class WaveGlowModel(GlowVocoder, Exportable):
"""Waveglow model used to convert betweeen spectrograms and audio"""
def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None):
if isinstance(cfg, dict):
cfg = OmegaConf.create(cfg)
super().__init__(cfg=cfg, trainer=trainer)
schema = OmegaConf.structured(WaveglowConfig)
# 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.sigma = self._cfg.sigma
self.audio_to_melspec_precessor = instantiate(self._cfg.preprocessor)
self.waveglow = instantiate(self._cfg.waveglow)
self.loss = WaveGlowLoss()
@GlowVocoder.mode.setter
def mode(self, new_mode):
if new_mode == OperationMode.training:
self.train()
else:
self.eval()
self._mode = new_mode
self.waveglow.mode = new_mode
@property
def input_types(self):
return {
"audio": NeuralType(('B', 'T'), AudioSignal()),
"audio_len": NeuralType(('B'), LengthsType()),
"run_inverse": NeuralType(optional=True),
}
@property
def output_types(self):
if self.mode == OperationMode.training or self.mode == OperationMode.validation:
output_dict = {
"pred_normal_dist": NeuralType(('B', 'flowgroup', 'T'), NormalDistributionSamplesType()),
"log_s_list": [NeuralType(('B', 'flowgroup', 'T'), VoidType())], # TODO: Figure out a good typing
"log_det_W_list": [NeuralType(elements_type=LogDeterminantType())],
}
if self.mode == OperationMode.validation:
output_dict["audio_pred"] = NeuralType(('B', 'T'), AudioSignal())
output_dict["spec"] = NeuralType(('B', 'T', 'D'), MelSpectrogramType())
output_dict["spec_len"] = NeuralType(('B'), LengthsType())
return output_dict
return {
"audio_pred": NeuralType(('B', 'T'), AudioSignal()),
}
@typecheck()
def forward(self, *, audio, audio_len, run_inverse=True):
if self.mode != self.waveglow.mode:
raise ValueError(
f"WaveGlowModel's mode {self.mode} does not match WaveGlowModule's mode {self.waveglow.mode}"
)
spec, spec_len = self.audio_to_melspec_precessor(audio, audio_len)
tensors = self.waveglow(spec=spec, audio=audio, run_inverse=run_inverse, sigma=self.sigma)
if self.mode == OperationMode.training:
return tensors[:-1] # z, log_s_list, log_det_W_list
elif self.mode == OperationMode.validation:
z, log_s_list, log_det_W_list, audio_pred = tensors
return z, log_s_list, log_det_W_list, audio_pred, spec, spec_len
return tensors # audio_pred
@typecheck(
input_types={
"spec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
"sigma": NeuralType(optional=True),
"denoise": NeuralType(optional=True),
"denoiser_strength": NeuralType(optional=True),
},
output_types={"audio": NeuralType(('B', 'T'), AudioSignal())},
)
def convert_spectrogram_to_audio(
self, spec: torch.Tensor, sigma: float = 1.0, denoise: bool = True, denoiser_strength: float = 0.01
) -> torch.Tensor:
with self.nemo_infer():
self.waveglow.remove_weightnorm()
audio = self.waveglow(
spec=spec.to(self.waveglow.upsample.weight.dtype), run_inverse=True, audio=None, sigma=sigma
)
if denoise:
audio = self.denoise(audio, denoiser_strength)
return audio
def training_step(self, batch, batch_idx):
self.mode = OperationMode.training
audio, audio_len = batch
z, log_s_list, log_det_W_list = self(audio=audio, audio_len=audio_len, run_inverse=False)
loss = self.loss(z=z, log_s_list=log_s_list, log_det_W_list=log_det_W_list, sigma=self.sigma)
output = {
'loss': loss,
'progress_bar': {'training_loss': loss},
'log': {'loss': loss},
}
return output
def validation_step(self, batch, batch_idx):
self.mode = OperationMode.validation
audio, audio_len = batch
z, log_s_list, log_det_W_list, audio_pred, spec, spec_len = self(
audio=audio, audio_len=audio_len, run_inverse=(batch_idx == 0)
)
loss = self.loss(z=z, log_s_list=log_s_list, log_det_W_list=log_det_W_list, sigma=self.sigma)
return {
"val_loss": loss,
"audio_pred": audio_pred,
"mel_target": spec,
"mel_len": spec_len,
}
def validation_epoch_end(self, outputs):
if self.logger is not None and self.logger.experiment is not 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
waveglow_log_to_tb_func(
tb_logger,
outputs[0].values(),
self.global_step,
tag="eval",
mel_fb=self.audio_to_melspec_precessor.fb,
)
avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
self.log('val_loss', avg_loss)
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) -> '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="WaveGlow-22050Hz",
location="https://api.ngc.nvidia.com/v2/models/nvidia/nemottsmodels/versions/1.0.0a5/files/WaveGlow-22050Hz.nemo",
description="This model is trained on LJSpeech sampled at 22050Hz, and can be used as an universal vocoder.",
class_=cls,
)
list_of_models.append(model)
return list_of_models
@property
def input_module(self):
return self.waveglow
@property
def output_module(self):
return self.waveglow
def _prepare_for_export(self):
self.update_bias_spect()
self.waveglow._prepare_for_export()
def forward_for_export(self, spec, z=None):
return self.waveglow(spec, z)
def output_types(self):
return {"mel_spec": NeuralType(('B', 'T', 'C'), MelSpectrogramType())}
def output_types(self):
return {
"out": NeuralType(('B', 'T', 'D'), EncodedRepresentation()),
"mask": NeuralType(('B', 'T', 'D'), MaskType()),
}
def input_types(self):
return {
"enc": NeuralType(('B', 'T', 'D'), EncodedRepresentation()),
"enc_mask": NeuralType(('B', 'T', 1), TokenDurationType()),
}
def input_types(self):
return {
"audio": NeuralType(('B', 'T'), AudioSignal()),
"audio_len": NeuralType(('B'), LengthsType()),
}
def input_types(self):
return {
"spect_predicted": NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
"spect_tgt": NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
}
def output_types(self):
return {
"loss": NeuralType(elements_type=LossType()),
}
class MixerTTSModel(SpectrogramGenerator, Exportable):
"""Mixer-TTS and Mixer-TTS-X models (https://arxiv.org/abs/2110.03584) that is used to generate mel spectrogram from text."""
