class FastSpeech2HifiGanE2EModel(TextToWaveform):
"""An end-to-end speech synthesis model based on FastSpeech2 and HiFiGan that converts strings to audio without
using the intermediate mel spectrogram representation."""
def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None):
if isinstance(cfg, dict):
cfg = OmegaConf.create(cfg)
super().__init__(cfg=cfg, trainer=trainer)
self.audio_to_melspec_precessor = instantiate(cfg.preprocessor)
self.encoder = instantiate(cfg.encoder)
self.variance_adapter = instantiate(cfg.variance_adaptor)
self.generator = instantiate(cfg.generator)
self.multiperioddisc = MultiPeriodDiscriminator()
self.multiscaledisc = MultiScaleDiscriminator()
self.melspec_fn = instantiate(cfg.preprocessor,
highfreq=None,
use_grads=True)
self.mel_val_loss = L1MelLoss()
self.durationloss = DurationLoss()
self.feat_matching_loss = FeatureMatchingLoss()
self.disc_loss = DiscriminatorLoss()
self.gen_loss = GeneratorLoss()
self.mseloss = torch.nn.MSELoss()
self.energy = cfg.add_energy_predictor
self.pitch = cfg.add_pitch_predictor
self.mel_loss_coeff = cfg.mel_loss_coeff
self.pitch_loss_coeff = cfg.pitch_loss_coeff
self.energy_loss_coeff = cfg.energy_loss_coeff
self.splice_length = cfg.splice_length
self.use_energy_pred = False
self.use_pitch_pred = False
self.log_train_images = False
self.logged_real_samples = False
self._tb_logger = None
self.sample_rate = cfg.sample_rate
self.hop_size = cfg.hop_size
# Parser and mappings are used for inference only.
self.parser = parsers.make_parser(name='en')
if 'mappings_filepath' in cfg:
mappings_filepath = cfg.get('mappings_filepath')
else:
logging.error(
"ERROR: You must specify a mappings.json file in the config file under model.mappings_filepath."
)
mappings_filepath = self.register_artifact('mappings_filepath',
mappings_filepath)
with open(mappings_filepath, 'r') as f:
mappings = json.load(f)
self.word2phones = mappings['word2phones']
self.phone2idx = mappings['phone2idx']
@property
def tb_logger(self):
if self._tb_logger is None:
if self.logger is None and self.logger.experiment is None:
return None
tb_logger = self.logger.experiment
if isinstance(self.logger, LoggerCollection):
for logger in self.logger:
if isinstance(logger, TensorBoardLogger):
tb_logger = logger.experiment
break
self._tb_logger = tb_logger
return self._tb_logger
def configure_optimizers(self):
gen_params = chain(
self.encoder.parameters(),
self.generator.parameters(),
self.variance_adapter.parameters(),
)
disc_params = chain(self.multiscaledisc.parameters(),
self.multiperioddisc.parameters())
opt1 = torch.optim.AdamW(disc_params, lr=self._cfg.lr)
opt2 = torch.optim.AdamW(gen_params, lr=self._cfg.lr)
num_procs = self._trainer.num_gpus * self._trainer.num_nodes
num_samples = len(self._train_dl.dataset)
batch_size = self._train_dl.batch_size
iter_per_epoch = np.ceil(num_samples / (num_procs * batch_size))
max_steps = iter_per_epoch * self._trainer.max_epochs
logging.info(f"MAX STEPS: {max_steps}")
sch1 = NoamAnnealing(opt1,
d_model=256,
warmup_steps=3000,
max_steps=max_steps,
min_lr=1e-5)
sch1_dict = {
'scheduler': sch1,
'interval': 'step',
}
sch2 = NoamAnnealing(opt2,
d_model=256,
warmup_steps=3000,
max_steps=max_steps,
min_lr=1e-5)
sch2_dict = {
'scheduler': sch2,
'interval': 'step',
}
return [opt1, opt2], [sch1_dict, sch2_dict]
@typecheck(
input_types={
"text":
NeuralType(('B', 'T'), TokenIndex()),
"text_length":
NeuralType(('B'), LengthsType()),
"splice":
NeuralType(optional=True),
"spec_len":
NeuralType(('B'), LengthsType(), optional=True),
"durations":
NeuralType(('B', 'T'), TokenDurationType(), optional=True),
"pitch":
NeuralType(('B', 'T'), RegressionValuesType(), optional=True),
"energies":
NeuralType(('B', 'T'), RegressionValuesType(), optional=True),
},
output_types={
"audio": NeuralType(('B', 'S', 'T'), MelSpectrogramType()),
"splices": NeuralType(),
"log_dur_preds": NeuralType(('B', 'T'), TokenLogDurationType()),
"pitch_preds": NeuralType(('B', 'T'), RegressionValuesType()),
"energy_preds": NeuralType(('B', 'T'), RegressionValuesType()),
"encoded_text_mask": NeuralType(('B', 'T', 'D'), MaskType()),
},
)
def forward(self,
*,
text,
text_length,
splice=True,
durations=None,
pitch=None,
energies=None,
spec_len=None):
encoded_text, encoded_text_mask = self.encoder(text=text,
text_length=text_length)
context, log_dur_preds, pitch_preds, energy_preds, spec_len = self.variance_adapter(
x=encoded_text,
x_len=text_length,
dur_target=durations,
pitch_target=pitch,
energy_target=energies,
spec_len=spec_len,
)
gen_in = context
splices = None
if splice:
# Splice generated spec
output = []
splices = []
for i, sample in enumerate(context):
start = np.random.randint(
low=0,
high=min(int(sample.size(0)), int(spec_len[i])) -
self.splice_length)
output.append(sample[start:start + self.splice_length, :])
splices.append(start)
gen_in = torch.stack(output)
output = self.generator(x=gen_in.transpose(1, 2))
return output, splices, log_dur_preds, pitch_preds, energy_preds, encoded_text_mask
def training_step(self, batch, batch_idx, optimizer_idx):
f, fl, t, tl, durations, pitch, energies = batch
spec, spec_len = self.audio_to_melspec_precessor(f, fl)
# train discriminator
if optimizer_idx == 0:
with torch.no_grad():
audio_pred, splices, _, _, _, _ = self(
spec=spec,
spec_len=spec_len,
text=t,
text_length=tl,
durations=durations,
pitch=pitch if not self.use_pitch_pred else None,
energies=energies if not self.use_energy_pred else None,
)
real_audio = []
for i, splice in enumerate(splices):
real_audio.append(
f[i, splice *
self.hop_size:(splice + self.splice_length) *
self.hop_size])
real_audio = torch.stack(real_audio).unsqueeze(1)
real_score_mp, gen_score_mp, _, _ = self.multiperioddisc(
real_audio, audio_pred)
real_score_ms, gen_score_ms, _, _ = self.multiscaledisc(
real_audio, audio_pred)
loss_mp, loss_mp_real, _ = self.disc_loss(real_score_mp,
gen_score_mp)
loss_ms, loss_ms_real, _ = self.disc_loss(real_score_ms,
gen_score_ms)
loss_mp /= len(loss_mp_real)
loss_ms /= len(loss_ms_real)
loss_disc = loss_mp + loss_ms
self.log("loss_discriminator", loss_disc, prog_bar=True)
self.log("loss_discriminator_ms", loss_ms)
self.log("loss_discriminator_mp", loss_mp)
return loss_disc
# train generator
elif optimizer_idx == 1:
audio_pred, splices, log_dur_preds, pitch_preds, energy_preds, encoded_text_mask = self(
spec=spec,
spec_len=spec_len,
text=t,
text_length=tl,
durations=durations,
pitch=pitch if not self.use_pitch_pred else None,
energies=energies if not self.use_energy_pred else None,
)
real_audio = []
for i, splice in enumerate(splices):
real_audio.append(
f[i, splice * self.hop_size:(splice + self.splice_length) *
self.hop_size])
real_audio = torch.stack(real_audio).unsqueeze(1)
# Do HiFiGAN generator loss
audio_length = torch.tensor([
self.splice_length * self.hop_size
for _ in range(real_audio.shape[0])
]).to(real_audio.device)
real_spliced_spec, _ = self.melspec_fn(real_audio.squeeze(),
