Spaces:
Running on Zero
Running on Zero
File size: 17,191 Bytes
cc348e2 | 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 | from abc import ABC
import torch
from Layers.Conformer import Conformer
from Layers.DurationPredictor import DurationPredictor
from Layers.LengthRegulator import LengthRegulator
from Layers.PostNet import PostNet
from Layers.VariancePredictor import VariancePredictor
from Utility.utils import make_non_pad_mask
from Utility.utils import make_pad_mask
from Layers.Conformer_accent_mha_mean import Conformer_accent_mha_mean # Added
from Layers.Conformer_integrate_lang_emb import Conformer_integrate_lang_emb # Added
class FastSpeech2(torch.nn.Module, ABC):
def __init__(self, # network structure related
weights,
idim=60,
odim=80,
adim=384,
aheads=4,
elayers=6,
eunits=1536,
dlayers=6,
dunits=1536,
postnet_layers=5,
postnet_chans=256,
postnet_filts=5,
positionwise_conv_kernel_size=1,
use_scaled_pos_enc=True,
use_batch_norm=True,
encoder_normalize_before=True,
decoder_normalize_before=True,
encoder_concat_after=False,
decoder_concat_after=False,
reduction_factor=1,
# encoder / decoder
use_macaron_style_in_conformer=True,
use_cnn_in_conformer=True,
conformer_enc_kernel_size=7,
conformer_dec_kernel_size=31,
# duration predictor
duration_predictor_layers=2,
duration_predictor_chans=256,
duration_predictor_kernel_size=3,
# energy predictor
energy_predictor_layers=2,
energy_predictor_chans=256,
energy_predictor_kernel_size=3,
energy_predictor_dropout=0.5,
energy_embed_kernel_size=1,
energy_embed_dropout=0.0,
stop_gradient_from_energy_predictor=True,
# pitch predictor
pitch_predictor_layers=5,
pitch_predictor_chans=256,
pitch_predictor_kernel_size=5,
pitch_predictor_dropout=0.5,
pitch_embed_kernel_size=1,
pitch_embed_dropout=0.0,
stop_gradient_from_pitch_predictor=True,
# training related
transformer_enc_dropout_rate=0.2,
transformer_enc_positional_dropout_rate=0.2,
transformer_enc_attn_dropout_rate=0.2,
transformer_dec_dropout_rate=0.2,
transformer_dec_positional_dropout_rate=0.2,
transformer_dec_attn_dropout_rate=0.2,
duration_predictor_dropout_rate=0.2,
postnet_dropout_rate=0.5,
# additional features
utt_embed_dim=704,
connect_utt_emb_at_encoder_out=True,
lang_embs=100):
super().__init__()
self.idim = idim
self.odim = odim
self.reduction_factor = reduction_factor
self.stop_gradient_from_pitch_predictor = stop_gradient_from_pitch_predictor
self.stop_gradient_from_energy_predictor = stop_gradient_from_energy_predictor
self.use_scaled_pos_enc = use_scaled_pos_enc
self.multilingual_model = lang_embs is not None
self.multispeaker_model = utt_embed_dim is not None
embed = torch.nn.Sequential(torch.nn.Linear(idim, 100),
torch.nn.Tanh(),
torch.nn.Linear(100, adim))
self.encoder = Conformer_integrate_lang_emb(idim=idim, attention_dim=adim, attention_heads=aheads, linear_units=eunits, num_blocks=elayers,
#self.encoder = Conformer(idim=idim, attention_dim=adim, attention_heads=aheads, linear_units=eunits, num_blocks=elayers,
input_layer=embed, dropout_rate=transformer_enc_dropout_rate,
positional_dropout_rate=transformer_enc_positional_dropout_rate, attention_dropout_rate=transformer_enc_attn_dropout_rate,
normalize_before=encoder_normalize_before, concat_after=encoder_concat_after,
positionwise_conv_kernel_size=positionwise_conv_kernel_size, macaron_style=use_macaron_style_in_conformer,
use_cnn_module=use_cnn_in_conformer, cnn_module_kernel=conformer_enc_kernel_size, zero_triu=False,
utt_embed=utt_embed_dim, connect_utt_emb_at_encoder_out=connect_utt_emb_at_encoder_out)
self.duration_predictor = DurationPredictor(idim=adim, n_layers=duration_predictor_layers,
n_chans=duration_predictor_chans,
kernel_size=duration_predictor_kernel_size,
dropout_rate=duration_predictor_dropout_rate, )
self.pitch_predictor = VariancePredictor(idim=adim, n_layers=pitch_predictor_layers,
n_chans=pitch_predictor_chans,
kernel_size=pitch_predictor_kernel_size,
dropout_rate=pitch_predictor_dropout)
