hanfish/LSai
0
1import math2import torch3from torch import nn4from torch.nn import functional as F5 6from module import commons7from module. modules import LayerNorm8 9 10class Encoder(nn.Module):11 def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., window_size=4,isflow=False, **kwargs):12 super().__init__()13 self.hidden_channels = hidden_channels14 self.filter_channels = filter_channels15 self.n_heads = n_heads16 self.n_layers = n_layers17 self.kernel_size = kernel_size18 self.p_dropout = p_dropout19 self.window_size = window_size20 21 self.drop = nn.Dropout(p_dropout)22 self.attn_layers = nn.ModuleList()23 self.norm_layers_1 = nn.ModuleList()24 self.ffn_layers = nn.ModuleList()25 self.norm_layers_2 = nn.ModuleList()26 for i in range(self.n_layers):27 self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size))28 self.norm_layers_1.append(LayerNorm(hidden_channels))29 self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout))30 self.norm_layers_2.append(LayerNorm(hidden_channels))31 if isflow:32 cond_layer = torch.nn.Conv1d(kwargs["gin_channels"], 2*hidden_channels*n_layers, 1)33 self.cond_pre = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, 1)34 self.cond_layer = weight_norm_modules(cond_layer, name='weight')35 self.gin_channels = kwargs["gin_channels"]36 def forward(self, x, x_mask, g=None):37 attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)38 x = x * x_mask39 if g is not None:40 g = self.cond_layer(g)41 42 for i in range(self.n_layers):43 if g is not None:44 x = self.cond_pre(x)45 cond_offset = i * 2 * self.hidden_channels46 g_l = g[:,cond_offset:cond_offset+2*self.hidden_channels,:]47 x = commons.fused_add_tanh_sigmoid_multiply(48 x,49 g_l,50 torch.IntTensor([self.hidden_channels]))51 y = self.attn_layers[i](x, x, attn_mask)52 y = self.drop(y)53 x = self.norm_layers_1[i](x + y)54 55 y = self.ffn_layers[i](x, x_mask)56 y = self.drop(y)57 x = self.norm_layers_2[i](x + y)58 x = x * x_mask59 return x60 61 62class Decoder(nn.Module):63 def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., proximal_bias=False, proximal_init=True, **kwargs):64 super().__init__()65 self.hidden_channels = hidden_channels66 self.filter_channels = filter_channels67 self.n_heads = n_heads68 self.n_layers = n_layers69 self.kernel_size = kernel_size70 self.p_dropout = p_dropout71 self.proximal_bias = proximal_bias72 self.proximal_init = proximal_init73 74 self.drop = nn.Dropout(p_dropout)75 self.self_attn_layers = nn.ModuleList()76 self.norm_layers_0 = nn.ModuleList()77 self.encdec_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 for i in range(self.n_layers):82 self.self_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias, proximal_init=proximal_init))83 self.norm_layers_0.append(LayerNorm(hidden_channels))84 self.encdec_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout))85 self.norm_layers_1.append(LayerNorm(hidden_channels))86 self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))87 self.norm_layers_2.append(LayerNorm(hidden_channels))88 89 def forward(self, x, x_mask, h, h_mask):90 """91 x: decoder input92 h: encoder output93 """94 self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(device=x.device, dtype=x.dtype)95 encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)96 x = x * x_mask97 for i in range(self.n_layers):98 y = self.self_attn_layers[i](x, x, self_attn_mask)99 y = self.drop(y)100 x = self.norm_layers_0[i](x + y)101 102 y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)103 y = self.drop(y)104 x = self.norm_layers_1[i](x + y)105 106 y = self.ffn_layers[i](x, x_mask)107 y = self.drop(y)108 x = self.norm_layers_2[i](x + y)109 x = x * x_mask110 return x111 112 113class MultiHeadAttention(nn.Module):114 def __init__(self, channels, out_channels, n_heads, p_dropout=0., window_size=None, heads_share=True, block_length=None, proximal_bias=False, proximal_init=False):115 super().