seq_len=audio_length)
pred_spliced_spec, _ = self.melspec_fn(audio_pred.squeeze(),
seq_len=audio_length)
loss_mel = torch.nn.functional.l1_loss(real_spliced_spec,
pred_spliced_spec)
loss_mel *= self.mel_loss_coeff
_, gen_score_mp, real_feat_mp, gen_feat_mp = self.multiperioddisc(
real_audio, audio_pred)
_, gen_score_ms, real_feat_ms, gen_feat_ms = self.multiscaledisc(
real_audio, audio_pred)
loss_gen_mp, list_loss_gen_mp = self.gen_loss(gen_score_mp)
loss_gen_ms, list_loss_gen_ms = self.gen_loss(gen_score_ms)
loss_gen_mp /= len(list_loss_gen_mp)
loss_gen_ms /= len(list_loss_gen_ms)
total_loss = loss_gen_mp + loss_gen_ms + loss_mel
loss_feat_mp = self.feat_matching_loss(real_feat_mp, gen_feat_mp)
loss_feat_ms = self.feat_matching_loss(real_feat_ms, gen_feat_ms)
total_loss += loss_feat_mp + loss_feat_ms
self.log(name="loss_gen_disc_feat",
value=loss_feat_mp + loss_feat_ms)
self.log(name="loss_gen_disc_feat_ms", value=loss_feat_ms)
self.log(name="loss_gen_disc_feat_mp", value=loss_feat_mp)
self.log(name="loss_gen_mel", value=loss_mel)
self.log(name="loss_gen_disc", value=loss_gen_mp + loss_gen_ms)
self.log(name="loss_gen_disc_mp", value=loss_gen_mp)
self.log(name="loss_gen_disc_ms", value=loss_gen_ms)
dur_loss = self.durationloss(log_duration_pred=log_dur_preds,
duration_target=durations.float(),
mask=encoded_text_mask)
self.log(name="loss_gen_duration", value=dur_loss)
total_loss += dur_loss
if self.pitch:
pitch_loss = self.mseloss(
pitch_preds, pitch.float()) * self.pitch_loss_coeff
total_loss += pitch_loss
self.log(name="loss_gen_pitch", value=pitch_loss)
if self.energy:
energy_loss = self.mseloss(energy_preds,
energies) * self.energy_loss_coeff
total_loss += energy_loss
self.log(name="loss_gen_energy", value=energy_loss)
# Log images to tensorboard
if self.log_train_images:
self.log_train_images = False
if self.logger is not None and self.logger.experiment is not None:
self.tb_logger.add_image(
"train_mel_target",
plot_spectrogram_to_numpy(
real_spliced_spec[0].data.cpu().numpy()),
self.global_step,
dataformats="HWC",
)
spec_predict = pred_spliced_spec[0].data.cpu().numpy()
self.tb_logger.add_image(
"train_mel_predicted",
plot_spectrogram_to_numpy(spec_predict),
self.global_step,
dataformats="HWC",
)
self.log(name="loss_gen", prog_bar=True, value=total_loss)
return total_loss
def validation_step(self, batch, batch_idx):
f, fl, t, tl, _, _, _ = batch
spec, spec_len = self.audio_to_melspec_precessor(f, fl)
audio_pred, _, _, _, _, _ = self(spec=spec,
spec_len=spec_len,
text=t,
text_length=tl,
splice=False)
audio_pred.squeeze_()
pred_spec, _ = self.melspec_fn(audio_pred, seq_len=spec_len)
loss = self.mel_val_loss(spec_pred=pred_spec,
spec_target=spec,
spec_target_len=spec_len,
pad_value=-11.52)
return {
"val_loss": loss,
"audio_target": f.squeeze() if batch_idx == 0 else None,
"audio_pred": audio_pred if batch_idx == 0 else None,
}
def on_train_epoch_start(self):
# Switch to using energy predictions after 50% of training
if not self.use_energy_pred and self.current_epoch >= np.ceil(
0.5 * self._trainer.max_epochs):
logging.info(
f"Using energy predictions after epoch: {self.current_epoch}")
self.use_energy_pred = True
# Switch to using pitch predictions after 62.5% of training
if not self.use_pitch_pred and self.current_epoch >= np.ceil(
0.625 * self._trainer.max_epochs):
logging.info(
f"Using pitch predictions after epoch: {self.current_epoch}")
self.use_pitch_pred = True
def validation_epoch_end(self, outputs):
if self.tb_logger is not None:
_, audio_target, audio_predict = outputs[0].values()
if not self.logged_real_samples:
self.tb_logger.add_audio("val_target",
audio_target[0].data.cpu(),
self.global_step, self.sample_rate)
self.logged_real_samples = True
audio_predict = audio_predict[0].data.cpu()
self.tb_logger.add_audio("val_pred", audio_predict,
self.global_step, self.sample_rate)
avg_loss = torch.stack([
x['val_loss'] for x in outputs
]).mean() # This reduces across batches, not workers!
self.log('val_loss', avg_loss, sync_dist=True)
self.log_train_images = True
def __setup_dataloader_from_config(self,
cfg,
shuffle_should_be: bool = True,
name: str = "train"):
if "dataset" not in cfg or not isinstance(cfg.dataset, DictConfig):
raise ValueError(f"No dataset for {name}")
if "dataloader_params" not in cfg or not isinstance(
cfg.dataloader_params, DictConfig):
raise ValueError(f"No dataloder_params for {name}")
if shuffle_should_be:
if 'shuffle' not in cfg.dataloader_params:
logging.warning(
f"Shuffle should be set to True for {self}'s {name} dataloader but was not found in its "
"config. Manually setting to True")
with open_dict(cfg.dataloader_params):
cfg.dataloader_params.shuffle = True
elif not cfg.dataloader_params.shuffle:
logging.error(
f"The {name} dataloader for {self} has shuffle set to False!!!"
)
elif not shuffle_should_be and cfg.dataloader_params.shuffle:
logging.error(
f"The {name} dataloader for {self} has shuffle set to True!!!")
dataset = instantiate(cfg.dataset)
return torch.utils.data.DataLoader(dataset,
collate_fn=dataset.collate_fn,
**cfg.dataloader_params)
def setup_training_data(self, cfg):
self._train_dl = self.__setup_dataloader_from_config(cfg)
def setup_validation_data(self, cfg):
self._validation_dl = self.__setup_dataloader_from_config(
cfg, shuffle_should_be=False, name="validation")
def parse(self,
str_input: str,
additional_word2phones=None) -> torch.tensor:
"""
Parses text input and converts them to phoneme indices.
str_input (str): The input text to be converted.
additional_word2phones (dict): Optional dictionary mapping words to phonemes for updating the model's
word2phones. This will not overwrite the existing dictionary, just update it with OOV or new mappings.
Defaults to None, which will keep the existing mapping.
"""
# Update model's word2phones if applicable
if additional_word2phones is not None:
self.word2phones.update(additional_word2phones)
# Convert text -> normalized text -> list of phones per word -> indices
if str_input[-1] not in [".", "!", "?"]:
str_input = str_input + "."
norm_text = re.findall(r"""[\w']+|[.,!?;"]""",
self.parser._normalize(str_input))
try:
phones = [self.word2phones[t] for t in norm_text]
except KeyError as error:
logging.error(
f"ERROR: The following word in the input is not in the model's dictionary and could not be converted"
f" to phonemes: ({error}).\n"
f"You can pass in an `additional_word2phones` dictionary with a conversion for"
f" this word, e.g. {{'{error}': \['phone1', 'phone2', ...\]}} to update the model's mapping."
)
raise
tokens = []
for phone_list in phones:
inds = [self.phone2idx[p] for p in phone_list]
tokens += inds
x = torch.tensor(tokens).unsqueeze_(0).long().to(self.device)
return x
def convert_text_to_waveform(self, *, tokens):
"""
Accepts tokens returned from self.parse() and returns a list of tensors. Note: The tensors in the list can have
different lengths.
