Spaces:
Running
Running
File size: 16,299 Bytes
6a62ffb |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 |
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import numpy as np
import torch
from fairseq.data.audio.speech_to_text_dataset import S2TDataConfig
class SpeechGenerator(object):
def __init__(self, model, vocoder, data_cfg: S2TDataConfig):
self.model = model
self.vocoder = vocoder
stats_npz_path = data_cfg.global_cmvn_stats_npz
self.gcmvn_stats = None
if stats_npz_path is not None:
self.gcmvn_stats = np.load(stats_npz_path)
def gcmvn_denormalize(self, x):
# x: B x T x C
if self.gcmvn_stats is None:
return x
mean = torch.from_numpy(self.gcmvn_stats["mean"]).to(x)
std = torch.from_numpy(self.gcmvn_stats["std"]).to(x)
assert len(x.shape) == 3 and mean.shape[0] == std.shape[0] == x.shape[2]
x = x * std.view(1, 1, -1).expand_as(x)
return x + mean.view(1, 1, -1).expand_as(x)
def get_waveform(self, feat):
# T x C -> T
return None if self.vocoder is None else self.vocoder(feat).squeeze(0)
class AutoRegressiveSpeechGenerator(SpeechGenerator):
def __init__(
self,
model,
vocoder,
data_cfg,
max_iter: int = 6000,
eos_prob_threshold: float = 0.5,
):
super().__init__(model, vocoder, data_cfg)
self.max_iter = max_iter
self.eos_prob_threshold = eos_prob_threshold
@torch.no_grad()
def generate(self, model, sample, has_targ=False, **kwargs):
model.eval()
src_tokens = sample["net_input"]["src_tokens"]
src_lengths = sample["net_input"]["src_lengths"]
bsz, src_len = src_tokens.size()[:2]
n_frames_per_step = model.decoder.n_frames_per_step
out_dim = model.decoder.out_dim
raw_dim = out_dim // n_frames_per_step
# initialize
encoder_out = model.forward_encoder(
src_tokens, src_lengths, speaker=sample["speaker"]
)
incremental_state = {}
feat, attn, eos_prob = [], [], []
finished = src_tokens.new_zeros((bsz,)).bool()
out_lens = src_lengths.new_zeros((bsz,)).long().fill_(self.max_iter)
prev_feat_out = encoder_out["encoder_out"][0].new_zeros(bsz, 1, out_dim)
for step in range(self.max_iter):
cur_out_lens = out_lens.clone()
cur_out_lens.masked_fill_(cur_out_lens.eq(self.max_iter), step + 1)
_, cur_eos_out, cur_extra = model.forward_decoder(
prev_feat_out,
encoder_out=encoder_out,
incremental_state=incremental_state,
target_lengths=cur_out_lens,
speaker=sample["speaker"],
**kwargs,
)
cur_eos_prob = torch.sigmoid(cur_eos_out).squeeze(2)
feat.append(cur_extra["feature_out"])
attn.append(cur_extra["attn"])
eos_prob.append(cur_eos_prob)
cur_finished = cur_eos_prob.squeeze(1) > self.eos_prob_threshold
out_lens.masked_fill_((~finished) & cur_finished, step + 1)
finished = finished | cur_finished
if finished.sum().item() == bsz:
break
prev_feat_out = cur_extra["feature_out"]
feat = torch.cat(feat, dim=1)
feat = model.decoder.postnet(feat) + feat
eos_prob = torch.cat(eos_prob, dim=1)
attn = torch.cat(attn, dim=2)
alignment = attn.max(dim=1)[1]
feat = feat.reshape(bsz, -1, raw_dim)
feat = self.gcmvn_denormalize(feat)
eos_prob = eos_prob.repeat_interleave(n_frames_per_step, dim=1)
attn = attn.repeat_interleave(n_frames_per_step, dim=2)
alignment = alignment.repeat_interleave(n_frames_per_step, dim=1)