def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None):
# Convert to Hydra 1.0 compatible DictConfig
cfg = model_utils.convert_model_config_to_dict_config(cfg)
cfg = model_utils.maybe_update_config_version(cfg)
# Setup normalizer
self.normalizer = None
self.text_normalizer_call = None
self.text_normalizer_call_kwargs = {}
self._setup_normalizer(cfg)
# Setup tokenizer
self.tokenizer = None
self._setup_tokenizer(cfg)
assert self.tokenizer is not None
num_tokens = len(self.tokenizer.tokens)
self.tokenizer_pad = self.tokenizer.pad
self.tokenizer_unk = self.tokenizer.oov
super().__init__(cfg=cfg, trainer=trainer)
self.pitch_loss_scale = cfg.pitch_loss_scale
self.durs_loss_scale = cfg.durs_loss_scale
self.mel_loss_scale = cfg.mel_loss_scale
self.aligner = instantiate(cfg.alignment_module)
self.forward_sum_loss = ForwardSumLoss()
self.bin_loss = BinLoss()
self.add_bin_loss = False
self.bin_loss_scale = 0.0
self.bin_loss_start_ratio = cfg.bin_loss_start_ratio
self.bin_loss_warmup_epochs = cfg.bin_loss_warmup_epochs
self.cond_on_lm_embeddings = cfg.get("cond_on_lm_embeddings", False)
if self.cond_on_lm_embeddings:
self.lm_padding_value = (self._train_dl.dataset.lm_padding_value
if self._train_dl is not None else
self._get_lm_padding_value(cfg.lm_model))
self.lm_embeddings = self._get_lm_embeddings(cfg.lm_model)
self.lm_embeddings.weight.requires_grad = False
self.self_attention_module = instantiate(
cfg.self_attention_module,
n_lm_tokens_channels=self.lm_embeddings.weight.shape[1])
self.encoder = instantiate(cfg.encoder,
num_tokens=num_tokens,
padding_idx=self.tokenizer_pad)
self.symbol_emb = self.encoder.to_embed
self.duration_predictor = instantiate(cfg.duration_predictor)
self.pitch_mean, self.pitch_std = float(cfg.pitch_mean), float(
cfg.pitch_std)
self.pitch_predictor = instantiate(cfg.pitch_predictor)
self.pitch_emb = instantiate(cfg.pitch_emb)
self.preprocessor = instantiate(cfg.preprocessor)
self.decoder = instantiate(cfg.decoder)
self.proj = nn.Linear(self.decoder.d_model, cfg.n_mel_channels)
def _setup_normalizer(self, cfg):
if "text_normalizer" in cfg:
normalizer_kwargs = {}
if "whitelist" in cfg.text_normalizer:
normalizer_kwargs["whitelist"] = self.register_artifact(
'text_normalizer.whitelist', cfg.text_normalizer.whitelist)
self.normalizer = instantiate(cfg.text_normalizer,
**normalizer_kwargs)
self.text_normalizer_call = self.normalizer.normalize
if "text_normalizer_call_kwargs" in cfg:
self.text_normalizer_call_kwargs = cfg.text_normalizer_call_kwargs
def _setup_tokenizer(self, cfg):
text_tokenizer_kwargs = {}
if "g2p" in cfg.text_tokenizer:
g2p_kwargs = {}
if "phoneme_dict" in cfg.text_tokenizer.g2p:
g2p_kwargs["phoneme_dict"] = self.register_artifact(
'text_tokenizer.g2p.phoneme_dict',
cfg.text_tokenizer.g2p.phoneme_dict,
)
if "heteronyms" in cfg.text_tokenizer.g2p:
g2p_kwargs["heteronyms"] = self.register_artifact(
'text_tokenizer.g2p.heteronyms',
cfg.text_tokenizer.g2p.heteronyms,
)
text_tokenizer_kwargs["g2p"] = instantiate(cfg.text_tokenizer.g2p,
**g2p_kwargs)
self.tokenizer = instantiate(cfg.text_tokenizer,
**text_tokenizer_kwargs)
def _get_lm_model_tokenizer(self, lm_model="albert"):
if getattr(self, "_lm_model_tokenizer", None) is not None:
return self._lm_model_tokenizer
if self._train_dl is not None and self._train_dl.dataset is not None:
self._lm_model_tokenizer = self._train_dl.dataset.lm_model_tokenizer
if lm_model == "albert":
self._lm_model_tokenizer = AlbertTokenizer.from_pretrained(
'albert-base-v2')
else:
raise NotImplementedError(
f"{lm_model} lm model is not supported. Only albert is supported at this moment."
)
return self._lm_model_tokenizer
def _get_lm_embeddings(self, lm_model="albert"):
if lm_model == "albert":
return transformers.AlbertModel.from_pretrained(
'albert-base-v2').embeddings.word_embeddings
else:
raise NotImplementedError(
f"{lm_model} lm model is not supported. Only albert is supported at this moment."
)
def _get_lm_padding_value(self, lm_model="albert"):
if lm_model == "albert":
return transformers.AlbertTokenizer.from_pretrained(
'albert-base-v2')._convert_token_to_id('<pad>')
else:
raise NotImplementedError(
f"{lm_model} lm model is not supported. Only albert is supported at this moment."
)
def _metrics(
self,
true_durs,
true_text_len,
pred_durs,
true_pitch,
pred_pitch,
true_spect=None,
pred_spect=None,
true_spect_len=None,
attn_logprob=None,
attn_soft=None,
attn_hard=None,
attn_hard_dur=None,
):
text_mask = get_mask_from_lengths(true_text_len)
mel_mask = get_mask_from_lengths(true_spect_len)
loss = 0.0
# Dur loss and metrics
durs_loss = F.mse_loss(pred_durs, (true_durs + 1).float().log(),
reduction='none')
durs_loss = durs_loss * text_mask.float()
durs_loss = durs_loss.sum() / text_mask.sum()
durs_pred = pred_durs.exp() - 1
durs_pred = torch.clamp_min(durs_pred, min=0)
durs_pred = durs_pred.round().long()
acc = ((true_durs == durs_pred) *
text_mask).sum().float() / text_mask.sum() * 100
acc_dist_1 = (((true_durs - durs_pred).abs() <= 1) *
text_mask).sum().float() / text_mask.sum() * 100
acc_dist_3 = (((true_durs - durs_pred).abs() <= 3) *
text_mask).sum().float() / text_mask.sum() * 100
pred_spect = pred_spect.transpose(1, 2)
# Mel loss
mel_loss = F.mse_loss(pred_spect, true_spect,
reduction='none').mean(dim=-2)
mel_loss = mel_loss * mel_mask.float()
mel_loss = mel_loss.sum() / mel_mask.sum()
loss = loss + self.durs_loss_scale * durs_loss + self.mel_loss_scale * mel_loss