"""
self.eval()
token_len = torch.tensor([len(i) for i in tokens]).to(self.device)
audio, _, log_dur_pred, _, _, _ = self(text=tokens,
text_length=token_len,
splice=False)
audio = audio.squeeze(1)
durations = torch.sum(torch.exp(log_dur_pred) - 1, 1).to(torch.int)
audio_list = []
for i, sample in enumerate(audio):
audio_list.append(sample[:durations[i] * self.hop_size])
return audio_list
@classmethod
def list_available_models(cls) -> 'List[PretrainedModelInfo]':
"""
This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud.
Returns:
List of available pre-trained models.
"""
list_of_models = []
model = PretrainedModelInfo(
pretrained_model_name="tts_en_e2e_fastspeech2hifigan",
location=
"https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_e2e_fastspeech2hifigan/versions/1.0.0/files/tts_en_e2e_fastspeech2hifigan.nemo",
description=
"This model is trained on LJSpeech sampled at 22050Hz with and can be used to generate female English voices with an American accent.",
class_=cls,
)
list_of_models.append(model)
return list_of_models
class HifiGanModel(Vocoder):
def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None):
if isinstance(cfg, dict):
cfg = OmegaConf.create(cfg)
super().__init__(cfg=cfg, trainer=trainer)
self.audio_to_melspec_precessor = instantiate(cfg.preprocessor)
# use a different melspec extractor because:
# 1. we need to pass grads
# 2. we need remove fmax limitation
self.trg_melspec_fn = instantiate(cfg.preprocessor,
highfreq=None,
use_grads=True)
self.generator = instantiate(cfg.generator)
self.mpd = MultiPeriodDiscriminator()
self.msd = MultiScaleDiscriminator()
self.feature_loss = FeatureMatchingLoss()
self.discriminator_loss = DiscriminatorLoss()
self.generator_loss = GeneratorLoss()
self.sample_rate = self._cfg.preprocessor.sample_rate
self.stft_bias = None
if isinstance(self._train_dl.dataset, MelAudioDataset):
self.finetune = True
logging.info("fine-tuning on pre-computed mels")
else:
self.finetune = False
logging.info("training on ground-truth mels")
def configure_optimizers(self):
self.optim_g = instantiate(
self._cfg.optim,
params=self.generator.parameters(),
)
self.optim_d = instantiate(
self._cfg.optim,
params=itertools.chain(self.msd.parameters(),
self.mpd.parameters()),
)
max_steps = self._cfg.max_steps
warmup_steps = 0 if self.finetune else np.ceil(0.2 * max_steps)
self.scheduler_g = CosineAnnealing(
self.optim_g,
max_steps=max_steps,
min_lr=1e-5,
warmup_steps=warmup_steps) # Use warmup to delay start
sch1_dict = {
'scheduler': self.scheduler_g,
'interval': 'step',
}
self.scheduler_d = CosineAnnealing(self.optim_d,
max_steps=max_steps,
min_lr=1e-5)
sch2_dict = {
'scheduler': self.scheduler_d,
'interval': 'step',
}
return [self.optim_g, self.optim_d], [sch1_dict, sch2_dict]
@property
def input_types(self):
return {
"spec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
}
@property
def output_types(self):
return {
"audio": NeuralType(('B', 'S', 'T'),
AudioSignal(self.sample_rate)),
}
@typecheck()
def forward(self, *, spec):
"""
Runs the generator, for inputs and outputs see input_types, and output_types
"""
return self.generator(x=spec)
@typecheck(output_types={"audio": NeuralType(('B', 'T'), AudioSignal())})
def convert_spectrogram_to_audio(self,
spec: 'torch.tensor') -> 'torch.tensor':
return self(spec=spec).squeeze(1)
def training_step(self, batch, batch_idx, optimizer_idx):
# if in finetune mode the mels are pre-computed using a
# spectrogram generator
if self.finetune:
audio, audio_len, audio_mel = batch
# else, we compute the mel using the ground truth audio
else:
audio, audio_len = batch
# mel as input for generator
audio_mel, _ = self.audio_to_melspec_precessor(audio, audio_len)
# mel as input for L1 mel loss
audio_trg_mel, _ = self.trg_melspec_fn(audio, audio_len)
audio = audio.unsqueeze(1)
audio_pred = self.generator(x=audio_mel)
audio_pred_mel, _ = self.trg_melspec_fn(audio_pred.squeeze(1),
audio_len)
# train discriminator
self.optim_d.zero_grad()
mpd_score_real, mpd_score_gen, _, _ = self.mpd(
y=audio, y_hat=audio_pred.detach())
loss_disc_mpd, _, _ = self.discriminator_loss(
disc_real_outputs=mpd_score_real,
disc_generated_outputs=mpd_score_gen)
msd_score_real, msd_score_gen, _, _ = self.msd(
y=audio, y_hat=audio_pred.detach())
loss_disc_msd, _, _ = self.discriminator_loss(
disc_real_outputs=msd_score_real,
disc_generated_outputs=msd_score_gen)
loss_d = loss_disc_msd + loss_disc_mpd
self.manual_backward(loss_d, self.optim_d)
self.optim_d.step()
# train generator
self.optim_g.zero_grad()
loss_mel = F.l1_loss(audio_pred_mel, audio_trg_mel) * 45
_, mpd_score_gen, fmap_mpd_real, fmap_mpd_gen = self.mpd(
y=audio, y_hat=audio_pred)
_, msd_score_gen, fmap_msd_real, fmap_msd_gen = self.msd(
y=audio, y_hat=audio_pred)
loss_fm_mpd = self.feature_loss(fmap_r=fmap_mpd_real,
fmap_g=fmap_mpd_gen)
loss_fm_msd = self.feature_loss(fmap_r=fmap_msd_real,
fmap_g=fmap_msd_gen)
loss_gen_mpd, _ = self.generator_loss(disc_outputs=mpd_score_gen)
loss_gen_msd, _ = self.generator_loss(disc_outputs=msd_score_gen)
loss_g = loss_gen_msd + loss_gen_mpd + loss_fm_msd + loss_fm_mpd + loss_mel
self.manual_backward(loss_g, self.optim_g)
self.optim_g.step()
metrics = {
"g_l1_loss": loss_mel,
"g_loss_fm_mpd": loss_fm_mpd,
"g_loss_fm_msd": loss_fm_msd,
"g_loss_gen_mpd": loss_gen_mpd,
"g_loss_gen_msd": loss_gen_msd,
"g_loss": loss_g,
"d_loss_mpd": loss_disc_mpd,
"d_loss_msd": loss_disc_msd,
"d_loss": loss_d,
"global_step": self.global_step,
"lr": self.optim_g.param_groups[0]['lr'],
}
self.log_dict(metrics, on_step=False, on_epoch=True, sync_dist=True)
def validation_step(self, batch, batch_idx):
if self.finetune:
audio, audio_len, audio_mel = batch
audio_mel_len = [audio_mel.shape[1]] * audio_mel.shape[0]
else:
audio, audio_len = batch
audio_mel, audio_mel_len = self.audio_to_melspec_precessor(
audio, audio_len)
audio_pred = self(spec=audio_mel)
# perform bias denoising
pred_denoised = self._bias_denoise(audio_pred, audio_mel).squeeze(1)
pred_denoised_mel, _ = self.audio_to_melspec_precessor(
pred_denoised, audio_len)
if self.finetune:
gt_mel, gt_mel_len = self.audio_to_melspec_precessor(
audio, audio_len)
audio_pred_mel, _ = self.audio_to_melspec_precessor(
audio_pred.squeeze(1), audio_len)
loss_mel = F.l1_loss(audio_mel, audio_pred_mel)
self.log("val_loss", loss_mel, prog_bar=True, sync_dist=True)
# plot audio once per epoch
if batch_idx == 0 and isinstance(self.logger,
