6from . import commons7import logging8 9logger = logging.getLogger(__name__)10 11 12class LayerNorm(nn.Module):13 def __init__(self, channels, eps=1e-5):14 super().__init__()15 self.channels = channels16 self.eps = eps17 18 self.gamma = nn.Parameter(torch.ones(channels))19 self.beta = nn.Parameter(torch.zeros(channels))20 21 def forward(self, x):22 x = x.transpose(1, -1)23 x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)24 return x.transpose(1, -1)25 26 27@torch.jit.script28def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):29 n_channels_int = n_channels[0]30 in_act = input_a + input_b31 t_act = torch.tanh(in_act[:, :n_channels_int, :])32 s_act = torch.sigmoid(in_act[:, n_channels_int:, :])33 acts = t_act * s_act34 return acts35 36 37class Encoder(nn.Module):38 def __init__(39 self,40 hidden_channels,41 filter_channels,42 n_heads,43 n_layers,44 kernel_size=1,45 p_dropout=0.0,46 window_size=4,47 isflow=True,48 **kwargs49 ):50 super().__init__()51 self.hidden_channels = hidden_channels52 self.filter_channels = filter_channels53 self.n_heads = n_heads54 self.n_layers = n_layers55 self.kernel_size = kernel_size56 self.p_dropout = p_dropout57 self.window_size = window_size58 # if isflow:59 # cond_layer = torch.nn.Conv1d(256, 2*hidden_channels*n_layers, 1)60 # self.cond_pre = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, 1)61 # self.cond_layer = weight_norm(cond_layer, name='weight')62 # self.gin_channels = 25663 self.cond_layer_idx = self.n_layers64 if "gin_channels" in kwargs:65 self.gin_channels = kwargs["gin_channels"]66 if self.gin_channels != 0:67 self.spk_emb_linear = nn.Linear(self.gin_channels, self.hidden_channels)68 # vits2 says 3rd block, so idx is 2 by default69 self.cond_layer_idx = (70 kwargs["cond_layer_idx"] if "cond_layer_idx" in kwargs else 271 )72 # logging.debug(self.gin_channels, self.cond_layer_idx)73 assert (74 self.cond_layer_idx < self.n_layers75 ), "cond_layer_idx should be less than n_layers"76 self.drop = nn.Dropout(p_dropout)77 self.attn_layers = nn.ModuleList()78 self.norm_layers_1 = nn.ModuleList()79 self.ffn_layers = nn.ModuleList()80 self.norm_layers_2 = nn.ModuleList()81 82 for i in range(self.n_layers):83 self.attn_layers.append(84 MultiHeadAttention(85 hidden_channels,86 hidden_channels,87 n_heads,88 p_dropout=p_dropout,89 window_size=window_size,90 )91 )92 self.norm_layers_1.append(LayerNorm(hidden_channels))93 self.ffn_layers.append(94 FFN(95 hidden_channels,96 hidden_channels,97 filter_channels,98 kernel_size,99 p_dropout=p_dropout,100 )101 )102 self.norm_layers_2.append(LayerNorm(hidden_channels))103 104 def forward(self, x, x_mask, g=None):105 attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)106 x = x * x_mask107 for i in range(self.n_layers):108 if i == self.cond_layer_idx and g is not None:109 g = self.spk_emb_linear(g.transpose(1, 2))110 g = g.transpose(1, 2)111 x = x + g112 x = x * x_mask113 y = self.attn_layers[i](x, x, attn_mask)114 y = self.drop(y)115 x = self.norm_layers_1[i](x + y)116 117 y = self.ffn_layers[i](x, x_mask)118 y = self.drop(y)119 x = self.norm_layers_2[i](x + y)120 x = x * x_mask121 return x122 123 124class Decoder(nn.Module):125 def __init__(126 self,127 hidden_channels,128 filter_channels,129 n_heads,130 n_layers,131 kernel_size=1,132 p_dropout=0.0,133 proximal_bias=False,134 proximal_init=True,135 **kwargs136 ):137 super().__init__()138 self.hidden_channels = hidden_channels139 self.filter_channels = filter_channels140 self.n_heads = n_heads141 self.n_layers = n_layers142 self.kernel_size = kernel_size143 self.p_dropout = p_dropout144 self.proximal_bias = proximal_bias145 self.proximal_init = proximal_init146 147 self.drop = nn.Dropout(p_dropout)148 self.self_attn_layers = nn.ModuleList()149 self.norm_layers_0 = nn.ModuleList()150 self.encdec_attn_layers = nn.ModuleList()151 self.norm_layers_1 = nn.ModuleList()152 self.ffn_layers = nn.ModuleList()153 self.norm_layers_2 = nn.ModuleList()154 for i in range(self.n_layers):155 self.self_attn_layers.append(156 MultiHeadAttention(157 hidden_channels,158 hidden_channels,159 n_heads,160 p_dropout=p_dropout,161 proximal_bias=proximal_bias,162 proximal_init=proximal_init,163 )164 )165 self.norm_layers_0.append(LayerNorm(hidden_channels))166 self.encdec_attn_layers.append(167 MultiHeadAttention(168 hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout169 )170 )171 self.norm_layers_1.append(LayerNorm(hidden_channels))172 