# Aligner loss
bin_loss, ctc_loss = None, None
ctc_loss = self.forward_sum_loss(attn_logprob=attn_logprob,
in_lens=true_text_len,
out_lens=true_spect_len)
loss = loss + ctc_loss
if self.add_bin_loss:
bin_loss = self.bin_loss(hard_attention=attn_hard,
soft_attention=attn_soft)
loss = loss + self.bin_loss_scale * bin_loss
true_avg_pitch = average_pitch(true_pitch.unsqueeze(1),
attn_hard_dur).squeeze(1)
# Pitch loss
pitch_loss = F.mse_loss(pred_pitch, true_avg_pitch,
reduction='none') # noqa
pitch_loss = (pitch_loss * text_mask).sum() / text_mask.sum()
loss = loss + self.pitch_loss_scale * pitch_loss
return loss, durs_loss, acc, acc_dist_1, acc_dist_3, pitch_loss, mel_loss, ctc_loss, bin_loss
@torch.jit.unused
def run_aligner(self, text, text_len, text_mask, spect, spect_len,
attn_prior):
text_emb = self.symbol_emb(text)
attn_soft, attn_logprob = self.aligner(
spect,
text_emb.permute(0, 2, 1),
mask=text_mask == 0,
attn_prior=attn_prior,
)
attn_hard = binarize_attention_parallel(attn_soft, text_len, spect_len)
attn_hard_dur = attn_hard.sum(2)[:, 0, :]
assert torch.all(torch.eq(attn_hard_dur.sum(dim=1), spect_len))
return attn_soft, attn_logprob, attn_hard, attn_hard_dur
@typecheck(
input_types={
"text":
NeuralType(('B', 'T_text'), TokenIndex()),
"text_len":
NeuralType(('B', ), LengthsType()),
"pitch":
NeuralType(('B', 'T_audio'), RegressionValuesType(),
optional=True),
"spect":
NeuralType(('B', 'D', 'T_spec'),
MelSpectrogramType(),
optional=True),
"spect_len":
NeuralType(('B', ), LengthsType(), optional=True),
"attn_prior":
NeuralType(('B', 'T_spec', 'T_text'), ProbsType(), optional=True),
"lm_tokens":
NeuralType(('B', 'T_lm_tokens'), TokenIndex(), optional=True),
},
output_types={
"pred_spect":
NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType()),
"durs_predicted":
NeuralType(('B', 'T_text'), TokenDurationType()),
"log_durs_predicted":
NeuralType(('B', 'T_text'), TokenLogDurationType()),
"pitch_predicted":
NeuralType(('B', 'T_text'), RegressionValuesType()),
"attn_soft":
NeuralType(('B', 'S', 'T_spec', 'T_text'), ProbsType()),
"attn_logprob":
NeuralType(('B', 'S', 'T_spec', 'T_text'), LogprobsType()),
"attn_hard":
NeuralType(('B', 'S', 'T_spec', 'T_text'), ProbsType()),
"attn_hard_dur":
NeuralType(('B', 'T_text'), TokenDurationType()),
},
)
def forward(self,
text,
text_len,
pitch=None,
spect=None,
spect_len=None,
attn_prior=None,
lm_tokens=None):
if self.training:
assert pitch is not None
text_mask = get_mask_from_lengths(text_len).unsqueeze(2)
enc_out, enc_mask = self.encoder(text, text_mask)
# Aligner
attn_soft, attn_logprob, attn_hard, attn_hard_dur = None, None, None, None
if spect is not None:
attn_soft, attn_logprob, attn_hard, attn_hard_dur = self.run_aligner(
text, text_len, text_mask, spect, spect_len, attn_prior)
if self.cond_on_lm_embeddings:
lm_emb = self.lm_embeddings(lm_tokens)
lm_features = self.self_attention_module(
enc_out,
lm_emb,
lm_emb,
q_mask=enc_mask.squeeze(2),
kv_mask=lm_tokens != self.lm_padding_value)
# Duration predictor
log_durs_predicted = self.duration_predictor(enc_out, enc_mask)
durs_predicted = torch.clamp(log_durs_predicted.exp() - 1, 0)
# Pitch predictor
pitch_predicted = self.pitch_predictor(enc_out, enc_mask)
# Avg pitch, add pitch_emb
if not self.training:
if pitch is not None:
pitch = average_pitch(pitch.unsqueeze(1),
attn_hard_dur).squeeze(1)
pitch_emb = self.pitch_emb(pitch.unsqueeze(1))
else:
pitch_emb = self.pitch_emb(pitch_predicted.unsqueeze(1))
else:
pitch = average_pitch(pitch.unsqueeze(1), attn_hard_dur).squeeze(1)
pitch_emb = self.pitch_emb(pitch.unsqueeze(1))
enc_out = enc_out + pitch_emb.transpose(1, 2)
if self.cond_on_lm_embeddings:
enc_out = enc_out + lm_features
# Regulate length
len_regulated_enc_out, dec_lens = regulate_len(attn_hard_dur, enc_out)
dec_out, dec_lens = self.decoder(
len_regulated_enc_out,
get_mask_from_lengths(dec_lens).unsqueeze(2))
pred_spect = self.proj(dec_out)
return (
pred_spect,
durs_predicted,
log_durs_predicted,
pitch_predicted,
attn_soft,
attn_logprob,
attn_hard,
attn_hard_dur,
)
def infer(
self,
text,
text_len=None,
text_mask=None,
spect=None,
spect_len=None,
attn_prior=None,
use_gt_durs=False,
lm_tokens=None,
pitch=None,
):
if text_mask is None:
text_mask = get_mask_from_lengths(text_len).unsqueeze(2)
enc_out, enc_mask = self.encoder(text, text_mask)
# Aligner
attn_hard_dur = None
if use_gt_durs:
attn_soft, attn_logprob, attn_hard, attn_hard_dur = self.run_aligner(
text, text_len, text_mask, spect, spect_len, attn_prior)
if self.cond_on_lm_embeddings:
lm_emb = self.lm_embeddings(lm_tokens)
lm_features = self.self_attention_module(
enc_out,
lm_emb,
lm_emb,
q_mask=enc_mask.squeeze(2),
kv_mask=lm_tokens != self.lm_padding_value)
# Duration predictor
log_durs_predicted = self.duration_predictor(enc_out, enc_mask)
durs_predicted = torch.clamp(log_durs_predicted.exp() - 1, 0)
# Avg pitch, pitch predictor
if use_gt_durs and pitch is not None:
pitch = average_pitch(pitch.unsqueeze(1), attn_hard_dur).squeeze(1)
pitch_emb = self.pitch_emb(pitch.unsqueeze(1))
else:
pitch_predicted = self.pitch_predictor(enc_out, enc_mask)
pitch_emb = self.pitch_emb(pitch_predicted.unsqueeze(1))
# Add pitch emb
enc_out = enc_out + pitch_emb.transpose(1, 2)
if self.cond_on_lm_embeddings:
enc_out = enc_out + lm_features
if use_gt_durs:
if attn_hard_dur is not None:
len_regulated_enc_out, dec_lens = regulate_len(
attn_hard_dur, enc_out)
else:
raise NotImplementedError