WandbLogger) and HAVE_WANDB:
clips = []
specs = []
for i in range(min(5, audio.shape[0])):
clips += [
wandb.Audio(
audio[i, :audio_len[i]].data.cpu().numpy(),
caption=f"real audio {i}",
sample_rate=self.sample_rate,
),
wandb.Audio(
audio_pred[i,
0, :audio_len[i]].data.cpu().numpy().astype(
'float32'),
caption=f"generated audio {i}",
sample_rate=self.sample_rate,
),
wandb.Audio(
pred_denoised[i, :audio_len[i]].data.cpu().numpy(),
caption=f"denoised audio {i}",
sample_rate=self.sample_rate,
),
]
specs += [
wandb.Image(
plot_spectrogram_to_numpy(audio_mel[
i, :, :audio_mel_len[i]].data.cpu().numpy()),
caption=f"input mel {i}",
),
wandb.Image(
plot_spectrogram_to_numpy(audio_pred_mel[
i, :, :audio_mel_len[i]].data.cpu().numpy()),
caption=f"output mel {i}",
),
wandb.Image(
plot_spectrogram_to_numpy(pred_denoised_mel[
i, :, :audio_mel_len[i]].data.cpu().numpy()),
caption=f"denoised mel {i}",
),
]
if self.finetune:
specs += [
wandb.Image(
plot_spectrogram_to_numpy(gt_mel[
i, :, :audio_mel_len[i]].data.cpu().numpy()),
caption=f"gt mel {i}",
),
]
self.logger.experiment.log({
"audio": clips,
"specs": specs
},
commit=False)
def _bias_denoise(self, audio, mel):
def stft(x):
comp = torch.stft(x.squeeze(1),
n_fft=1024,
hop_length=256,
win_length=1024)
real, imag = comp[..., 0], comp[..., 1]
mags = torch.sqrt(real**2 + imag**2)
phase = torch.atan2(imag, real)
return mags, phase
def istft(mags, phase):
comp = torch.stack(
[mags * torch.cos(phase), mags * torch.sin(phase)], dim=-1)
x = torch.istft(comp, n_fft=1024, hop_length=256, win_length=1024)
return x
# create bias tensor
if self.stft_bias is None:
audio_bias = self(spec=torch.zeros_like(mel, device=mel.device))
self.stft_bias, _ = stft(audio_bias)
self.stft_bias = self.stft_bias[:, :, 0][:, :, None]
audio_mags, audio_phase = stft(audio)
audio_mags = audio_mags - self.cfg.denoise_strength * self.stft_bias
audio_mags = torch.clamp(audio_mags, 0.0)
audio_denoised = istft(audio_mags, audio_phase).unsqueeze(1)
return audio_denoised
def __setup_dataloader_from_config(self,
cfg,
shuffle_should_be: bool = True,
name: str = "train"):
if "dataset" not in cfg or not isinstance(cfg.dataset, DictConfig):
raise ValueError(f"No dataset for {name}")
if "dataloader_params" not in cfg or not isinstance(
cfg.dataloader_params, DictConfig):
raise ValueError(f"No dataloder_params for {name}")
if shuffle_should_be:
if 'shuffle' not in cfg.dataloader_params:
logging.warning(
f"Shuffle should be set to True for {self}'s {name} dataloader but was not found in its "
"config. Manually setting to True")
with open_dict(cfg["dataloader_params"]):
cfg.dataloader_params.shuffle = True
elif not cfg.dataloader_params.shuffle:
logging.error(
f"The {name} dataloader for {self} has shuffle set to False!!!"
)
elif not shuffle_should_be and cfg.dataloader_params.shuffle:
logging.error(
f"The {name} dataloader for {self} has shuffle set to True!!!")
dataset = instantiate(cfg.dataset)
return torch.utils.data.DataLoader(dataset,
collate_fn=dataset.collate_fn,
**cfg.dataloader_params)
def setup_training_data(self, cfg):
self._train_dl = self.__setup_dataloader_from_config(cfg)
def setup_validation_data(self, cfg):
self._validation_dl = self.__setup_dataloader_from_config(
cfg, shuffle_should_be=False, name="validation")
@classmethod
def list_available_models(cls) -> 'Optional[Dict[str, str]]':
# TODO
pass
class FastPitchHifiGanE2EModel(TextToWaveform):
"""An end-to-end speech synthesis model based on FastPitch and HiFiGan that converts strings to audio without using
the intermediate mel spectrogram representation.
"""
def __init__(self, cfg: DictConfig, trainer: Trainer = None):
if isinstance(cfg, dict):
cfg = OmegaConf.create(cfg)
self._parser = parsers.make_parser(
labels=cfg.labels,
name='en',
unk_id=-1,
blank_id=-1,
do_normalize=True,
abbreviation_version="fastpitch",
make_table=False,
)
super().__init__(cfg=cfg, trainer=trainer)
schema = OmegaConf.structured(FastPitchHifiGanE2EConfig)
# ModelPT ensures that cfg is a DictConfig, but do this second check in case ModelPT changes
if isinstance(cfg, dict):
cfg = OmegaConf.create(cfg)
elif not isinstance(cfg, DictConfig):
raise ValueError(
f"cfg was type: {type(cfg)}. Expected either a dict or a DictConfig"
)
# Ensure passed cfg is compliant with schema
OmegaConf.merge(cfg, schema)
self.preprocessor = instantiate(cfg.preprocessor)
self.melspec_fn = instantiate(cfg.preprocessor,
highfreq=None,
use_grads=True)
self.encoder = instantiate(cfg.input_fft)
self.duration_predictor = instantiate(cfg.duration_predictor)
self.pitch_predictor = instantiate(cfg.pitch_predictor)
self.generator = instantiate(cfg.generator)
self.multiperioddisc = MultiPeriodDiscriminator()
self.multiscaledisc = MultiScaleDiscriminator()
self.mel_val_loss = L1MelLoss()
self.feat_matching_loss = FeatureMatchingLoss()
self.disc_loss = DiscriminatorLoss()
self.gen_loss = GeneratorLoss()
self.max_token_duration = cfg.max_token_duration
self.pitch_emb = torch.nn.Conv1d(
1,
cfg.symbols_embedding_dim,
kernel_size=cfg.pitch_embedding_kernel_size,
padding=int((cfg.pitch_embedding_kernel_size - 1) / 2),
)
# Store values precomputed from training data for convenience
self.register_buffer('pitch_mean', torch.zeros(1))
self.register_buffer('pitch_std', torch.zeros(1))
self.pitchloss = PitchLoss()
self.durationloss = DurationLoss()
self.mel_loss_coeff = cfg.mel_loss_coeff
self.log_train_images = False
self.logged_real_samples = False
self._tb_logger = None
self.hann_window = None
self.splice_length = cfg.splice_length
self.sample_rate = cfg.sample_rate
self.hop_size = cfg.hop_size
@property
def tb_logger(self):
if self._tb_logger is None:
if self.logger is None and self.logger.experiment is None:
return None
tb_logger = self.logger.experiment
if isinstance(self.logger, LoggerCollection):
for logger in self.logger:
if isinstance(logger, TensorBoardLogger):
tb_logger = logger.experiment
break
self._tb_logger = tb_logger
return self._tb_logger
@property
def parser(self):
if self._parser is not None:
return self._parser
self._parser = parsers.make_parser(
labels=self._cfg.labels,
name='en',
unk_id=-1,
blank_id=-1,
do_normalize=True,
abbreviation_version="fastpitch",
make_table=False,
)
return self._parser
def parse(self, str_input: str) -> torch.tensor:
if str_input[-1] not in [".", "!", "?"]:
str_input = str_input + "."
