Spaces:
Running
Running
File size: 12,932 Bytes
2c7b92a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 |
import math
import torch
import torch.nn as nn
from torch.nn.modules.utils import consume_prefix_in_state_dict_if_present
class CustomLSTM(nn.Module):
def __init__(self, input_sz, hidden_sz):
super().__init__()
self.input_sz = input_sz
self.hidden_size = hidden_sz
self.W = nn.Parameter(torch.Tensor(input_sz, hidden_sz * 4))
self.U = nn.Parameter(torch.Tensor(hidden_sz, hidden_sz * 4))
self.bias = nn.Parameter(torch.Tensor(hidden_sz * 4))
self.init_weights()
def init_weights(self):
stdv = 1.0 / math.sqrt(self.hidden_size)
for weight in self.parameters():
weight.data.uniform_(-stdv, stdv)
def forward(self, x,
init_states=None):
"""Assumes x is of shape (batch, sequence, feature)"""
#print(type(x))
#print(x.shape)
bs, seq_sz, _ = x.size()
hidden_seq = []
if init_states is None:
h_t, c_t = (torch.zeros(bs, self.hidden_size).to(x.device),
torch.zeros(bs, self.hidden_size).to(x.device))
else:
h_t, c_t = init_states
HS = self.hidden_size
for t in range(seq_sz):
x_t = x[:, t, :]
# batch the computations into a single matrix multiplication
gates = x_t @ self.W + h_t @ self.U + self.bias
i_t, f_t, g_t, o_t = (
torch.sigmoid(gates[:, :HS]), # input
torch.sigmoid(gates[:, HS:HS*2]), # forget
torch.tanh(gates[:, HS*2:HS*3]),
torch.sigmoid(gates[:, HS*3:]), # output
)
c_t = f_t * c_t + i_t * g_t
h_t = o_t * torch.tanh(c_t)
hidden_seq.append(h_t.unsqueeze(0))
hidden_seq = torch.cat(hidden_seq, dim=0)
# reshape from shape (sequence, batch, feature) to (batch, sequence, feature)
hidden_seq = hidden_seq.transpose(0, 1).contiguous()
return hidden_seq, (h_t, c_t)
hparams = {
'n_mel_channels': 128, # From LogMelSpectrogram
'postnet_embedding_dim': 512, # Common choice, adjust as needed
'postnet_kernel_size': 5, # Common choice, adjust as needed
'postnet_n_convolutions': 5, # Typical number of Postnet convolutions
}
class ConvNorm(torch.nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=1, stride=1,
padding=None, dilation=1, bias=True, w_init_gain='linear'):
super(ConvNorm, self).__init__()
if padding is None:
assert(kernel_size % 2 == 1)
padding = int(dilation * (kernel_size - 1) / 2)
self.conv = torch.nn.Conv1d(in_channels, out_channels,
kernel_size=kernel_size, stride=stride,
padding=padding, dilation=dilation,
bias=bias)
torch.nn.init.xavier_uniform_(
self.conv.weight, gain=torch.nn.init.calculate_gain(w_init_gain))
def forward(self, signal):
conv_signal = self.conv(signal)
return conv_signal
URLS = {
"hubert-discrete": "https://github.com/bshall/acoustic-model/releases/download/v0.1/hubert-discrete-d49e1c77.pt",
"hubert-soft": "https://github.com/bshall/acoustic-model/releases/download/v0.1/hubert-soft-0321fd7e.pt",
}
class AcousticModel(nn.Module):
def __init__(self, discrete: bool = False, upsample: bool = True, use_custom_lstm=False):
super().__init__()
# self.spk_projection = nn.Linear(512+512, 512)
self.encoder = Encoder(discrete, upsample)
self.decoder = Decoder(use_custom_lstm=use_custom_lstm)
self.postnet = Postnet(hparams) # Add this line. Ensure hparams is defined or pass explicit parameters
def forward(self, x: torch.Tensor, spk_embs, mels: torch.Tensor) -> torch.Tensor:
x = self.encoder(x)
exp_spk_embs = spk_embs.unsqueeze(1).expand(-1, x.size(1), -1)
concat_x = torch.cat([x, exp_spk_embs], dim=-1)
# x = self.spk_projection(concat_x)
output = self.decoder(concat_x, mels)
postnet_output = self.postnet(output) + output
return postnet_output
#def forward(self, x: torch.Tensor, mels: torch.Tensor) -> torch.Tensor:
# x = self.encoder(x)
# return self.decoder(x, mels)
def forward_test(self, x, spk_embs, mels):
print('x shape', x.shape)
print('se shape', spk_embs.shape)
print('mels shape', mels.shape)
x = self.encoder(x)