else:
len_regulated_enc_out, dec_lens = regulate_len(
durs_predicted, enc_out)
dec_out, _ = self.decoder(len_regulated_enc_out,
get_mask_from_lengths(dec_lens).unsqueeze(2))
pred_spect = self.proj(dec_out)
return pred_spect
def on_train_epoch_start(self):
bin_loss_start_epoch = np.ceil(self.bin_loss_start_ratio *
self._trainer.max_epochs)
# Add bin loss when current_epoch >= bin_start_epoch
if not self.add_bin_loss and self.current_epoch >= bin_loss_start_epoch:
logging.info(
f"Using hard attentions after epoch: {self.current_epoch}")
self.add_bin_loss = True
if self.add_bin_loss:
self.bin_loss_scale = min(
(self.current_epoch - bin_loss_start_epoch) /
self.bin_loss_warmup_epochs, 1.0)
def training_step(self, batch, batch_idx):
attn_prior, lm_tokens = None, None
if self.cond_on_lm_embeddings:
audio, audio_len, text, text_len, attn_prior, pitch, _, lm_tokens = batch
else:
audio, audio_len, text, text_len, attn_prior, pitch, _ = batch
spect, spect_len = self.preprocessor(input_signal=audio,
length=audio_len)
# pitch normalization
zero_pitch_idx = pitch == 0
pitch = (pitch - self.pitch_mean) / self.pitch_std
pitch[zero_pitch_idx] = 0.0
(
pred_spect,
_,
pred_log_durs,
pred_pitch,
attn_soft,
attn_logprob,
attn_hard,
attn_hard_dur,
) = self(
text=text,
text_len=text_len,
pitch=pitch,
spect=spect,
spect_len=spect_len,
attn_prior=attn_prior,
lm_tokens=lm_tokens,
)
(
loss,
durs_loss,
acc,
acc_dist_1,
acc_dist_3,
pitch_loss,
mel_loss,
ctc_loss,
bin_loss,
) = self._metrics(
pred_durs=pred_log_durs,
pred_pitch=pred_pitch,
true_durs=attn_hard_dur,
true_text_len=text_len,
true_pitch=pitch,
true_spect=spect,
pred_spect=pred_spect,
true_spect_len=spect_len,
attn_logprob=attn_logprob,
attn_soft=attn_soft,
attn_hard=attn_hard,
attn_hard_dur=attn_hard_dur,
)
train_log = {
'train_loss':
loss,
'train_durs_loss':
durs_loss,
'train_pitch_loss':
torch.tensor(1.0).to(durs_loss.device)
if pitch_loss is None else pitch_loss,
'train_mel_loss':
mel_loss,
'train_durs_acc':
acc,
'train_durs_acc_dist_3':
acc_dist_3,
'train_ctc_loss':
torch.tensor(1.0).to(durs_loss.device)
if ctc_loss is None else ctc_loss,
'train_bin_loss':
torch.tensor(1.0).to(durs_loss.device)
if bin_loss is None else bin_loss,
}
return {
'loss': loss,
'progress_bar': {k: v.detach()
for k, v in train_log.items()},
'log': train_log
}
def validation_step(self, batch, batch_idx):
attn_prior, lm_tokens = None, None
if self.cond_on_lm_embeddings:
audio, audio_len, text, text_len, attn_prior, pitch, _, lm_tokens = batch
else:
audio, audio_len, text, text_len, attn_prior, pitch, _ = batch
spect, spect_len = self.preprocessor(input_signal=audio,
length=audio_len)
# pitch normalization
zero_pitch_idx = pitch == 0
pitch = (pitch - self.pitch_mean) / self.pitch_std
pitch[zero_pitch_idx] = 0.0
(
pred_spect,
_,
pred_log_durs,
pred_pitch,
attn_soft,
attn_logprob,
attn_hard,
attn_hard_dur,
) = self(
text=text,
text_len=text_len,
pitch=pitch,
spect=spect,
spect_len=spect_len,
attn_prior=attn_prior,
lm_tokens=lm_tokens,
)
(
loss,
durs_loss,
acc,
acc_dist_1,
acc_dist_3,
pitch_loss,
mel_loss,
ctc_loss,
bin_loss,
) = self._metrics(
pred_durs=pred_log_durs,
pred_pitch=pred_pitch,
true_durs=attn_hard_dur,
true_text_len=text_len,
true_pitch=pitch,
true_spect=spect,
pred_spect=pred_spect,
true_spect_len=spect_len,
attn_logprob=attn_logprob,
attn_soft=attn_soft,
attn_hard=attn_hard,
attn_hard_dur=attn_hard_dur,
)
# without ground truth internal features except for durations
pred_spect, _, pred_log_durs, pred_pitch, attn_soft, attn_logprob, attn_hard, attn_hard_dur = self(
text=text,
text_len=text_len,
pitch=None,
spect=spect,
spect_len=spect_len,
attn_prior=attn_prior,
lm_tokens=lm_tokens,
)
*_, with_pred_features_mel_loss, _, _ = self._metrics(
pred_durs=pred_log_durs,
pred_pitch=pred_pitch,
true_durs=attn_hard_dur,
true_text_len=text_len,
true_pitch=pitch,
true_spect=spect,
pred_spect=pred_spect,
true_spect_len=spect_len,
attn_logprob=attn_logprob,
attn_soft=attn_soft,
attn_hard=attn_hard,
attn_hard_dur=attn_hard_dur,
)
val_log = {
'val_loss':
loss,
'val_durs_loss':
durs_loss,
'val_pitch_loss':
torch.tensor(1.0).to(durs_loss.device)
if pitch_loss is None else pitch_loss,
'val_mel_loss':
mel_loss,
'val_with_pred_features_mel_loss':
with_pred_features_mel_loss,
'val_durs_acc':
acc,
'val_durs_acc_dist_3':
acc_dist_3,
'val_ctc_loss':
torch.tensor(1.0).to(durs_loss.device)
if ctc_loss is None else ctc_loss,
'val_bin_loss':
torch.tensor(1.0).to(durs_loss.device)
if bin_loss is None else bin_loss,
}
self.log_dict(val_log,
prog_bar=False,
on_epoch=True,
logger=True,
sync_dist=True)
if batch_idx == 0 and self.current_epoch % 5 == 0 and isinstance(
self.logger, WandbLogger):
specs = []
pitches = []
for i in range(min(3, spect.shape[0])):
specs += [
wandb.Image(
plot_spectrogram_to_numpy(
spect[i, :, :spect_len[i]].data.cpu().numpy()),
caption=f"gt mel {i}",
),
wandb.Image(
plot_spectrogram_to_numpy(
pred_spect.transpose(
1, 2)[i, :, :spect_len[i]].data.cpu().numpy()),
caption=f"pred mel {i}",
),
]
pitches += [
wandb.Image(
plot_pitch_to_numpy(
average_pitch(pitch.unsqueeze(1),
attn_hard_dur).squeeze(1)
[i, :text_len[i]].data.cpu().numpy(),
ylim_range=[-2.5, 2.5],
),
caption=f"gt pitch {i}",
),
]
pitches += [
wandb.Image(
plot_pitch_to_numpy(
pred_pitch[i, :text_len[i]].data.cpu().numpy(),
ylim_range=[-2.5, 2.5]),
caption=f"pred pitch {i}",
),
]
self.logger.experiment.log({"specs": specs, "pitches": pitches})
@typecheck(