tokens = self.parser(str_input)
x = torch.tensor(tokens).unsqueeze_(0).long().to(self.device)
return x
def configure_optimizers(self):
gen_params = chain(
self.pitch_emb.parameters(),
self.encoder.parameters(),
self.duration_predictor.parameters(),
self.pitch_predictor.parameters(),
self.generator.parameters(),
)
disc_params = chain(self.multiscaledisc.parameters(),
self.multiperioddisc.parameters())
opt1 = torch.optim.AdamW(disc_params, lr=self._cfg.lr)
opt2 = torch.optim.AdamW(gen_params, lr=self._cfg.lr)
num_procs = self._trainer.num_gpus * self._trainer.num_nodes
num_samples = len(self._train_dl.dataset)
batch_size = self._train_dl.batch_size
iter_per_epoch = np.ceil(num_samples / (num_procs * batch_size))
max_steps = iter_per_epoch * self._trainer.max_epochs
logging.info(f"MAX STEPS: {max_steps}")
sch1 = NoamAnnealing(opt1,
d_model=1,
warmup_steps=1000,
max_steps=max_steps,
last_epoch=-1)
sch1_dict = {
'scheduler': sch1,
'interval': 'step',
}
sch2 = NoamAnnealing(opt2,
d_model=1,
warmup_steps=1000,
max_steps=max_steps,
last_epoch=-1)
sch2_dict = {
'scheduler': sch2,
'interval': 'step',
}
return [opt1, opt2], [sch1_dict, sch2_dict]
@typecheck(
input_types={
"text": NeuralType(('B', 'T'), TokenIndex()),
"durs": NeuralType(('B', 'T'), TokenDurationType(), optional=True),
"pitch": NeuralType(('B', 'T'),
RegressionValuesType(),
optional=True),
"pace": NeuralType(optional=True),
"splice": NeuralType(optional=True),
},
output_types={
"audio": NeuralType(('B', 'S', 'T'), AudioSignal()),
"splices": NeuralType(),
"log_dur_preds": NeuralType(('B', 'T'), TokenLogDurationType()),
"pitch_preds": NeuralType(('B', 'T'), RegressionValuesType()),
},
)
def forward(self, *, text, durs=None, pitch=None, pace=1.0, splice=True):
if self.training:
assert durs is not None
assert pitch is not None
# Input FFT
enc_out, enc_mask = self.encoder(input=text, conditioning=0)
# Embedded for predictors
pred_enc_out, pred_enc_mask = enc_out, enc_mask
# Predict durations
log_durs_predicted = self.duration_predictor(pred_enc_out,
pred_enc_mask)
durs_predicted = torch.clamp(
torch.exp(log_durs_predicted) - 1, 0, self.max_token_duration)
# Predict pitch
pitch_predicted = self.pitch_predictor(enc_out, enc_mask)
if pitch is None:
pitch_emb = self.pitch_emb(pitch_predicted.unsqueeze(1))
else:
pitch_emb = self.pitch_emb(pitch.unsqueeze(1))
enc_out = enc_out + pitch_emb.transpose(1, 2)
if durs is None:
len_regulated, dec_lens = regulate_len(durs_predicted, enc_out,
pace)
else:
len_regulated, dec_lens = regulate_len(durs, enc_out, pace)
gen_in = len_regulated
splices = []
if splice:
output = []
for i, sample in enumerate(len_regulated):
start = np.random.randint(
low=0,
high=min(int(sample.size(0)), int(dec_lens[i])) -
self.splice_length)
# Splice generated spec
output.append(sample[start:start + self.splice_length, :])
splices.append(start)
gen_in = torch.stack(output)
output = self.generator(x=gen_in.transpose(1, 2))
return output, torch.tensor(
splices), log_durs_predicted, pitch_predicted
def training_step(self, batch, batch_idx, optimizer_idx):
audio, _, text, text_lens, durs, pitch, _ = batch
# train discriminator
if optimizer_idx == 0:
with torch.no_grad():
audio_pred, splices, _, _ = self(text=text,
durs=durs,
pitch=pitch)
real_audio = []
for i, splice in enumerate(splices):
real_audio.append(
audio[i, splice *
self.hop_size:(splice + self.splice_length) *
self.hop_size])
real_audio = torch.stack(real_audio).unsqueeze(1)
real_score_mp, gen_score_mp, _, _ = self.multiperioddisc(
real_audio, audio_pred)
real_score_ms, gen_score_ms, _, _ = self.multiscaledisc(
real_audio, audio_pred)
loss_mp, loss_mp_real, _ = self.disc_loss(
disc_real_outputs=real_score_mp,
disc_generated_outputs=gen_score_mp)
loss_ms, loss_ms_real, _ = self.disc_loss(
disc_real_outputs=real_score_ms,
disc_generated_outputs=gen_score_ms)
loss_mp /= len(loss_mp_real)
loss_ms /= len(loss_ms_real)
loss_disc = loss_mp + loss_ms
self.log("loss_discriminator", loss_disc, prog_bar=True)
self.log("loss_discriminator_ms", loss_ms)
self.log("loss_discriminator_mp", loss_mp)
return loss_disc
# train generator
elif optimizer_idx == 1:
audio_pred, splices, log_dur_preds, pitch_preds = self(text=text,
durs=durs,
pitch=pitch)
real_audio = []
for i, splice in enumerate(splices):
real_audio.append(
audio[i, splice *
self.hop_size:(splice + self.splice_length) *
self.hop_size])
real_audio = torch.stack(real_audio).unsqueeze(1)
dur_loss = self.durationloss(log_durs_predicted=log_dur_preds,
durs_tgt=durs,
len=text_lens)
pitch_loss = self.pitchloss(
pitch_predicted=pitch_preds,
pitch_tgt=pitch,
)
# Do HiFiGAN generator loss
audio_length = torch.tensor([
self.splice_length * self.hop_size
for _ in range(real_audio.shape[0])
]).to(real_audio.device)
real_spliced_spec, _ = self.melspec_fn(real_audio.squeeze(),
audio_length)
pred_spliced_spec, _ = self.melspec_fn(audio_pred.squeeze(),
audio_length)
loss_mel = torch.nn.functional.l1_loss(real_spliced_spec,
pred_spliced_spec)
loss_mel *= self.mel_loss_coeff
_, gen_score_mp, _, _ = self.multiperioddisc(
real_audio, audio_pred)
_, gen_score_ms, _, _ = self.multiscaledisc(y=real_audio,
y_hat=audio_pred)
loss_gen_mp, list_loss_gen_mp = self.gen_loss(
disc_outputs=gen_score_mp)
loss_gen_ms, list_loss_gen_ms = self.gen_loss(
disc_outputs=gen_score_ms)
loss_gen_mp /= len(list_loss_gen_mp)
loss_gen_ms /= len(list_loss_gen_ms)
total_loss = loss_gen_mp + loss_gen_ms + loss_mel
total_loss += dur_loss
total_loss += pitch_loss
self.log(name="loss_gen_mel", value=loss_mel)
self.log(name="loss_gen_disc", value=loss_gen_mp + loss_gen_ms)
self.log(name="loss_gen_disc_mp", value=loss_gen_mp)
self.log(name="loss_gen_disc_ms", value=loss_gen_ms)
self.log(name="loss_gen_duration", value=dur_loss)
self.log(name="loss_gen_pitch", value=pitch_loss)
# Log images to tensorboard
if self.log_train_images:
self.log_train_images = False
if self.logger is not None and self.logger.experiment is not None:
self.tb_logger.add_image(
"train_mel_target",
plot_spectrogram_to_numpy(
real_spliced_spec[0].data.cpu().numpy()),
self.global_step,
dataformats="HWC",
)
spec_predict = pred_spliced_spec[0].data.cpu().numpy()
self.tb_logger.add_image(
"train_mel_predicted",
plot_spectrogram_to_numpy(spec_predict),
self.global_step,
dataformats="HWC",
)
self.log(name="loss_gen", prog_bar=True, value=total_loss)
return total_loss
def validation_step(self, batch, batch_idx):
audio, audio_lens, text, _, _, _, _ = batch
mels, mel_lens = self.preprocessor(audio, audio_lens)
audio_pred, _, log_durs_predicted, _ = self(text=text,
durs=None,
pitch=None,
splice=False)
audio_length = torch.sum(torch.clamp(torch.exp(log_durs_predicted - 1),
0),
axis=1)
audio_pred.squeeze_()
pred_spec, _ = self.melspec_fn(audio_pred, audio_length)
loss = self.mel_val_loss(spec_pred=pred_spec,
spec_target=mels,
spec_target_len=mel_lens,
pad_value=-11.52,
transpose=False)
return {
"val_loss": loss,
"audio_target": audio if batch_idx == 0 else None,
"audio_pred": audio_pred.squeeze() if batch_idx == 0 else None,
}
def validation_epoch_end(self, outputs):
if self.tb_logger is not None:
_, audio_target, audio_predict = outputs[0].values()
if not self.logged_real_samples:
self.tb_logger.add_audio("val_target",
audio_target[0].data.cpu(),
self.global_step, self.sample_rate)
self.logged_real_samples = True
audio_predict = audio_predict[0].data.cpu()
self.tb_logger.add_audio("val_pred", audio_predict,
self.global_step, self.sample_rate)
avg_loss = torch.stack([
x['val_loss'] for x in outputs
]).mean() # This reduces across batches, not workers!