self.ffn_layers.append(173 FFN(174 hidden_channels,175 hidden_channels,176 filter_channels,177 kernel_size,178 p_dropout=p_dropout,179 causal=True,180 )181 )182 self.norm_layers_2.append(LayerNorm(hidden_channels))183 184 def forward(self, x, x_mask, h, h_mask):185 """186 x: decoder input187 h: encoder output188 """189 self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(190 device=x.device, dtype=x.dtype191 )192 encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)193 x = x * x_mask194 for i in range(self.n_layers):195 y = self.self_attn_layers[i](x, x, self_attn_mask)196 y = self.drop(y)197 x = self.norm_layers_0[i](x + y)198 199 y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)200 y = self.drop(y)201 x = self.norm_layers_1[i](x + y)202 203 y = self.ffn_layers[i](x, x_mask)204 y = self.drop(y)205 x = self.norm_layers_2[i](x + y)206 x = x * x_mask207 return x208 209 210class MultiHeadAttention(nn.Module):211 def __init__(212 self,213 channels,214 out_channels,215 n_heads,216 p_dropout=0.0,217 window_size=None,218 heads_share=True,219 block_length=None,220 proximal_bias=False,221 proximal_init=False,222 ):223 super().__init__()224 assert channels % n_heads == 0225 226 self.channels = channels227 self.out_channels = out_channels228 self.n_heads = n_heads229 self.p_dropout = p_dropout230 self.window_size = window_size231 self.heads_share = heads_share232 self.block_length = block_length233 self.proximal_bias = proximal_bias234 self.proximal_init = proximal_init235 self.attn = None236 237 self.k_channels = channels // n_heads238 self.conv_q = nn.Conv1d(channels, channels, 1)239 self.conv_k = nn.Conv1d(channels, channels, 1)240 self.conv_v = nn.Conv1d(channels, channels, 1)241 self.conv_o = nn.Conv1d(channels, out_channels, 1)242 self.drop = nn.Dropout(p_dropout)243 244 if window_size is not None:245 n_heads_rel = 1 if heads_share else n_heads246 rel_stddev = self.k_channels**-0.5247 self.emb_rel_k = nn.Parameter(248 torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)249 * rel_stddev250 )251 self.emb_rel_v = nn.Parameter(252 torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)253 * rel_stddev254 )255 256 nn.init.xavier_uniform_(self.conv_q.weight)257 nn.init.xavier_uniform_(self.conv_k.weight)258 nn.init.xavier_uniform_(self.conv_v.weight)259 if proximal_init:260 with torch.no_grad():261 self.conv_k.weight.copy_(self.conv_q.weight)262 self.conv_k.bias.copy_(self.conv_q.bias)263 264 def forward(self, x, c, attn_mask=None):265 q = self.conv_q(x)266 k = self.conv_k(c)267 v = self.conv_v(c)268 269 x, self.attn = self.attention(q, k, v, mask=attn_mask)270 271 x = self.conv_o(x)272 return x273 274 def attention(self, query, key, value, mask=None):275 # reshape [b, d, t] -> [b, n_h, t, d_k]276 b, d, t_s, t_t = (*key.size(), query.size(2))277 query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)278 key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)279 value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)280 281 scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))282 if self.window_size is not None:283 assert (284 t_s == t_t285 ), "Relative attention is only available for self-attention."286 key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)287 rel_logits = self._matmul_with_relative_keys(288 query / math.sqrt(self.k_channels), key_relative_embeddings289 )290 scores_local = self._relative_position_to_absolute_position(rel_logits)291 scores = scores + scores_local292 if self.proximal_bias:293 assert t_s == t_t, "Proximal bias is only available for self-attention."294 scores = scores + self._attention_bias_proximal(t_s).to(295 device=scores.device, dtype=scores.dtype296 )297 if mask is not None:298 scores = scores.masked_fill(mask == 0, -1e4)299 if self.block_length is not None:300 assert (301 t_s == t_t302 ), "Local attention is only available for self-attention."303 block_mask = (304 torch.ones_like(scores)305 .triu(-self.block_length)306 .tril(self.block_length)307 )308 scores = scores.masked_fill(block_mask == 0, -1e4)309 p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s]310 p_attn = self.drop(p_attn)311 output = torch.matmul(p_attn, value)312 if self.window_size is not None:313 relative_weights = self._absolute_position_to_relative_position(p_attn)314 value_relative_embeddings = self._get_relative_embeddings(315 self.emb_rel_v, t_s316 )317 output = output + self._matmul_with_relative_values(318 relative_weights, value_relative_embeddings319 )320 output = (321 output.transpose(2, 3).contiguous().view(b, d, t_t)322 ) # [b, n_h, t_t, d_k] -> [b, d, t_t]323 return output, p_attn324 325 def _matmul_with_relative_values(self, x, y):326 """327 x: [b, h, l, m]328 y: [h or 1, m, d]329 ret: [b, h, l, d]330 """331 ret = torch.matmul(x, y.unsqueeze(0))332 return ret333 334 def _matmul_with_relative_keys(self, x, y):335 """336 x: [b, h, l, d]337 y: [h or 1, m, d]338 ret: [b, h, l, m]339 """340 ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))341 return ret342 343 def _get_relative_embeddings(self, relative_embeddings, length):344 2 * self.window_size + 1345 # Pad first before slice to avoid using cond ops.346 pad_length = max(length - (self.window_size + 1), 0)347 slice_start_position = max((self.window_size + 1) - length, 0)348 slice_end_position = slice_start_position + 2 * length - 1349 if pad_length > 0:350 padded_relative_embeddings = F.pad(351 relative_embeddings,352 commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]),353 )354 else:355 padded_relative_embeddings = relative_embeddings356 used_relative_embeddings = padded_relative_embeddings[357 :, slice_start_position:slice_end_position358 ]359 return used_relative_embeddings360 361 def _relative_position_to_absolute_position(self, x):362 """363 x: [b, h, l, 2*l-1]364 ret: [b, h, l, l]365 """366 batch, heads, length, _ = x.size()367 # Concat columns of pad to shift from relative to absolute indexing.368 x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]]))369 370 # Concat extra elements so to add up to shape (len+1, 2*len-1).371 x_flat = x.view([batch, heads, length * 2 * length])372 x_flat = F.pad(373 x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [0, length - 1]])374 )375 376 # Reshape and slice out the padded elements.377 x_final = x_flat.view([batch, heads, length + 1, 2 * length - 1])[378 :, :, :length, length - 1 :379 ]380 return x_final381 382 def _absolute_position_to_relative_position(self, x):383 """384 x: [b, h, l, l]385 ret: [b, h, l, 2*l-1]386 """387 batch, heads, length, _ = x.size()388 # pad along column389 x = F.pad(390 x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length - 1]])391 )392 x_flat = x.view([batch, heads, length**2 + length * (length - 1)])393 # add 0's in the beginning that will skew the elements after reshape394 x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))395 x_final = x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:]396 return x_final397 398 def _attention_bias_proximal(self, length):399 """Bias for self-attention to encourage attention to close positions.400 Args:401 length: an integer scalar.402 Returns:403 a Tensor with shape [1, 1, length, length]404 """405 r = torch.arange(length, dtype=torch.float32)406 diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)407 return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)408 409 410class FFN(nn.Module):411 def __init__(412 self,413 in_channels,414 out_channels,415 filter_channels,416 kernel_size,417 p_dropout=0.0,418 activation=None,419 causal=False,420 ):421 super().__init__()422 self.in_channels = in_channels423 self.out_channels = out_channels424 self.filter_channels = filter_channels425 self.kernel_size = kernel_size426 self.p_dropout = p_dropout427 self.activation = activation428 self.causal = causal429 430 if causal:431 self.padding = self._causal_padding432 else:433 self.padding = self._same_padding434 435 self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)436 self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)437 self.drop = nn.Dropout(p_dropout)438 439 def forward(self, x, x_mask):440 x = self.conv_1(self.padding(x * x_mask))441 if self.activation == "gelu":442 x = x * torch.sigmoid(1.702 * x)443 else:444 x = torch.relu(x)445 x = self.drop(x)446 x = self.conv_2(self.padding(x * x_mask))447 return x * x_mask448 449 def _causal_padding(self, x):450 if self.kernel_size == 1:451 return x452 pad_l = self.kernel_size - 1453 pad_r = 0454 padding = [[0, 0], [0, 0], [pad_l, pad_r]]455 x = F.pad(x, commons.convert_pad_shape(padding))456 return x457 458 def _same_padding(self, x):459 if self.kernel_size == 1:460 return x461 pad_l = (self.kernel_size - 1) // 2462 pad_r = self.kernel_size // 2463 padding = [[0, 0], [0, 0], [pad_l, pad_r]]464 x = F.pad(x, commons.convert_pad_shape(padding))465 return x466