input_types={
"tokens":
NeuralType(('B', 'T_text'), TokenIndex(), optional=True),
"tokens_len":
NeuralType(('B'), LengthsType(), optional=True),
"lm_tokens":
NeuralType(('B', 'T_lm_tokens'), TokenIndex(), optional=True),
"raw_texts": [NeuralType(optional=True)],
"lm_model":
NeuralType(optional=True),
},
output_types={
"spect": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType()),
},
)
def generate_spectrogram(
self,
tokens: Optional[torch.Tensor] = None,
tokens_len: Optional[torch.Tensor] = None,
lm_tokens: Optional[torch.Tensor] = None,
raw_texts: Optional[List[str]] = None,
norm_text_for_lm_model: bool = True,
lm_model: str = "albert",
):
if tokens is not None:
if tokens_len is None:
# It is assumed that padding is consecutive and only at the end
tokens_len = (tokens != self.tokenizer.pad).sum(dim=-1)
else:
if raw_texts is None:
raise ValueError(
"raw_texts must be specified if tokens is None")
t_seqs = [self.tokenizer(t) for t in raw_texts]
tokens = torch.nn.utils.rnn.pad_sequence(
sequences=[
torch.tensor(t, dtype=torch.long, device=self.device)
for t in t_seqs
],
batch_first=True,
padding_value=self.tokenizer.pad,
)
tokens_len = torch.tensor([len(t) for t in t_seqs],
dtype=torch.long,
device=tokens.device)
if self.cond_on_lm_embeddings and lm_tokens is None:
if raw_texts is None:
raise ValueError(
"raw_texts must be specified if lm_tokens is None")
lm_model_tokenizer = self._get_lm_model_tokenizer(lm_model)
lm_padding_value = lm_model_tokenizer._convert_token_to_id('<pad>')
lm_space_value = lm_model_tokenizer._convert_token_to_id('▁')
assert isinstance(self.tokenizer,
EnglishCharsTokenizer) or isinstance(
self.tokenizer, EnglishPhonemesTokenizer)
if norm_text_for_lm_model and self.text_normalizer_call is not None:
raw_texts = [
self.text_normalizer_call(
t, **self.text_normalizer_call_kwargs)
for t in raw_texts
]
preprocess_texts_as_tts_input = [
self.tokenizer.text_preprocessing_func(t) for t in raw_texts
]
lm_tokens_as_ids_list = [
lm_model_tokenizer.encode(t, add_special_tokens=False)
for t in preprocess_texts_as_tts_input
]
if self.tokenizer.pad_with_space:
lm_tokens_as_ids_list = [[lm_space_value] + t +
[lm_space_value]
for t in lm_tokens_as_ids_list]
lm_tokens = torch.full(
(len(lm_tokens_as_ids_list),
max([len(t) for t in lm_tokens_as_ids_list])),
fill_value=lm_padding_value,
device=tokens.device,
)
for i, lm_tokens_i in enumerate(lm_tokens_as_ids_list):
lm_tokens[i, :len(lm_tokens_i)] = torch.tensor(
lm_tokens_i, device=tokens.device)
pred_spect = self.infer(tokens, tokens_len,
lm_tokens=lm_tokens).transpose(1, 2)
return pred_spect
def parse(self, text: str, normalize=True) -> torch.Tensor:
if self.training:
logging.warning("parse() is meant to be called in eval mode.")
if normalize and self.text_normalizer_call is not None:
text = self.text_normalizer_call(
text, **self.text_normalizer_call_kwargs)
eval_phon_mode = contextlib.nullcontext()
if hasattr(self.tokenizer, "set_phone_prob"):
eval_phon_mode = self.tokenizer.set_phone_prob(prob=1.0)
with eval_phon_mode:
tokens = self.tokenizer.encode(text)
return torch.tensor(tokens).long().unsqueeze(0).to(self.device)
def _loader(self, cfg):
try:
_ = cfg.dataset.manifest_filepath
except omegaconf.errors.MissingMandatoryValue:
logging.warning(
"manifest_filepath was skipped. No dataset for this model.")
return None
dataset = instantiate(
cfg.dataset,
text_normalizer=self.normalizer,
text_normalizer_call_kwargs=self.text_normalizer_call_kwargs,
text_tokenizer=self.tokenizer,
)
return torch.utils.data.DataLoader( # noqa
dataset=dataset,
collate_fn=dataset.collate_fn,
**cfg.dataloader_params,
)
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_lj_mixertts",
location=
"https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_lj_mixertts/versions/1.6.0/files/tts_en_lj_mixertts.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, # noqa
)
list_of_models.append(model)
model = PretrainedModelInfo(
pretrained_model_name="tts_en_lj_mixerttsx",
location=
"https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_lj_mixerttsx/versions/1.6.0/files/tts_en_lj_mixerttsx.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, # noqa
)
list_of_models.append(model)
return list_of_models
# Methods for model exportability
@property
def input_types(self):
return {
"text":
NeuralType(('B', 'T_text'), TokenIndex()),
"lm_tokens":
NeuralType(('B', 'T_lm_tokens'), TokenIndex(), optional=True),
}
@property
def output_types(self):
return {
"spect": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType()),
}
def input_example(self, max_text_len=10, max_lm_tokens_len=10):
text = torch.randint(
low=0,
high=len(self.tokenizer.tokens),
size=(1, max_text_len),
device=self.device,
dtype=torch.long,
)
inputs = {'text': text}
if self.cond_on_lm_embeddings:
inputs['lm_tokens'] = torch.randint(
low=0,
high=self.lm_embeddings.weight.shape[0],
size=(1, max_lm_tokens_len),
device=self.device,
dtype=torch.long,
)
return (inputs, )
def forward_for_export(self, text, lm_tokens=None):
text_mask = (text != self.tokenizer_pad).unsqueeze(2)
spect = self.infer(text=text, text_mask=text_mask,
lm_tokens=lm_tokens).transpose(1, 2)
return spect.to(torch.float)
def output_types(self):
return {
"spect": NeuralType(('B', 'D', 'T_spec'), MelSpectrogramType()),
}
class UnivNetModel(Vocoder, Exportable):
"""UnivNet model (https://arxiv.org/abs/2106.07889) that is used to generate audio from mel spectrogram."""