print('x_enc shape', x.shape)
return
@torch.inference_mode()
def generate(self, x: torch.Tensor, spk_embs) -> torch.Tensor:
x = self.encoder(x)
exp_spk_embs = spk_embs.unsqueeze(1).expand(-1, x.size(1), -1)
concat_x = torch.cat([x, exp_spk_embs], dim=-1)
# x = self.spk_projection(concat_x)
mels = self.decoder.generate(concat_x)
postnet_mels = self.postnet(mels) + mels
return postnet_mels
class Encoder(nn.Module):
def __init__(self, discrete: bool = False, upsample: bool = True):
super().__init__()
self.embedding = nn.Embedding(100 + 1, 256) if discrete else None
self.prenet = PreNet(256, 256, 256)
self.convs = nn.Sequential(
nn.Conv1d(256, 512, 5, 1, 2),
nn.ReLU(),
nn.Dropout(0.3),
nn.InstanceNorm1d(512),
nn.ConvTranspose1d(512, 512, 4, 2, 1) if upsample else nn.Identity(),
nn.Dropout(0.3),
nn.Conv1d(512, 512, 5, 1, 2),
nn.ReLU(),
nn.Dropout(0.3),
nn.InstanceNorm1d(512),
nn.Conv1d(512, 512, 5, 1, 2),
nn.ReLU(),
nn.Dropout(0.3),
nn.InstanceNorm1d(512),
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
if self.embedding is not None:
x = self.embedding(x)
x = self.prenet(x)
x = self.convs(x.transpose(1, 2))
return x.transpose(1, 2)
class Decoder(nn.Module):
def __init__(self, use_custom_lstm=False):
super().__init__()
self.use_custom_lstm = use_custom_lstm
self.prenet = PreNet(128, 256, 256)
if use_custom_lstm:
self.lstm1 = CustomLSTM(1024 + 256, 1024)
self.lstm2 = CustomLSTM(1024, 1024)
self.lstm3 = CustomLSTM(1024, 1024)
else:
self.lstm1 = nn.LSTM(1024 + 256, 1024)
self.lstm2 = nn.LSTM(1024, 1024)
self.lstm3 = nn.LSTM(1024, 1024)
self.proj = nn.Linear(1024, 128, bias=False)
self.dropout = nn.Dropout(0.3)
def forward(self, x: torch.Tensor, mels: torch.Tensor) -> torch.Tensor:
mels = self.prenet(mels)
x, _ = self.lstm1(torch.cat((x, mels), dim=-1))
x = self.dropout(x)
res = x
x, _ = self.lstm2(x)
x = self.dropout(x)
x = res + x
res = x
x, _ = self.lstm3(x)
x = self.dropout(x)
x = res + x
return self.proj(x)
@torch.inference_mode()
def generate(self, xs: torch.Tensor) -> torch.Tensor:
m = torch.zeros(xs.size(0), 128, device=xs.device)
if self.use_custom_lstm:
h1 = torch.zeros(xs.size(0), 1024, device=xs.device)
c1 = torch.zeros(xs.size(0), 1024, device=xs.device)
h2 = torch.zeros(xs.size(0), 1024, device=xs.device)
c2 = torch.zeros(xs.size(0), 1024, device=xs.device)
h3 = torch.zeros(xs.size(0), 1024, device=xs.device)
c3 = torch.zeros(xs.size(0), 1024, device=xs.device)
else:
h1 = torch.zeros(1, xs.size(0), 1024, device=xs.device)
c1 = torch.zeros(1, xs.size(0), 1024, device=xs.device)
h2 = torch.zeros(1, xs.size(0), 1024, device=xs.device)
c2 = torch.zeros(1, xs.size(0), 1024, device=xs.device)
h3 = torch.zeros(1, xs.size(0), 1024, device=xs.device)
c3 = torch.zeros(1, xs.size(0), 1024, device=xs.device)
mel = []
for x in torch.unbind(xs, dim=1):
m = self.prenet(m)
x = torch.cat((x, m), dim=1).unsqueeze(1)
x1, (h1, c1) = self.lstm1(x, (h1, c1))
x2, (h2, c2) = self.lstm2(x1, (h2, c2))
x = x1 + x2
x3, (h3, c3) = self.lstm3(x, (h3, c3))
x = x + x3
m = self.proj(x).squeeze(1)
mel.append(m)
return torch.stack(mel, dim=1)
class PreNet(nn.Module):
def __init__(
self,
input_size: int,
hidden_size: int,
output_size: int,
dropout: float = 0.5,
):
super().__init__()
self.net = nn.Sequential(
nn.Linear(input_size, hidden_size),
nn.ReLU(),
nn.Dropout(dropout),
nn.Linear(hidden_size, output_size),
nn.ReLU(),
nn.Dropout(dropout),
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.net(x)
def _acoustic(
name: str,
discrete: bool,
upsample: bool,
pretrained: bool = True,
progress: bool = True,
) -> AcousticModel:
acoustic = AcousticModel(discrete, upsample)
if pretrained:
checkpoint = torch.hub.load_state_dict_from_url(URLS[name], progress=progress)
consume_prefix_in_state_dict_if_present(checkpoint["acoustic-model"], "module.")