self.log('val_loss', avg_loss, sync_dist=True)
self.log_train_images = True
def _loader(self, cfg):
dataset = FastPitchDataset(
manifest_filepath=cfg['manifest_filepath'],
parser=self.parser,
sample_rate=cfg['sample_rate'],
int_values=cfg.get('int_values', False),
max_duration=cfg.get('max_duration', None),
min_duration=cfg.get('min_duration', None),
max_utts=cfg.get('max_utts', 0),
trim=cfg.get('trim_silence', True),
)
return torch.utils.data.DataLoader(
dataset=dataset,
batch_size=cfg['batch_size'],
collate_fn=dataset.collate_fn,
drop_last=cfg.get('drop_last', True),
shuffle=cfg['shuffle'],
num_workers=cfg.get('num_workers', 16),
)
def setup_training_data(self, cfg):
self._train_dl = self._loader(cfg)
def setup_validation_data(self, cfg):
self._validation_dl = self._loader(cfg)
def setup_test_data(self, cfg):
"""Omitted."""
pass
@classmethod
def list_available_models(cls) -> 'List[PretrainedModelInfo]':
"""
This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud.
Returns:
List of available pre-trained models.
"""
list_of_models = []
model = PretrainedModelInfo(
pretrained_model_name="tts_en_e2e_fastpitchhifigan",
location=
"https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_e2e_fastpitchhifigan/versions/1.0.0/files/tts_en_e2e_fastpitchhifigan.nemo",
description=
"This model is trained on LJSpeech sampled at 22050Hz with and can be used to generate female English voices with an American accent.",
class_=cls,
)
list_of_models.append(model)
return list_of_models
def convert_text_to_waveform(self, *, tokens):
"""
Accepts tokens returned from self.parse() and returns a list of tensors. Note: The tensors in the list can have
different lengths.
"""
self.eval()
audio, _, log_dur_pred, _ = self(text=tokens, splice=False)
audio = audio.squeeze(1)
durations = torch.sum(
torch.clamp(
torch.exp(log_dur_pred) - 1, 0, self.max_token_duration),
1).to(torch.int)
audio_list = []
for i, sample in enumerate(audio):
audio_list.append(sample[:durations[i] * self.hop_size])
return audio_list
class HifiGanModel(Vocoder, Exportable):
def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None):
if isinstance(cfg, dict):
cfg = OmegaConf.create(cfg)
super().__init__(cfg=cfg, trainer=trainer)
self.audio_to_melspec_precessor = instantiate(cfg.preprocessor)
# use a different melspec extractor because:
# 1. we need to pass grads
# 2. we need remove fmax limitation
self.trg_melspec_fn = instantiate(cfg.preprocessor,
highfreq=None,
use_grads=True)
self.generator = instantiate(cfg.generator)
self.mpd = MultiPeriodDiscriminator()
self.msd = MultiScaleDiscriminator()
self.feature_loss = FeatureMatchingLoss()
self.discriminator_loss = DiscriminatorLoss()
self.generator_loss = GeneratorLoss()
self.l1_factor = cfg.get("l1_loss_factor", 45)
self.sample_rate = self._cfg.preprocessor.sample_rate
self.stft_bias = None
if self._train_dl and isinstance(self._train_dl.dataset,
MelAudioDataset):
self.input_as_mel = True
else:
self.input_as_mel = False
self.automatic_optimization = False
def configure_optimizers(self):
self.optim_g = instantiate(
self._cfg.optim,
params=self.generator.parameters(),
)
self.optim_d = instantiate(
self._cfg.optim,
params=itertools.chain(self.msd.parameters(),
self.mpd.parameters()),
)
self.scheduler_g = CosineAnnealing(
optimizer=self.optim_g,
max_steps=self._cfg.max_steps,
min_lr=self._cfg.sched.min_lr,
warmup_steps=self._cfg.sched.warmup_ratio * self._cfg.max_steps,
) # Use warmup to delay start
sch1_dict = {
'scheduler': self.scheduler_g,
'interval': 'step',
}
self.scheduler_d = CosineAnnealing(
optimizer=self.optim_d,
max_steps=self._cfg.max_steps,
min_lr=self._cfg.sched.min_lr,
)
sch2_dict = {
'scheduler': self.scheduler_d,
'interval': 'step',
}
return [self.optim_g, self.optim_d], [sch1_dict, sch2_dict]
@property
def input_types(self):
return {
"spec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
}
@property
def output_types(self):
return {
"audio": NeuralType(('B', 'S', 'T'),
AudioSignal(self.sample_rate)),
}
@typecheck()
def forward(self, *, spec):
"""
Runs the generator, for inputs and outputs see input_types, and output_types
"""
return self.generator(x=spec)
def forward_for_export(self, spec):
return self.generator(x=spec)
@typecheck(
input_types={
"spec": NeuralType(('B', 'C', 'T'), MelSpectrogramType())
},
output_types={"audio": NeuralType(('B', 'T'), AudioSignal())},
)
def convert_spectrogram_to_audio(self,
spec: 'torch.tensor') -> 'torch.tensor':
return self(spec=spec).squeeze(1)
def training_step(self, batch, batch_idx, optimizer_idx):
# if in finetune mode the mels are pre-computed using a
# spectrogram generator
if self.input_as_mel:
audio, audio_len, audio_mel = batch
# else, we compute the mel using the ground truth audio
else:
audio, audio_len = batch
# mel as input for generator
audio_mel, _ = self.audio_to_melspec_precessor(audio, audio_len)
# mel as input for L1 mel loss
audio_trg_mel, _ = self.trg_melspec_fn(audio, audio_len)
audio = audio.unsqueeze(1)
audio_pred = self.generator(x=audio_mel)
audio_pred_mel, _ = self.trg_melspec_fn(audio_pred.squeeze(1),
audio_len)
# train discriminator
self.optim_d.zero_grad()
mpd_score_real, mpd_score_gen, _, _ = self.mpd(
y=audio, y_hat=audio_pred.detach())
loss_disc_mpd, _, _ = self.discriminator_loss(
disc_real_outputs=mpd_score_real,
disc_generated_outputs=mpd_score_gen)
msd_score_real, msd_score_gen, _, _ = self.msd(
y=audio, y_hat=audio_pred.detach())
loss_disc_msd, _, _ = self.discriminator_loss(
disc_real_outputs=msd_score_real,
disc_generated_outputs=msd_score_gen)
loss_d = loss_disc_msd + loss_disc_mpd
self.manual_backward(loss_d)
self.optim_d.step()
# train generator
self.optim_g.zero_grad()
loss_mel = F.l1_loss(audio_pred_mel, audio_trg_mel)
_, mpd_score_gen, fmap_mpd_real, fmap_mpd_gen = self.mpd(
y=audio, y_hat=audio_pred)
_, msd_score_gen, fmap_msd_real, fmap_msd_gen = self.msd(
y=audio, y_hat=audio_pred)
loss_fm_mpd = self.feature_loss(fmap_r=fmap_mpd_real,
fmap_g=fmap_mpd_gen)
loss_fm_msd = self.feature_loss(fmap_r=fmap_msd_real,
fmap_g=fmap_msd_gen)
loss_gen_mpd, _ = self.generator_loss(disc_outputs=mpd_score_gen)
loss_gen_msd, _ = self.generator_loss(disc_outputs=msd_score_gen)
loss_g = loss_gen_msd + loss_gen_mpd + loss_fm_msd + loss_fm_mpd + loss_mel * self.l1_factor