def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None):
# Convert to Hydra 1.0 compatible DictConfig
cfg = model_utils.convert_model_config_to_dict_config(cfg)
cfg = model_utils.maybe_update_config_version(cfg)
super().__init__(cfg=cfg, trainer=trainer)
self.audio_to_melspec_precessor = instantiate(cfg.preprocessor)
# We use separate preprocessor for training, because we need to pass grads and remove pitch fmax limitation
self.trg_melspec_fn = instantiate(cfg.preprocessor,
highfreq=None,
use_grads=True)
self.generator = instantiate(
cfg.generator,
n_mel_channels=cfg.preprocessor.nfilt,
hop_length=cfg.preprocessor.n_window_stride)
self.mpd = MultiPeriodDiscriminator(
cfg.discriminator.mpd,
debug=cfg.debug if "debug" in cfg else False)
self.mrd = MultiResolutionDiscriminator(
cfg.discriminator.mrd,
debug=cfg.debug if "debug" in cfg else False)
self.discriminator_loss = DiscriminatorLoss()
self.generator_loss = GeneratorLoss()
# Reshape MRD resolutions hyperparameter and apply them to MRSTFT loss
self.stft_resolutions = cfg.discriminator.mrd.resolutions
self.fft_sizes = [res[0] for res in self.stft_resolutions]
self.hop_sizes = [res[1] for res in self.stft_resolutions]
self.win_lengths = [res[2] for res in self.stft_resolutions]
self.mrstft_loss = MultiResolutionSTFTLoss(self.fft_sizes,
self.hop_sizes,
self.win_lengths)
self.stft_lamb = cfg.stft_lamb
self.sample_rate = self._cfg.preprocessor.sample_rate
self.stft_bias = None
self.input_as_mel = False
if self._train_dl:
self.input_as_mel = self._train_dl.dataset.load_precomputed_mel
self.automatic_optimization = False
def _get_max_steps(self):
return compute_max_steps(
max_epochs=self._cfg.max_epochs,
accumulate_grad_batches=self.trainer.accumulate_grad_batches,
limit_train_batches=self.trainer.limit_train_batches,
num_workers=get_num_workers(self.trainer),
num_samples=len(self._train_dl.dataset),
batch_size=get_batch_size(self._train_dl),
drop_last=self._train_dl.drop_last,
)
def configure_optimizers(self):
optim_g = instantiate(
self._cfg.optim,
params=self.generator.parameters(),
)
optim_d = instantiate(
self._cfg.optim,
params=itertools.chain(self.mrd.parameters(),
self.mpd.parameters()),
)
return [optim_g, optim_d]
@typecheck()
def forward(self, *, spec):
"""
Runs the generator, for inputs and outputs see input_types, and output_types
"""
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):
if self.input_as_mel:
# Pre-computed spectrograms will be used as input
audio, audio_len, audio_mel = batch
else:
audio, audio_len = batch
audio_mel, _ = self.audio_to_melspec_precessor(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)
optim_g, optim_d = self.optimizers()
# Train discriminator
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)
mrd_score_real, mrd_score_gen, _, _ = self.mrd(
y=audio, y_hat=audio_pred.detach())
loss_disc_mrd, _, _ = self.discriminator_loss(
disc_real_outputs=mrd_score_real,
disc_generated_outputs=mrd_score_gen)
loss_d = loss_disc_mrd + loss_disc_mpd
self.manual_backward(loss_d)
optim_d.step()
# Train generator
optim_g.zero_grad()
loss_sc, loss_mag = self.mrstft_loss(x=audio_pred.squeeze(1),
y=audio.squeeze(1),
input_lengths=audio_len)
loss_sc = torch.stack(loss_sc).mean()
loss_mag = torch.stack(loss_mag).mean()
loss_mrstft = (loss_sc + loss_mag) * self.stft_lamb
_, mpd_score_gen, _, _ = self.mpd(y=audio, y_hat=audio_pred)
_, mrd_score_gen, _, _ = self.mrd(y=audio, y_hat=audio_pred)
loss_gen_mpd, _ = self.generator_loss(disc_outputs=mpd_score_gen)
loss_gen_mrd, _ = self.generator_loss(disc_outputs=mrd_score_gen)
loss_g = loss_gen_mrd + loss_gen_mpd + loss_mrstft
self.manual_backward(loss_g)
optim_g.step()
metrics = {
"g_loss_sc": loss_sc,
"g_loss_mag": loss_mag,
"g_loss_mrstft": loss_mrstft,
"g_loss_gen_mpd": loss_gen_mpd,
"g_loss_gen_mrd": loss_gen_mrd,
"g_loss": loss_g,
"d_loss_mpd": loss_disc_mpd,
"d_loss_mrd": loss_disc_mrd,
"d_loss": loss_d,
"global_step": self.global_step,
"lr": optim_g.param_groups[0]['lr'],
}
self.log_dict(metrics, on_step=True, sync_dist=True)
self.log("g_mrstft_loss",
loss_mrstft,
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 = 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 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")
def setup_test_data(self, cfg):
pass
@classmethod
def list_available_models(cls) -> 'Optional[Dict[str, str]]':
list_of_models = []
model = PretrainedModelInfo(
pretrained_model_name="tts_en_lj_univnet",
location=
"https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_lj_univnet/versions/1.7.0/files/tts_en_lj_univnet.nemo",
description=
"This model is trained on LJSpeech sampled at 22050Hz, and has been tested on generating female English voices with an American accent.",
class_=cls,
)
list_of_models.append(model)
model = PretrainedModelInfo(
pretrained_model_name="tts_en_libritts_univnet",
location=
"https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_libritts_univnet/versions/1.7.0/files/tts_en_libritts_multispeaker_univnet.nemo",
description=
"This model is trained on all LibriTTS training data (train-clean-100, train-clean-360, and train-other-500) sampled at 22050Hz, and has been tested on generating English voices.",
class_=cls,
)
list_of_models.append(model)
return list_of_models
# Methods for model exportability
def _prepare_for_export(self, **kwargs):
if self.generator is not None:
try:
self.generator.remove_weight_norm()
except ValueError:
return
@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)),
}
def input_example(self, max_batch=1, max_dim=256):
"""
Generates input examples for tracing etc.
Returns:
A tuple of input examples.
"""
par = next(self.parameters())
mel = torch.randn(
(max_batch, self.cfg['preprocessor']['nfilt'], max_dim),
device=par.device,
dtype=par.dtype)
return ({'spec': mel}, )
def forward_for_export(self, spec):
"""
Runs the generator, for inputs and outputs see input_types, and output_types
"""
return self.generator(x=spec)
class FastSpeech2Model(SpectrogramGenerator):
"""FastSpeech 2 model used to convert from text (phonemes) to mel-spectrograms."""