out_lens = out_lens * n_frames_per_step
finalized = [
{
"feature": feat[b, :out_len],
"eos_prob": eos_prob[b, :out_len],
"attn": attn[b, :, :out_len],
"alignment": alignment[b, :out_len],
"waveform": self.get_waveform(feat[b, :out_len]),
}
for b, out_len in zip(range(bsz), out_lens)
]
if has_targ:
assert sample["target"].size(-1) == out_dim
tgt_feats = sample["target"].view(bsz, -1, raw_dim)
tgt_feats = self.gcmvn_denormalize(tgt_feats)
tgt_lens = sample["target_lengths"] * n_frames_per_step
for b, (f, l) in enumerate(zip(tgt_feats, tgt_lens)):
finalized[b]["targ_feature"] = f[:l]
finalized[b]["targ_waveform"] = self.get_waveform(f[:l])
return finalized
class MultiDecoderSpeechGenerator(SpeechGenerator):
def __init__(
self,
models,
args,
vocoder,
data_cfg,
tgt_dict_mt,
max_iter: int = 6000,
eos_prob_threshold: float = 0.5,
eos_mt=None,
symbols_to_strip_from_output=None,
):
super().__init__(models[0], vocoder, data_cfg)
self.max_iter = max_iter
self.eos_prob_threshold = eos_prob_threshold
self.tgt_dict_mt = tgt_dict_mt
self.eos_mt = eos_mt
from examples.speech_to_speech.unity.sequence_generator import SequenceGenerator
from fairseq import search
self.text_generator = SequenceGenerator(
models,
tgt_dict_mt,
beam_size=max(1, getattr(args, "beam", 5)),
max_len_a=getattr(args, "max_len_a", 0),
max_len_b=getattr(args, "max_len_b", 200),
min_len=getattr(args, "min_len", 1),
normalize_scores=(not getattr(args, "unnormalized", False)),
len_penalty=getattr(args, "lenpen", 1),
unk_penalty=getattr(args, "unkpen", 0),
temperature=getattr(args, "temperature", 1.0),
match_source_len=getattr(args, "match_source_len", False),
no_repeat_ngram_size=getattr(args, "no_repeat_ngram_size", 0),
search_strategy=search.BeamSearch(tgt_dict_mt),
eos=eos_mt,
symbols_to_strip_from_output=symbols_to_strip_from_output,
)
@torch.no_grad()
def generate(self, model, sample, has_targ=False, **kwargs):
model.eval()
src_tokens = sample["net_input"]["src_tokens"]
src_lengths = sample["net_input"]["src_lengths"]
bsz, src_len = src_tokens.size()[:2]
n_frames_per_step = model.decoder.n_frames_per_step
out_dim = model.decoder.out_dim
raw_dim = out_dim // n_frames_per_step
# initialize
encoder_out = model.forward_encoder(
src_tokens, src_lengths, speaker=sample["speaker"]
)
prefix_tokens = None
constraints = None
bos_token = None
mt_decoder = getattr(model, f"{model.mt_task_name}_decoder")
# 1. MT decoder
finalized_mt = self.text_generator.generate_decoder(
[encoder_out],
src_tokens,
src_lengths,
sample,
prefix_tokens,
constraints,
bos_token,
aux_task_name=model.mt_task_name,
)
# extract decoder output corresponding to the best hypothesis
max_tgt_len = max([len(hypo[0]["tokens"]) for hypo in finalized_mt])
prev_output_tokens_mt = (
src_tokens.new_zeros(src_tokens.shape[0], max_tgt_len)
.fill_(mt_decoder.padding_idx)
.int()
) # B x T
for i, hypo in enumerate(finalized_mt):
i_beam = 0
tmp = hypo[i_beam]["tokens"].int() # hyp + eos
prev_output_tokens_mt[i, 0] = self.text_generator.eos
if tmp[-1] == self.text_generator.eos:
tmp = tmp[:-1]
prev_output_tokens_mt[i, 1 : len(tmp) + 1] = tmp