__init__()116 assert channels % n_heads == 0117 118 self.channels = channels119 self.out_channels = out_channels120 self.n_heads = n_heads121 self.p_dropout = p_dropout122 self.window_size = window_size123 self.heads_share = heads_share124 self.block_length = block_length125 self.proximal_bias = proximal_bias126 self.proximal_init = proximal_init127 self.attn = None128 129 self.k_channels = channels // n_heads130 self.conv_q = nn.Conv1d(channels, channels, 1)131 self.conv_k = nn.Conv1d(channels, channels, 1)132 self.conv_v = nn.Conv1d(channels, channels, 1)133 self.conv_o = nn.Conv1d(channels, out_channels, 1)134 self.drop = nn.Dropout(p_dropout)135 136 if window_size is not None:137 n_heads_rel = 1 if heads_share else n_heads138 rel_stddev = self.k_channels**-0.5139 self.emb_rel_k = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)140 self.emb_rel_v = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)141 142 nn.init.xavier_uniform_(self.conv_q.weight)143 nn.init.xavier_uniform_(self.conv_k.weight)144 nn.init.xavier_uniform_(self.conv_v.weight)145 if proximal_init:146 with torch.no_grad():147 self.conv_k.weight.copy_(self.conv_q.weight)148 self.conv_k.bias.copy_(self.conv_q.bias)149 150 def forward(self, x, c, attn_mask=None):151 q = self.conv_q(x)152 k = self.conv_k(c)153 v = self.conv_v(c)154 155 x, self.attn = self.attention(q, k, v, mask=attn_mask)156 157 x = self.conv_o(x)158 return x159 160 def attention(self, query, key, value, mask=None):161 # reshape [b, d, t] -> [b, n_h, t, d_k]162 b, d, t_s, t_t = (*key.size(), query.size(2))163 query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)164 key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)165 value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)166 167 scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))168 if self.window_size is not None:169 assert t_s == t_t, "Relative attention is only available for self-attention."170 key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)171 rel_logits = self._matmul_with_relative_keys(query /math.sqrt(self.k_channels), key_relative_embeddings)172 scores_local = self._relative_position_to_absolute_position(rel_logits)173 scores = scores + scores_local174 if self.proximal_bias:175 assert t_s == t_t, "Proximal bias is only available for self-attention."176 scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype)177 if mask is not None:178 scores = scores.masked_fill(mask == 0, -1e4)179 if self.block_length is not None:180 assert t_s == t_t, "Local attention is only available for self-attention."181 block_mask = torch.ones_like(scores).triu(-self.block_length).tril(self.block_length)182 scores = scores.masked_fill(block_mask == 0, -1e4)183 p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s]184 p_attn = self.drop(p_attn)185 output = torch.matmul(p_attn, value)186 if self.window_size is not None:187 relative_weights = self._absolute_position_to_relative_position(p_attn)188 value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)189 output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings)190 output = output.transpose(2, 3).contiguous().view(b, d, t_t) # [b, n_h, t_t, d_k] -> [b, d, t_t]191 return output, p_attn192 193 def _matmul_with_relative_values(self, x, y):194 """195 x: [b, h, l, m]196 y: [h or 1, m, d]197 ret: [b, h, l, d]198 """199 ret = torch.matmul(x, y.unsqueeze(0))200 return ret201 202 def _matmul_with_relative_keys(self, x, y):203 """204 x: [b, h, l, d]205 y: [h or 1, m, d]206 ret: [b, h, l, m]207 """208 ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))209 return ret210 211 def _get_relative_embeddings(self, relative_embeddings, length):212 max_relative_position = 2 * self.window_size + 1213 # Pad first before slice to avoid using cond ops.214 pad_length = max(length - (self.window_size + 1), 0)215 slice_start_position = max((self.window_size + 1) - length, 0)216 slice_end_position = slice_start_position + 2 * length - 1217 if pad_length > 0:218 padded_relative_embeddings = F.pad(219 relative_embeddings,220 commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]))221 else:222 padded_relative_embeddings = relative_embeddings223 used_relative_embeddings = padded_relative_embeddings[:,slice_start_position:slice_end_position]224 return used_relative_embeddings225 226 def _relative_position_to_absolute_position(self, x):227 """228 x: [b, h, l, 2*l-1]229 ret: [b, h, l, l]230 """231 batch, heads, length, _ = x.size()232 # Concat columns of pad to shift from relative to absolute indexing.233 x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]]))234 235 # Concat extra elements so to add up to shape (len+1, 2*len-1).236 x_flat = x.view([batch, heads, length * 2 * length])237 x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [0, length - 1]]))238 239 # Reshape and slice out the padded elements.240 x_final = x_flat.view([batch, heads, length+1, 2*length-1])[:, :, :length, length-1:]241 return x_final242 243 def _absolute_position_to_relative_position(self, x):244 """245 x: [b, h, l, l]246 ret: [b, h, l, 2*l-1]247 """248 batch, heads, length, _ = x.size()249 # padd along column250 x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length - 1]]))251 x_flat = x.view([batch, heads, length**2 + length*(length -1)])252 # add 0's in the beginning that will skew the elements after reshape253 x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))254 x_final = x_flat.view([batch, heads, length, 2*length])[:,:,:,1:]255 return x_final256 257 def _attention_bias_proximal(self, length):258 """Bias for self-attention to encourage attention to close positions.259 Args:260 length: an integer scalar.261 Returns:262 a Tensor with shape [1, 1, length, length]263 """264 r = torch.arange(length, dtype=torch.float32)265 diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)266 return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)267 268 269class FFN(nn.Module):270 def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0., activation=None, causal=False):271 super().__init__()272 self.in_channels = in_channels273 self.out_channels = out_channels274 self.filter_channels = filter_channels275 self.kernel_size = kernel_size276 self.p_dropout = p_dropout277 self.activation = activation278 self.causal = causal279 280 if causal:281 self.padding = self._causal_padding282 else:283 self.padding = self._same_padding284 285 self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)286 self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)287 self.drop = nn.Dropout(p_dropout)288 289 def forward(self, x, x_mask):290 x = self.conv_1(self.padding(x * x_mask))291 if self.activation == "gelu":292 x = x * torch.sigmoid(1.702 * x)293 else:294 x = torch.relu(x)295 x = self.drop(x)296 x = self.conv_2(self.padding(x * x_mask))297 return x * x_mask298 299 def _causal_padding(self, x):300 if self.kernel_size == 1:301 return x302 pad_l = self.kernel_size - 1303 pad_r = 0304 padding = [[0, 0], [0, 0], [pad_l, pad_r]]305 x = F.pad(x, commons.convert_pad_shape(padding))306 return x307 308 def _same_padding(self, x):309 if self.kernel_size == 1:310 return x311 pad_l = (self.kernel_size - 1) // 2312 pad_r = self.kernel_size // 2313 padding = [[0, 0], [0, 0], [pad_l, pad_r]]314 x = F.pad(x, commons.convert_pad_shape(padding))315 return x316 317 318import torch.nn as nn319from torch.nn.utils import remove_weight_norm, weight_norm320 321 322class Depthwise_Separable_Conv1D(nn.Module):323 def __init__(324 self,325 in_channels,326 out_channels,327 kernel_size,328 stride=1,329 padding=0,330 dilation=1,331 bias=True,332 padding_mode='zeros', # TODO: refine this type333 device=None,334 dtype=None335 ):336 super().__init__()337 self.depth_conv = nn.Conv1d(in_channels=in_channels, out_channels=in_channels, kernel_size=kernel_size,338 groups=in_channels, stride=stride, padding=padding, dilation=dilation, bias=bias,339 padding_mode=padding_mode, device=device, dtype=dtype)340 self.point_conv = nn.Conv1d(in_channels=in_channels, out_channels=out_channels, kernel_size=1, bias=bias,341 device=device, dtype=dtype)342 343 def forward(self, input):344 return self.point_conv(self.depth_conv(input))345 346 def weight_norm(self):347 self.depth_conv = weight_norm(self.depth_conv, name='weight')348 self.point_conv = weight_norm(self.point_conv, name='weight')349 350 def remove_weight_norm(self):351 self.depth_conv = remove_weight_norm(self.depth_conv, name='weight')352 self.point_conv = remove_weight_norm(self.point_conv, name='weight')353 354 355class Depthwise_Separable_TransposeConv1D(nn.Module):356 def __init__(357 self,358 in_channels,359 out_channels,360 kernel_size,361 stride=1,362 padding=0,363 output_padding=0,364 bias=True,365 dilation=1,366 padding_mode='zeros', # TODO: refine this type367 device=None,368 dtype=None369 ):370 super().