def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None):
if isinstance(cfg, dict):
cfg = OmegaConf.create(cfg)
super().__init__(cfg=cfg, trainer=trainer)
schema = OmegaConf.structured(FastSpeech2Config)
# 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.pitch = cfg.add_pitch_predictor
self.energy = cfg.add_energy_predictor
self.duration_coeff = cfg.duration_coeff
self.audio_to_melspec_preprocessor = instantiate(
self._cfg.preprocessor)
self.encoder = instantiate(self._cfg.encoder)
self.mel_decoder = instantiate(self._cfg.decoder)
self.variance_adapter = instantiate(self._cfg.variance_adaptor)
self.loss = L2MelLoss()
self.mseloss = torch.nn.MSELoss()
self.durationloss = DurationLoss()
self.log_train_images = False
# 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']
@typecheck(
input_types={
"text":
NeuralType(('B', 'T'), TokenIndex()),
"text_length":
NeuralType(('B'), LengthsType()),
"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={
"mel_spec":
NeuralType(('B', 'T', 'C'), MelSpectrogramType()),
"log_dur_preds":
NeuralType(('B', 'T'), TokenDurationType(), optional=True),
"pitch_preds":
NeuralType(('B', 'T'), RegressionValuesType(), optional=True),
"energy_preds":
NeuralType(('B', 'T'), RegressionValuesType(), optional=True),
"encoded_text_mask":
NeuralType(('B', 'T', 'D'), MaskType()),
},
)
def forward(self,
*,
text,
text_length,
spec_len=None,
durations=None,
pitch=None,
energies=None):
encoded_text, encoded_text_mask = self.encoder(text=text,
text_length=text_length)
aligned_text, 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,
)
mel = self.mel_decoder(decoder_input=aligned_text, lengths=spec_len)
return mel, log_dur_preds, pitch_preds, energy_preds, encoded_text_mask
def training_step(self, batch, batch_idx):
f, fl, t, tl, durations, pitch, energies = batch
spec, spec_len = self.audio_to_melspec_preprocessor(f, fl)
mel, log_dur_preds, pitch_preds, energy_preds, encoded_text_mask = self(
text=t,
text_length=tl,
spec_len=spec_len,
durations=durations,
pitch=pitch,
energies=energies)
total_loss = self.loss(spec_pred=mel.transpose(1, 2),
spec_target=spec,
spec_target_len=spec_len,
pad_value=-11.52)
self.log(name="train_mel_loss", value=total_loss.clone().detach())
# Duration prediction loss
dur_loss = self.durationloss(log_duration_pred=log_dur_preds,
duration_target=durations.float(),
mask=encoded_text_mask)
dur_loss *= self.duration_coeff
self.log(name="train_dur_loss", value=dur_loss)
total_loss += dur_loss
# Pitch prediction loss
if self.pitch:
pitch_loss = self.mseloss(pitch_preds, pitch)
total_loss += pitch_loss
self.log(name="train_pitch_loss", value=pitch_loss)
# Energy prediction loss
if self.energy:
energy_loss = self.mseloss(energy_preds, energies)
total_loss += energy_loss
self.log(name="train_energy_loss", value=energy_loss)
self.log(name="train_loss", value=total_loss)
return {"loss": total_loss, "outputs": [spec, mel]}
def training_epoch_end(self, outputs):
if self.log_train_images and self.logger is not None and self.logger.experiment is not 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
spec_target, spec_predict = outputs[0]["outputs"]
tb_logger.add_image(
"train_mel_target",
plot_spectrogram_to_numpy(spec_target[0].data.cpu().numpy()),
self.global_step,
dataformats="HWC",
)
spec_predict = spec_predict[0].data.cpu().numpy()
tb_logger.add_image(
"train_mel_predicted",
plot_spectrogram_to_numpy(spec_predict.T),
self.global_step,
dataformats="HWC",
)
self.log_train_images = False
return super().training_epoch_end(outputs)
def validation_step(self, batch, batch_idx):
f, fl, t, tl, _, _, _ = batch
spec, spec_len = self.audio_to_melspec_preprocessor(f, fl)
mel, _, _, _, _ = self(text=t, text_length=tl, spec_len=spec_len)
loss = self.loss(spec_pred=mel.transpose(1, 2),
spec_target=spec,
spec_target_len=spec_len,
pad_value=-11.52)
return {
"val_loss": loss,
"mel_target": spec,
"mel_pred": mel,
}
def validation_epoch_end(self, outputs):
if self.logger is not None and self.logger.experiment is not 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
_, spec_target, spec_predict = outputs[0].values()
tb_logger.add_image(
"val_mel_target",
plot_spectrogram_to_numpy(spec_target[0].data.cpu().numpy()),
self.global_step,
dataformats="HWC",
)
spec_predict = spec_predict[0].data.cpu().numpy()
tb_logger.add_image(
"val_mel_predicted",
plot_spectrogram_to_numpy(spec_predict.T),
self.global_step,
dataformats="HWC",
)
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 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
@typecheck(output_types={
"spect": NeuralType(('B', 'C', 'T'), MelSpectrogramType())
})
def generate_spectrogram(self, tokens: torch.Tensor) -> torch.Tensor:
self.eval()
token_len = torch.tensor([tokens.shape[1]]).to(self.device)
spect, *_ = self(text=tokens, text_length=token_len)
return spect.transpose(1, 2)
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) -> 'List[PretrainedModelInfo]':
"""
This method returns a list of pre-trained models 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_fastspeech2",
location=
"https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_fastspeech_2/versions/1.0.0/files/tts_en_fastspeech2.nemo",
description=
"This model is trained on LJSpeech sampled at 22050Hz, and can be used to generate female English voices with an American accent.",
class_=cls,
aliases=["FastSpeech2-22050Hz"],
)
list_of_models.append(model)
return list_of_models
def output_types(self):
return {
"audio": NeuralType(('B', 'S', 'T'),
AudioSignal(self.sample_rate)),
}
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(debug=cfg.debug if "debug" in cfg else False)
self.msd = MultiScaleDiscriminator(debug=cfg.debug if "debug" in cfg else False)
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()),
)
if hasattr(self._cfg, 'sched'):
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]
else:
return [self.optim_g, self.optim_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(
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):
# 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()
# run schedulers
schedulers = self.lr_schedulers()
if schedulers is not None:
sch1, sch2 = schedulers
sch1.step()
sch2.step()
metrics = {
"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 = 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 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)
"""
if self.generator is not None:
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 ({'spec': mel},)
def forward_for_export(self, spec):
"""
Runs the generator, for inputs and outputs see input_types, and output_types
"""
return self.generator(x=spec)
def input_types(self):
return {
"audio": NeuralType(('B', 'T'), AudioSignal()),
"audio_len": NeuralType(('B'), LengthsType()),
"run_inverse": NeuralType(optional=True),
}
class GlowVocoder(Vocoder):
""" Base class for all Vocoders that use a Glow or reversible Flow-based setup. """
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self._mode = OperationMode.infer
self.bias_spect = None
self.stft = None # Required to be defined in children classes
self.n_mel = None # Required to be defined in children classes
@property
def mode(self):
return self._mode
@contextmanager