text = "".join([self.tgt_dict_mt[c] for c in tmp])
text = text.replace("_", " ")
text = text.replace("▁", " ")
text = text.replace("<unk>", " ")
text = text.replace("<s>", "")
text = text.replace("</s>", "")
if len(text) > 0 and text[0] == " ":
text = text[1:]
sample_id = sample["id"].tolist()[i]
print("{} (None-{})".format(text, sample_id))
mt_decoder_out = mt_decoder(
prev_output_tokens_mt,
encoder_out=encoder_out,
features_only=True,
)
x = mt_decoder_out[0].transpose(0, 1)
mt_decoder_padding_mask = None
if prev_output_tokens_mt.eq(mt_decoder.padding_idx).any():
mt_decoder_padding_mask = prev_output_tokens_mt.eq(mt_decoder.padding_idx)
# 2. TTS encoder
if getattr(model, "synthesizer_encoder", None) is not None:
synthesizer_encoder_out = model.synthesizer_encoder(
x,
mt_decoder_padding_mask,
)
else:
synthesizer_encoder_out = {
"encoder_out": [x], # T x B x C
"encoder_padding_mask": [mt_decoder_padding_mask]
if mt_decoder_padding_mask is not None
else [], # B x T
"encoder_embedding": [],
"encoder_states": [],
"src_tokens": [],
"src_lengths": [],
}
# 3. TTS decoder
incremental_state = {}
feat, attn, eos_prob = [], [], []
finished = src_tokens.new_zeros((bsz,)).bool()
out_lens = src_lengths.new_zeros((bsz,)).long().fill_(self.max_iter)
prev_feat_out = encoder_out["encoder_out"][0].new_zeros(bsz, 1, out_dim)
for step in range(self.max_iter):
cur_out_lens = out_lens.clone()
cur_out_lens.masked_fill_(cur_out_lens.eq(self.max_iter), step + 1)
_, cur_eos_out, cur_extra = model.forward_decoder(
prev_feat_out,
encoder_out=synthesizer_encoder_out,
incremental_state=incremental_state,
target_lengths=cur_out_lens,
speaker=sample["speaker"],
**kwargs,
)
cur_eos_prob = torch.sigmoid(cur_eos_out).squeeze(2)
feat.append(cur_extra["feature_out"])
attn.append(cur_extra["attn"])
eos_prob.append(cur_eos_prob)
cur_finished = cur_eos_prob.squeeze(1) > self.eos_prob_threshold
out_lens.masked_fill_((~finished) & cur_finished, step + 1)
finished = finished | cur_finished
if finished.sum().item() == bsz:
break
prev_feat_out = cur_extra["feature_out"]
feat = torch.cat(feat, dim=1)
feat = model.decoder.postnet(feat) + feat
eos_prob = torch.cat(eos_prob, dim=1)
attn = torch.cat(attn, dim=2)
alignment = attn.max(dim=1)[1]
feat = feat.reshape(bsz, -1, raw_dim)
feat = self.gcmvn_denormalize(feat)
eos_prob = eos_prob.repeat_interleave(n_frames_per_step, dim=1)
attn = attn.repeat_interleave(n_frames_per_step, dim=2)
alignment = alignment.repeat_interleave(n_frames_per_step, dim=1)
out_lens = out_lens * n_frames_per_step
finalized = [
{
"feature": feat[b, :out_len],
"eos_prob": eos_prob[b, :out_len],
"attn": attn[b, :, :out_len],
"alignment": alignment[b, :out_len],
"waveform": self.get_waveform(feat[b, :out_len]),
}
for b, out_len in zip(range(bsz), out_lens)
]
if has_targ:
assert sample["target"].size(-1) == out_dim
tgt_feats = sample["target"].view(bsz, -1, raw_dim)
tgt_feats = self.gcmvn_denormalize(tgt_feats)
tgt_lens = sample["target_lengths"] * n_frames_per_step
for b, (f, l) in enumerate(zip(tgt_feats, tgt_lens)):
finalized[b]["targ_feature"] = f[:l]
finalized[b]["targ_waveform"] = self.get_waveform(f[:l])