self.manual_backward(loss_g)
self.optim_g.step()
metrics = {
"g_l1_loss": loss_mel,
"g_loss_fm_mpd": loss_fm_mpd,
"g_loss_fm_msd": loss_fm_msd,
"g_loss_gen_mpd": loss_gen_mpd,
"g_loss_gen_msd": loss_gen_msd,
"g_loss": loss_g,
"d_loss_mpd": loss_disc_mpd,
"d_loss_msd": loss_disc_msd,
"d_loss": loss_d,
"global_step": self.global_step,
"lr": self.optim_g.param_groups[0]['lr'],
}
self.log_dict(metrics, on_step=True, sync_dist=True)
self.log("g_l1_loss",
loss_mel,
prog_bar=True,
logger=False,
sync_dist=True)
def validation_step(self, batch, batch_idx):
if self.input_as_mel:
audio, audio_len, audio_mel = batch
audio_mel_len = [audio_mel.shape[1]] * audio_mel.shape[0]
else:
audio, audio_len = batch
audio_mel, audio_mel_len = self.audio_to_melspec_precessor(
audio, audio_len)
audio_pred = self(spec=audio_mel)
# perform bias denoising
pred_denoised = self._bias_denoise(audio_pred, audio_mel).squeeze(1)
pred_denoised_mel, _ = self.audio_to_melspec_precessor(
pred_denoised, audio_len)
if self.input_as_mel:
gt_mel, gt_mel_len = self.audio_to_melspec_precessor(
audio, audio_len)
audio_pred_mel, _ = self.audio_to_melspec_precessor(
audio_pred.squeeze(1), audio_len)
loss_mel = F.l1_loss(audio_mel, audio_pred_mel)
self.log_dict({"val_loss": loss_mel}, on_epoch=True, sync_dist=True)
# plot audio once per epoch
if batch_idx == 0 and isinstance(self.logger,
WandbLogger) and HAVE_WANDB:
clips = []
specs = []
for i in range(min(5, audio.shape[0])):
clips += [
wandb.Audio(
audio[i, :audio_len[i]].data.cpu().numpy(),
caption=f"real audio {i}",
sample_rate=self.sample_rate,
),
wandb.Audio(
audio_pred[i,
0, :audio_len[i]].data.cpu().numpy().astype(
'float32'),
caption=f"generated audio {i}",
sample_rate=self.sample_rate,
),
wandb.Audio(
pred_denoised[i, :audio_len[i]].data.cpu().numpy(),
caption=f"denoised audio {i}",
sample_rate=self.sample_rate,
),
]
specs += [
wandb.Image(
plot_spectrogram_to_numpy(audio_mel[
i, :, :audio_mel_len[i]].data.cpu().numpy()),
caption=f"input mel {i}",
),
wandb.Image(
plot_spectrogram_to_numpy(audio_pred_mel[
i, :, :audio_mel_len[i]].data.cpu().numpy()),
caption=f"output mel {i}",
),
wandb.Image(
plot_spectrogram_to_numpy(pred_denoised_mel[
i, :, :audio_mel_len[i]].data.cpu().numpy()),
caption=f"denoised mel {i}",
),
]
if self.input_as_mel:
specs += [
wandb.Image(
plot_spectrogram_to_numpy(gt_mel[
i, :, :audio_mel_len[i]].data.cpu().numpy()),
caption=f"gt mel {i}",
),
]
self.logger.experiment.log({"audio": clips, "specs": specs})
def _bias_denoise(self, audio, mel):
def stft(x):
comp = stft_patch(x.squeeze(1),
n_fft=1024,
hop_length=256,
win_length=1024)
real, imag = comp[..., 0], comp[..., 1]
mags = torch.sqrt(real**2 + imag**2)
phase = torch.atan2(imag, real)
return mags, phase
def istft(mags, phase):
comp = torch.stack(
[mags * torch.cos(phase), mags * torch.sin(phase)], dim=-1)
x = torch.istft(comp, n_fft=1024, hop_length=256, win_length=1024)
return x
# create bias tensor
if self.stft_bias is None or self.stft_bias.shape[0] != audio.shape[0]:
audio_bias = self(spec=torch.zeros_like(mel, device=mel.device))
self.stft_bias, _ = stft(audio_bias)
self.stft_bias = self.stft_bias[:, :, 0][:, :, None]
audio_mags, audio_phase = stft(audio)
audio_mags = audio_mags - self.cfg.get("denoise_strength",
0.0025) * self.stft_bias
audio_mags = torch.clamp(audio_mags, 0.0)
audio_denoised = istft(audio_mags, audio_phase).unsqueeze(1)
return audio_denoised
def __setup_dataloader_from_config(self,
cfg,
shuffle_should_be: bool = True,
name: str = "train"):
if "dataset" not in cfg or not isinstance(cfg.dataset, DictConfig):
raise ValueError(f"No dataset for {name}")
if "dataloader_params" not in cfg or not isinstance(
cfg.dataloader_params, DictConfig):
raise ValueError(f"No dataloder_params for {name}")
if shuffle_should_be:
if 'shuffle' not in cfg.dataloader_params:
logging.warning(
f"Shuffle should be set to True for {self}'s {name} dataloader but was not found in its "
"config. Manually setting to True")
with open_dict(cfg["dataloader_params"]):
cfg.dataloader_params.shuffle = True
elif not cfg.dataloader_params.shuffle:
logging.error(
f"The {name} dataloader for {self} has shuffle set to False!!!"
)
elif not shuffle_should_be and cfg.dataloader_params.shuffle:
logging.error(
f"The {name} dataloader for {self} has shuffle set to True!!!")
dataset = instantiate(cfg.dataset)
return torch.utils.data.DataLoader(dataset,
collate_fn=dataset.collate_fn,
**cfg.dataloader_params)
def setup_training_data(self, cfg):
self._train_dl = self.__setup_dataloader_from_config(cfg)
def setup_validation_data(self, cfg):
self._validation_dl = self.__setup_dataloader_from_config(
cfg, shuffle_should_be=False, name="validation")
@classmethod
def list_available_models(cls) -> 'Optional[Dict[str, str]]':
list_of_models = []
model = PretrainedModelInfo(
pretrained_model_name="tts_hifigan",
location=
"https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_hifigan/versions/1.0.0rc1/files/tts_hifigan.nemo",
description=
"This model is trained on LJSpeech audio sampled at 22050Hz and mel spectrograms generated from Tacotron2, TalkNet, and FastPitch. This model has been tested on generating female English voices with an American accent.",
class_=cls,
)
list_of_models.append(model)
return list_of_models
def load_state_dict(self, state_dict, strict=True):
# override load_state_dict to give us some flexibility to be backward-compatible
# with old checkpoints
new_state_dict = {}
num_resblocks = len(self.cfg['generator']['resblock_kernel_sizes'])
for k, v in state_dict.items():
new_k = k
if 'resblocks' in k:
parts = k.split(".")
# only do this is the checkpoint type is older
if len(parts) == 6:
layer = int(parts[2])
new_layer = f"{layer//num_resblocks}.{layer%num_resblocks}"
new_k = f"generator.resblocks.{new_layer}.{'.'.join(parts[3:])}"
new_state_dict[new_k] = v
super().load_state_dict(new_state_dict, strict=strict)
def _prepare_for_export(self, **kwargs):
"""
Override this method to prepare module for export. This is in-place operation.
Base version does common necessary module replacements (Apex etc)
"""
self.generator.remove_weight_norm()
def input_example(self):
"""
Generates input examples for tracing etc.
Returns:
A tuple of input examples.