__init__()371 self.depth_conv = nn.ConvTranspose1d(in_channels=in_channels, out_channels=in_channels, kernel_size=kernel_size,372 groups=in_channels, stride=stride, output_padding=output_padding,373 padding=padding, dilation=dilation, bias=bias, padding_mode=padding_mode,374 device=device, dtype=dtype)375 self.point_conv = nn.Conv1d(in_channels=in_channels, out_channels=out_channels, kernel_size=1, bias=bias,376 device=device, dtype=dtype)377 378 def forward(self, input):379 return self.point_conv(self.depth_conv(input))380 381 def weight_norm(self):382 self.depth_conv = weight_norm(self.depth_conv, name='weight')383 self.point_conv = weight_norm(self.point_conv, name='weight')384 385 def remove_weight_norm(self):386 remove_weight_norm(self.depth_conv, name='weight')387 remove_weight_norm(self.point_conv, name='weight')388 389 390def weight_norm_modules(module, name='weight', dim=0):391 if isinstance(module, Depthwise_Separable_Conv1D) or isinstance(module, Depthwise_Separable_TransposeConv1D):392 module.weight_norm()393 return module394 else:395 return weight_norm(module, name, dim)396 397 398def remove_weight_norm_modules(module, name='weight'):399 if isinstance(module, Depthwise_Separable_Conv1D) or isinstance(module, Depthwise_Separable_TransposeConv1D):400 module.remove_weight_norm()401 else:402 remove_weight_norm(module, name)403 404 405class FFT(nn.Module):406 def __init__(self, hidden_channels, filter_channels, n_heads, n_layers=1, kernel_size=1, p_dropout=0.,407 proximal_bias=False, proximal_init=True, isflow = False, **kwargs):408 super().__init__()409 self.hidden_channels = hidden_channels410 self.filter_channels = filter_channels411 self.n_heads = n_heads412 self.n_layers = n_layers413 self.kernel_size = kernel_size414 self.p_dropout = p_dropout415 self.proximal_bias = proximal_bias416 self.proximal_init = proximal_init417 if isflow:418 cond_layer = torch.nn.Conv1d(kwargs["gin_channels"], 2*hidden_channels*n_layers, 1)419 self.cond_pre = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, 1)420 self.cond_layer = weight_norm_modules(cond_layer, name='weight')421 self.gin_channels = kwargs["gin_channels"]422 self.drop = nn.Dropout(p_dropout)423 self.self_attn_layers = nn.ModuleList()424 self.norm_layers_0 = nn.ModuleList()425 self.ffn_layers = nn.ModuleList()426 self.norm_layers_1 = nn.ModuleList()427 for i in range(self.n_layers):428 self.self_attn_layers.append(429 MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias,430 proximal_init=proximal_init))431 self.norm_layers_0.append(LayerNorm(hidden_channels))432 self.ffn_layers.append(433 FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))434 self.norm_layers_1.append(LayerNorm(hidden_channels))435 436 def forward(self, x, x_mask, g = None):437 """438 x: decoder input439 h: encoder output440 """441 if g is not None:442 g = self.cond_layer(g)443 444 self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(device=x.device, dtype=x.dtype)445 x = x * x_mask446 for i in range(self.n_layers):447 if g is not None:448 x = self.cond_pre(x)449 cond_offset = i * 2 * self.hidden_channels450 g_l = g[:,cond_offset:cond_offset+2*self.hidden_channels,:]451 x = commons.fused_add_tanh_sigmoid_multiply(452 x,453 g_l,454 torch.IntTensor([self.hidden_channels]))455 y = self.self_attn_layers[i](x, x, self_attn_mask)456 y = self.drop(y)457 x = self.norm_layers_0[i](x + y)458 459 y = self.ffn_layers[i](x, x_mask)460 y = self.drop(y)461 x = self.norm_layers_1[i](x + y)462 x = x * x_mask463 return x464 465 466 467class TransformerCouplingLayer(nn.Module):468 def __init__(self,469 channels,470 hidden_channels,471 kernel_size,472 n_layers,473 n_heads,474 p_dropout=0,475 filter_channels=0,476 mean_only=False,477 wn_sharing_parameter=None,478 gin_channels = 0479 ):480 assert channels % 2 == 0, "channels should be divisible by 2"481 super().__init__()482 self.channels = channels483 self.hidden_channels = hidden_channels484 self.kernel_size = kernel_size485 self.n_layers = n_layers486 self.half_channels = channels // 2487 self.mean_only = mean_only488 489 self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)490 self.enc = Encoder(hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, isflow = True, gin_channels = gin_channels) if wn_sharing_parameter is None else wn_sharing_parameter491 self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)492 self.post.weight.data.zero_()493 self.post.bias.data.zero_()494 495 def forward(self, x, x_mask, g=None, reverse=False):496 x0, x1 = torch.split(x, [self.half_channels]*2, 1)497 h = self.pre(x0) * x_mask498 h = self.enc(h, x_mask, g=g)499 stats = self.post(h) * x_mask500 if not self.mean_only:501 m, logs = torch.split(stats, [self.half_channels]*2, 1)502 else:503 m = stats504 logs = torch.zeros_like(m)505 506 if not reverse:507 x1 = m + x1 * torch.exp(logs) * x_mask508 x = torch.cat([x0, x1], 1)509 logdet = torch.sum(logs, [1,2])510 return x, logdet511 else:512 x1 = (x1 - m) * torch.exp(-logs) * x_mask513 x = torch.cat([x0, x1], 1)514 return x