return finalized
class NonAutoregressiveSpeechGenerator(SpeechGenerator):
@torch.no_grad()
def generate(self, model, sample, has_targ=False, **kwargs):
model.eval()
bsz, max_src_len = sample["net_input"]["src_tokens"].size()
n_frames_per_step = model.encoder.n_frames_per_step
out_dim = model.encoder.out_dim
raw_dim = out_dim // n_frames_per_step
feat, feat_post, out_lens, log_dur_out, _, _ = model(
src_tokens=sample["net_input"]["src_tokens"],
src_lengths=sample["net_input"]["src_lengths"],
prev_output_tokens=sample["net_input"]["prev_output_tokens"],
incremental_state=None,
target_lengths=sample["target_lengths"],
speaker=sample["speaker"],
)
if feat_post is not None:
feat = feat_post
feat = feat.view(bsz, -1, raw_dim)
feat = self.gcmvn_denormalize(feat)
dur_out = torch.clamp(torch.round(torch.exp(log_dur_out) - 1).long(), min=0)
def get_dur_plot_data(d):
r = []
for i, dd in enumerate(d):
r += [i + 1] * dd.item()
return r
out_lens = out_lens * n_frames_per_step
finalized = [
{
"feature": feat[b, :l] if l > 0 else feat.new_zeros([1, raw_dim]),
"waveform": self.get_waveform(
feat[b, :l] if l > 0 else feat.new_zeros([1, raw_dim])
),
"attn": feat.new_tensor(get_dur_plot_data(dur_out[b])),
}
for b, l in zip(range(bsz), out_lens)
]
if has_targ:
tgt_feats = sample["target"].view(bsz, -1, raw_dim)
tgt_feats = self.gcmvn_denormalize(tgt_feats)
tgt_lens = sample["target_lengths"] * n_frames_per_step
for b, (f, l) in enumerate(zip(tgt_feats, tgt_lens)):
finalized[b]["targ_feature"] = f[:l]
finalized[b]["targ_waveform"] = self.get_waveform(f[:l])
return finalized
class TeacherForcingAutoRegressiveSpeechGenerator(AutoRegressiveSpeechGenerator):
@torch.no_grad()
def generate(self, model, sample, has_targ=False, **kwargs):
model.eval()
src_tokens = sample["net_input"]["src_tokens"]
src_lens = sample["net_input"]["src_lengths"]
prev_out_tokens = sample["net_input"]["prev_output_tokens"]
tgt_lens = sample["target_lengths"]
n_frames_per_step = model.decoder.n_frames_per_step
raw_dim = model.decoder.out_dim // n_frames_per_step
bsz = src_tokens.shape[0]
feat, eos_prob, extra = model(
src_tokens,
src_lens,
prev_out_tokens,
incremental_state=None,
target_lengths=tgt_lens,
speaker=sample["speaker"],
)
attn = extra["attn"] # B x T_s x T_t
alignment = attn.max(dim=1)[1]
feat = feat.reshape(bsz, -1, raw_dim)
feat = self.gcmvn_denormalize(feat)
eos_prob = eos_prob.repeat_interleave(n_frames_per_step, dim=1)
attn = attn.repeat_interleave(n_frames_per_step, dim=2)
alignment = alignment.repeat_interleave(n_frames_per_step, dim=1)
tgt_lens = sample["target_lengths"] * n_frames_per_step
finalized = [
{
"feature": feat[b, :tgt_len],
"eos_prob": eos_prob[b, :tgt_len],
"attn": attn[b, :, :tgt_len],
"alignment": alignment[b, :tgt_len],
"waveform": self.get_waveform(feat[b, :tgt_len]),
}
for b, tgt_len in zip(range(bsz), tgt_lens)
]
if has_targ:
tgt_feats = sample["target"].view(bsz, -1, raw_dim)
tgt_feats = self.gcmvn_denormalize(tgt_feats)
for b, (f, l) in enumerate(zip(tgt_feats, tgt_lens)):
finalized[b]["targ_feature"] = f[:l]
finalized[b]["targ_waveform"] = self.get_waveform(f[:l])
return finalized
|