self.pitch_embed = torch.nn.Sequential(torch.nn.Conv1d(in_channels=1, out_channels=adim,
kernel_size=pitch_embed_kernel_size,
padding=(pitch_embed_kernel_size - 1) // 2),
torch.nn.Dropout(pitch_embed_dropout))
self.energy_predictor = VariancePredictor(idim=adim, n_layers=energy_predictor_layers,
n_chans=energy_predictor_chans,
kernel_size=energy_predictor_kernel_size,
dropout_rate=energy_predictor_dropout)
self.energy_embed = torch.nn.Sequential(torch.nn.Conv1d(in_channels=1, out_channels=adim,
kernel_size=energy_embed_kernel_size,
padding=(energy_embed_kernel_size - 1) // 2),
torch.nn.Dropout(energy_embed_dropout))
self.length_regulator = LengthRegulator()
self.decoder = Conformer_integrate_lang_emb(idim=0,
# self.decoder = Conformer(idim=0,
attention_dim=adim,
attention_heads=aheads,
linear_units=dunits,
num_blocks=dlayers,
input_layer=None,
dropout_rate=transformer_dec_dropout_rate,
positional_dropout_rate=transformer_dec_positional_dropout_rate,
attention_dropout_rate=transformer_dec_attn_dropout_rate,
normalize_before=decoder_normalize_before,
concat_after=decoder_concat_after,
positionwise_conv_kernel_size=positionwise_conv_kernel_size,
macaron_style=use_macaron_style_in_conformer,
use_cnn_module=use_cnn_in_conformer,
cnn_module_kernel=conformer_dec_kernel_size)
self.feat_out = torch.nn.Linear(adim, odim * reduction_factor)
self.postnet = PostNet(idim=idim,
odim=odim,
n_layers=postnet_layers,
n_chans=postnet_chans,
n_filts=postnet_filts,
use_batch_norm=use_batch_norm,
dropout_rate=postnet_dropout_rate)
self.load_state_dict(weights)
def _forward(self, text_tensors, text_lens, gold_speech=None, speech_lens=None,
gold_durations=None, gold_pitch=None, gold_energy=None,
is_inference=False, duration_scaling_factor=1.0, utterance_embedding=None, lang_embs=None,
pitch_variance_scale=1.0, energy_variance_scale=1.0):
# if not self.multilingual_model:
# lang_ids = None
if not self.multispeaker_model:
utterance_embedding = None
# forward encoder
text_masks = self._source_mask(text_lens)
encoded_texts, _ = self.encoder(text_tensors, text_masks, utterance_embedding=utterance_embedding, lang_embs=lang_embs) # (B, Tmax, adim)
# forward duration predictor and variance predictors
duration_masks = make_pad_mask(text_lens, device=text_lens.device)
if self.stop_gradient_from_pitch_predictor:
pitch_predictions = self.pitch_predictor(encoded_texts.detach(), duration_masks.unsqueeze(-1))
else:
pitch_predictions = self.pitch_predictor(encoded_texts, duration_masks.unsqueeze(-1))
if self.stop_gradient_from_energy_predictor:
energy_predictions = self.energy_predictor(encoded_texts.detach(), duration_masks.unsqueeze(-1))
else:
energy_predictions = self.energy_predictor(encoded_texts, duration_masks.unsqueeze(-1))
if is_inference:
if gold_durations is not None:
duration_predictions = gold_durations
else:
duration_predictions = self.duration_predictor.inference(encoded_texts, duration_masks)
if gold_pitch is not None:
pitch_predictions = gold_pitch
if gold_energy is not None:
energy_predictions = gold_energy
pitch_predictions = _scale_variance(pitch_predictions, pitch_variance_scale)
energy_predictions = _scale_variance(energy_predictions, energy_variance_scale)
pitch_embeddings = self.pitch_embed(pitch_predictions.transpose(1, 2)).transpose(1, 2)
energy_embeddings = self.energy_embed(energy_predictions.transpose(1, 2)).transpose(1, 2)
encoded_texts = encoded_texts + energy_embeddings + pitch_embeddings
encoded_texts = self.length_regulator(encoded_texts, duration_predictions, duration_scaling_factor)
else:
duration_predictions = self.duration_predictor(encoded_texts, duration_masks)
pitch_predictions = _scale_variance(pitch_predictions, pitch_variance_scale)
energy_predictions = _scale_variance(energy_predictions, energy_variance_scale)
# use groundtruth to clone
pitch_embeddings = self.pitch_embed(gold_pitch.transpose(1, 2)).transpose(1, 2)
energy_embeddings = self.energy_embed(gold_energy.transpose(1, 2)).transpose(1, 2)