"""
par = next(self.parameters())
mel = torch.randn((1, self.cfg['preprocessor']['nfilt'], 96),
device=par.device,
dtype=par.dtype)
return mel
class HifiGanModel(Vocoder):
def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None):
if isinstance(cfg, dict):
cfg = OmegaConf.create(cfg)
super().__init__(cfg=cfg, trainer=trainer)
self.audio_to_melspec_precessor = instantiate(cfg.preprocessor)
# use a different melspec extractor because:
# 1. we need to pass grads
# 2. we need remove fmax limitation
self.trg_melspec_fn = instantiate(cfg.preprocessor,
highfreq=None,
use_grads=True)
self.generator = instantiate(cfg.generator)
self.mpd = MultiPeriodDiscriminator()
self.msd = MultiScaleDiscriminator()
self.feature_loss = FeatureMatchingLoss()
self.discriminator_loss = DiscriminatorLoss()
self.generator_loss = GeneratorLoss()
self.sample_rate = self._cfg.preprocessor.sample_rate
def configure_optimizers(self):
self.optim_g = torch.optim.AdamW(
self.generator.parameters(),
self._cfg.optim.lr,
betas=[self._cfg.optim.adam_b1, self._cfg.optim.adam_b2])
self.optim_d = torch.optim.AdamW(
itertools.chain(self.msd.parameters(), self.mpd.parameters()),
self._cfg.optim.lr,
betas=[self._cfg.optim.adam_b1, self._cfg.optim.adam_b2],
)
self.scheduler_g = torch.optim.lr_scheduler.StepLR(
self.optim_g,
step_size=self._cfg.optim.lr_step,
gamma=self._cfg.optim.lr_decay,
)
self.scheduler_d = torch.optim.lr_scheduler.StepLR(
self.optim_d,
step_size=self._cfg.optim.lr_step,
gamma=self._cfg.optim.lr_decay,
)
return [self.optim_g,
self.optim_d], [self.scheduler_g, self.scheduler_d]
@property
def input_types(self):
return {
"spec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
}
@property
def output_types(self):
return {
"audio": NeuralType(('B', 'S', 'T'),
AudioSignal(self.sample_rate)),
}
@typecheck()
def forward(self, *, spec):
"""
Runs the generator, for inputs and outputs see input_types, and output_types
"""
return self.generator(x=spec)
@typecheck(output_types={"audio": NeuralType(('B', 'T'), AudioSignal())})
def convert_spectrogram_to_audio(self,
spec: 'torch.tensor') -> 'torch.tensor':
return self(spec=spec).squeeze(1)
def training_step(self, batch, batch_idx, optimizer_idx):
audio, audio_len = batch
# mel as input for generator
audio_mel, _ = self.audio_to_melspec_precessor(audio, audio_len)
# mel as input for L1 mel loss
audio_trg_mel, _ = self.trg_melspec_fn(audio, audio_len)
audio = audio.unsqueeze(1)
audio_pred = self.generator(x=audio_mel)
audio_pred_mel, _ = self.trg_melspec_fn(audio_pred.squeeze(1),
audio_len)
# train discriminator
self.optim_d.zero_grad()
mpd_score_real, mpd_score_gen, _, _ = self.mpd(
y=audio, y_hat=audio_pred.detach())
loss_disc_mpd, _, _ = self.discriminator_loss(
disc_real_outputs=mpd_score_real,
disc_generated_outputs=mpd_score_gen)
msd_score_real, msd_score_gen, _, _ = self.msd(
y=audio, y_hat=audio_pred.detach())
loss_disc_msd, _, _ = self.discriminator_loss(
disc_real_outputs=msd_score_real,
disc_generated_outputs=msd_score_gen)
loss_d = loss_disc_msd + loss_disc_mpd
self.manual_backward(loss_d, self.optim_d)
self.optim_d.step()
# train generator
self.optim_g.zero_grad()
loss_mel = F.l1_loss(audio_pred_mel, audio_trg_mel) * 45
_, mpd_score_gen, fmap_mpd_real, fmap_mpd_gen = self.mpd(
y=audio, y_hat=audio_pred)
_, msd_score_gen, fmap_msd_real, fmap_msd_gen = self.msd(
y=audio, y_hat=audio_pred)
loss_fm_mpd = self.feature_loss(fmap_r=fmap_mpd_real,
fmap_g=fmap_mpd_gen)
loss_fm_msd = self.feature_loss(fmap_r=fmap_msd_real,
fmap_g=fmap_msd_gen)
loss_gen_mpd, _ = self.generator_loss(disc_outputs=mpd_score_gen)
loss_gen_msd, _ = self.generator_loss(disc_outputs=msd_score_gen)
loss_g = loss_gen_msd + loss_gen_mpd + loss_fm_msd + loss_fm_mpd + loss_mel
self.manual_backward(loss_g, self.optim_g)
self.optim_g.step()
metrics = {
"g_l1_loss": loss_mel,
"g_loss_fm_mpd": loss_fm_mpd,
"g_loss_fm_msd": loss_fm_msd,
"g_loss_gen_mpd": loss_gen_mpd,
"g_loss_gen_msd": loss_gen_msd,
"g_loss": loss_g,
"d_loss_mpd": loss_disc_mpd,
"d_loss_msd": loss_disc_msd,
"d_loss": loss_d,
"global_step": self.global_step,
"lr": self.optim_g.param_groups[0]['lr'],
}
self.log_dict(metrics, on_step=False, on_epoch=True, sync_dist=True)
def validation_step(self, batch, batch_idx):
audio, audio_len = batch
audio_mel, audio_mel_len = self.audio_to_melspec_precessor(
audio, audio_len)
audio_pred = self(spec=audio_mel)
audio_pred_mel, _ = self.audio_to_melspec_precessor(
audio_pred.squeeze(1), audio_len)
loss_mel = F.l1_loss(audio_mel, audio_pred_mel)
self.log("val_loss", loss_mel, prog_bar=True, sync_dist=True)
# plot audio once per epoch
if batch_idx == 0 and isinstance(self.logger, WandbLogger):
clips = []
specs = []
for i in range(min(5, audio.shape[0])):
clips += [
wandb.Audio(
audio[i, :audio_len[i]].data.cpu().numpy(),
caption=f"real audio {i}",
sample_rate=self.sample_rate,
),
wandb.Audio(
audio_pred[i,
0, :audio_len[i]].data.cpu().numpy().astype(
'float32'),
caption=f"generated audio {i}",
sample_rate=self.sample_rate,
),
]
specs += [
wandb.Image(
plot_spectrogram_to_numpy(audio_mel[
i, :, :audio_mel_len[i]].data.cpu().numpy()),
caption=f"real audio {i}",
),
wandb.Image(
plot_spectrogram_to_numpy(audio_pred_mel[
i, :, :audio_mel_len[i]].data.cpu().numpy()),
caption=f"generated audio {i}",
),
]
self.logger.experiment.log({
"audio": clips,
"specs": specs
},
commit=False)
def __setup_dataloader_from_config(self,
cfg,
shuffle_should_be: bool = True,
name: str = "train"):
if "dataset" not in cfg or not isinstance(cfg.dataset, DictConfig):
raise ValueError(f"No dataset for {name}")
if "dataloader_params" not in cfg or not isinstance(
cfg.dataloader_params, DictConfig):
raise ValueError(f"No dataloder_params for {name}")
if shuffle_should_be:
if 'shuffle' not in cfg.dataloader_params:
logging.warning(
f"Shuffle should be set to True for {self}'s {name} dataloader but was not found in its "
"config. Manually setting to True")
with open_dict(cfg["dataloader_params"]):
cfg.dataloader_params.shuffle = True
elif not cfg.dataloader_params.shuffle:
logging.error(
f"The {name} dataloader for {self} has shuffle set to False!!!"
)
elif not shuffle_should_be and cfg.dataloader_params.shuffle:
logging.error(
f"The {name} dataloader for {self} has shuffle set to True!!!")
dataset = instantiate(cfg.dataset)
return torch.utils.data.DataLoader(dataset,
collate_fn=dataset.collate_fn,
**cfg.dataloader_params)
def setup_training_data(self, cfg):
self._train_dl = self.__setup_dataloader_from_config(cfg)
def setup_validation_data(self, cfg):
self._validation_dl = self.__setup_dataloader_from_config(
cfg, shuffle_should_be=False, name="validation")
@classmethod
def list_available_models(cls) -> 'Optional[Dict[str, str]]':
# TODO
pass