acoustic.load_state_dict(checkpoint["acoustic-model"])
acoustic.eval()
return acoustic
def hubert_discrete(
pretrained: bool = True,
progress: bool = True,
) -> AcousticModel:
r"""HuBERT-Discrete acoustic model from `"A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion"`.
Args:
pretrained (bool): load pretrained weights into the model
progress (bool): show progress bar when downloading model
"""
return _acoustic(
"hubert-discrete",
discrete=True,
upsample=True,
pretrained=pretrained,
progress=progress,
)
def hubert_soft(
pretrained: bool = True,
progress: bool = True,
) -> AcousticModel:
r"""HuBERT-Soft acoustic model from `"A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion"`.
Args:
pretrained (bool): load pretrained weights into the model
progress (bool): show progress bar when downloading model
"""
return _acoustic(
"hubert-soft",
discrete=False,
upsample=True,
pretrained=pretrained,
progress=progress,
)
class Postnet(nn.Module):
def __init__(self, hparams):
super(Postnet, self).__init__()
self.convolutions = nn.ModuleList()
self.convolutions.append(
nn.Sequential(
ConvNorm(in_channels=hparams['n_mel_channels'], # Adjusted input channels
out_channels=hparams['postnet_embedding_dim'], # Output channels remain the same
kernel_size=hparams['postnet_kernel_size'], stride=1,
padding=int((hparams['postnet_kernel_size'] - 1) / 2), # Dynamic padding
dilation=1, bias=True, w_init_gain='tanh'),
nn.BatchNorm1d(hparams['postnet_embedding_dim'])
)
)
for i in range(1, hparams['postnet_n_convolutions'] - 1):
self.convolutions.append(
nn.Sequential(
ConvNorm(hparams['postnet_embedding_dim'],
hparams['postnet_embedding_dim'],
kernel_size=hparams['postnet_kernel_size'], stride=1,
padding=int((hparams['postnet_kernel_size'] - 1) / 2), # Dynamic padding
dilation=1, w_init_gain='tanh'),
nn.BatchNorm1d(hparams['postnet_embedding_dim'])
)
)
self.convolutions.append(
nn.Sequential(
ConvNorm(hparams['postnet_embedding_dim'], hparams['n_mel_channels'],
kernel_size=hparams['postnet_kernel_size'], stride=1,
padding=int((hparams['postnet_kernel_size'] - 1) / 2), # Dynamic padding
dilation=1, w_init_gain='linear'),
nn.BatchNorm1d(hparams['n_mel_channels'])
)
)
def forward(self, x):
#print(f"Input shape to Postnet: {x.shape}")
x = x.transpose(1, 2)
for i, conv in enumerate(self.convolutions[:-1]):
x = conv(x)
#print(f"Shape after Convolution {i+1}: {x.shape}")
x = torch.tanh(x)
x = F.dropout(x, 0.5, self.training)
# Last layer
x = self.convolutions[-1](x)
#print(f"Shape after last Convolution: {x.shape}")
x = F.dropout(x, 0.5, self.training)
x = x.transpose(1, 2)
return x |