encoded_texts = encoded_texts + energy_embeddings + pitch_embeddings
encoded_texts = self.length_regulator(encoded_texts, gold_durations) # (B, Lmax, adim)
# forward decoder
if speech_lens is not None and not is_inference:
if self.reduction_factor > 1:
olens_in = speech_lens.new([olen // self.reduction_factor for olen in speech_lens])
else:
olens_in = speech_lens
h_masks = self._source_mask(olens_in)
else:
h_masks = None
zs, _ = self.decoder(encoded_texts, h_masks) # (B, Lmax, adim)
before_outs = self.feat_out(zs).view(zs.size(0), -1, self.odim) # (B, Lmax, odim)
# postnet -> (B, Lmax//r * r, odim)
after_outs = before_outs + self.postnet(before_outs.transpose(1, 2)).transpose(1, 2)
return before_outs, after_outs, duration_predictions, pitch_predictions, energy_predictions
@torch.no_grad()
def forward(self,
text,
speech=None,
durations=None,
pitch=None,
energy=None,
utterance_embedding=None,
return_duration_pitch_energy=False,
lang_emb=None,
duration_scaling_factor=1.0,
pitch_variance_scale=1.0,
energy_variance_scale=1.0):
"""
Generate the sequence of spectrogram frames given the sequence of vectorized phonemes.
Args:
text: input sequence of vectorized phonemes
speech: feature sequence to extract style from (not used for now, placeholder for future plans)
durations: durations to be used (optional, if not provided, they will be predicted)
pitch: token-averaged pitch curve to be used (optional, if not provided, it will be predicted)
energy: token-averaged energy curve to be used (optional, if not provided, it will be predicted)
return_duration_pitch_energy: whether to return the list of predicted durations for nicer plotting
utterance_embedding: embedding of speaker information
lang_id: id to be fed into the embedding layer that contains language information
duration_scaling_factor: reasonable values are 0.8 < scale < 1.2.
1.0 means no scaling happens, higher values increase durations for the whole
utterance, lower values decrease durations for the whole utterance.
pitch_variance_scale: reasonable values are 0.6 < scale < 1.4.
1.0 means no scaling happens, higher values increase variance of the pitch curve,
lower values decrease variance of the pitch curve.
energy_variance_scale: reasonable values are 0.6 < scale < 1.4.
1.0 means no scaling happens, higher values increase variance of the energy curve,
lower values decrease variance of the energy curve.
Returns:
mel spectrogram
"""
self.eval()
# setup batch axis
ilens = torch.tensor([text.shape[0]], dtype=torch.long, device=text.device)
if speech is not None:
gold_speech = speech.unsqueeze(0).to(text.device)
else:
gold_speech = None
if durations is not None:
durations = durations.unsqueeze(0).to(text.device)
if pitch is not None:
pitch = pitch.unsqueeze(0).to(text.device)
if energy is not None:
energy = energy.unsqueeze(0).to(text.device)
# if lang_id is not None:
# lang_id = lang_id.unsqueeze(0).to(text.device)
before_outs, after_outs, d_outs, pitch_predictions, energy_predictions = self._forward(text.unsqueeze(0),
ilens,
gold_speech=gold_speech,
gold_durations=durations,
is_inference=True,
gold_pitch=pitch,
gold_energy=energy,
utterance_embedding=utterance_embedding.unsqueeze(0),
lang_embs=lang_emb,
duration_scaling_factor=duration_scaling_factor,
pitch_variance_scale=pitch_variance_scale,
energy_variance_scale=energy_variance_scale)
self.train()
if return_duration_pitch_energy:
return after_outs[0], d_outs[0], pitch_predictions[0], energy_predictions[0]
return after_outs[0]
def _source_mask(self, ilens):
x_masks = make_non_pad_mask(ilens).to(next(self.parameters()).device)
return x_masks.unsqueeze(-2)
def _scale_variance(sequence, scale):
if scale == 1.0:
return sequence
average = sequence[0][sequence[0] != 0.0].mean()
sequence = sequence - average # center sequence around 0
sequence = sequence * scale # scale the variance
sequence = sequence + average # move center back to original with changed variance
return sequence
|