def temp_mode(self, mode):
old_mode = self.mode
self.mode = mode
try:
yield
finally:
self.mode = old_mode
@contextmanager
def nemo_infer(self): # Prepend with nemo to avoid any .infer() clashes with lightning or pytorch
with ExitStack() as stack:
stack.enter_context(self.temp_mode(OperationMode.infer))
stack.enter_context(torch.no_grad())
yield
def check_children_attributes(self):
if self.stft is None or self.n_mel is None:
try:
self.stft = self.audio_to_melspec_precessor.stft
self.n_mel = self.audio_to_melspec_precessor.nfilt
except AttributeError:
raise AttributeError(
f"{self} did not have stft and n_mel defined. These two parameters are required for GlowVocoder's "
"methods to work"
)
def update_bias_spect(self):
self.check_children_attributes() # Ensure self.n_mel and self.stft are defined
with self.nemo_infer():
spect = torch.zeros((1, self.n_mel, 88)).to(self.device)
bias_audio = self.convert_spectrogram_to_audio(spec=spect, sigma=0.0, denoise=False)
bias_spect, _ = self.stft.transform(bias_audio)
self.bias_spect = bias_spect[:, :, 0][:, :, None]
@typecheck(
input_types={"audio": NeuralType(('B', 'T'), AudioSignal()), "strength": NeuralType(optional=True)},
output_types={"audio": NeuralType(('B', 'T'), AudioSignal())},
)
def denoise(self, audio: 'torch.tensor', strength: float = 0.01):
self.check_children_attributes() # Ensure self.n_mel and self.stft are defined
if self.bias_spect is None:
self.update_bias_spect()
audio_spect, audio_angles = self.stft.transform(audio)
audio_spect_denoised = audio_spect - self.bias_spect.to(audio.device) * strength
audio_spect_denoised = torch.clamp(audio_spect_denoised, 0.0)
audio_denoised = self.stft.inverse(audio_spect_denoised, audio_angles)
return audio_denoised
def output_types(self):
return {
"spec": NeuralType(('B', 'C', 'D', 'T'), SpectrogramType()),
}
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
def input_types(self):
return {
"mel_spec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
}
def input_types(self):
return {
"log_durs_predicted": NeuralType(('B', 'T'), TokenLogDurationType()),
"durs_tgt": NeuralType(('B', 'T'), TokenDurationType()),
"len": NeuralType(('B'), LengthsType()),
}
def input_types(self):
return {
"token_embedding": NeuralType(('B', 'D', 'T'), EmbeddedTextType()),
"token_len": NeuralType(('B'), LengthsType()),
}
def input_types(self):
return {
"pitch_predicted": NeuralType(('B', 'T'), RegressionValuesType()),
"pitch_tgt": NeuralType(('B', 'T'), RegressionValuesType()),
"len": NeuralType(('B'), LengthsType()),
}
def output_types(self):
return {
"encoder_embedding": NeuralType(('B', 'T', 'D'),
EmbeddedTextType()),
}
class UniGlowModel(Vocoder):
"""Waveglow model used to convert betweeen spectrograms and audio"""
def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None):
if isinstance(cfg, dict):
cfg = OmegaConf.create(cfg)
super().__init__(cfg=cfg, trainer=trainer)
schema = OmegaConf.structured(WaveglowConfig)
# 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.sigma = self._cfg.sigma
self.audio_to_melspec_precessor = instantiate(self._cfg.preprocessor)
self.model = UniGlowModule(
self._cfg.uniglow.n_mel_channels,
self._cfg.uniglow.n_flows,
self._cfg.uniglow.n_group,
self._cfg.uniglow.n_wn_channels,
self._cfg.uniglow.n_wn_layers,
self._cfg.uniglow.wn_kernel_size,
self.get_upsample_factor(),
)
self.mode = OperationMode.infer
self.loss = UniGlowLoss(self._cfg.uniglow.stft_loss_coef)
self.removed_weightnorm = False
@property
def mode(self):
return self._mode
@mode.setter
def mode(self, new_mode):
if new_mode == OperationMode.training:
self.train()
else:
self.eval()
self._mode = new_mode
self.model.mode = new_mode
@property
def input_types(self):
return {
"audio": NeuralType(('B', 'T'), AudioSignal()),
"audio_len": NeuralType(('B'), LengthsType()),
}
@property
def output_types(self):
if self.mode == OperationMode.training or self.mode == OperationMode.validation:
output_dict = {
"pred_normal_dist":
NeuralType(('B', 'flowgroup', 'T'),
NormalDistributionSamplesType()),
"logdet":
NeuralType(elements_type=LogDeterminantType()),
"predicted_audio":
NeuralType(('B', 'T'), AudioSignal()),
}
if self.mode == OperationMode.validation:
output_dict["spec"] = NeuralType(('B', 'T', 'D'),
MelSpectrogramType())
output_dict["spec_len"] = NeuralType(('B'), LengthsType())
return output_dict
return {
"audio_pred": NeuralType(('B', 'T'), AudioSignal()),
}
@typecheck()
def forward(self, *, audio, audio_len):
if self.mode != self.model.mode:
raise ValueError(
f"WaveGlowModel's mode {self.mode} does not match WaveGlowModule's mode {self.model.mode}"
)
spec, spec_len = self.audio_to_melspec_precessor(audio, audio_len)
tensors = self.model(spec=spec, audio=audio, sigma=self.sigma)
if self.mode == OperationMode.training:
return tensors # z, logdet, audio_pred
elif self.mode == OperationMode.validation:
z, logdet, audio_pred = tensors
return z, logdet, audio_pred, spec, spec_len
return tensors
@typecheck(
input_types={
"spec": NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
"sigma": NeuralType(optional=True)
},
output_types={"audio": NeuralType(('B', 'T'), AudioSignal())},
)
def convert_spectrogram_to_audio(self,
spec: torch.Tensor,
sigma: float = 1.0) -> torch.Tensor:
if not self.removed_weightnorm:
self.waveglow.remove_weightnorm()
self.removed_weightnorm = True
self.mode = OperationMode.infer
with torch.no_grad():
audio = self.model(spec=spec, audio=None, sigma=sigma)
return audio
def training_step(self, batch, batch_idx):
self.mode = OperationMode.training
audio, audio_len = batch
z, logdet, predicted_audio = self(audio=audio, audio_len=audio_len)
loss = self.loss(z=z,
logdet=logdet,
gt_audio=audio,
predicted_audio=predicted_audio,
sigma=self.sigma)
output = {
'loss': loss,
'progress_bar': {
'training_loss': loss
},
'log': {
'loss': loss
},
}
return output
def validation_step(self, batch, batch_idx):
self.mode = OperationMode.validation
audio, audio_len = batch
z, logdet, predicted_audio, spec, spec_len = self(audio=audio,
audio_len=audio_len)
loss = self.loss(z=z,
logdet=logdet,
gt_audio=audio,
predicted_audio=predicted_audio,
sigma=self.sigma)
# compute average stoi score for batch
stoi_score = 0
sr = self._cfg.preprocessor.sample_rate
for audio_i, audio_recon_i in zip(audio.cpu(), predicted_audio.cpu()):
stoi_score += stoi(audio_i, audio_recon_i, sr)
stoi_score /= audio.shape[0]
return {
"val_loss": loss,
"predicted_audio": predicted_audio,
"mel_target": spec,
"mel_len": spec_len,
"stoi": stoi_score,
}
def validation_epoch_end(self, outputs):
if self.logger is not None and self.logger.experiment is not 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
waveglow_log_to_tb_func(
tb_logger,
tuple(outputs[0].values())[:-1],
self.global_step,
tag="eval",
mel_fb=self.audio_to_melspec_precessor.fb,
)
avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
avg_stoi = torch.FloatTensor([x['stoi'] for x in outputs]).mean()
tensorboard_logs = {'val_loss': avg_loss, 'stoi': avg_stoi}
logging.info(
f"Validation summary | Epoch {self.current_epoch} | NLL {avg_loss:.2f} | STOI: {avg_stoi:.2f}"
)
return {'val_loss': avg_loss, 'log': tensorboard_logs}
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) -> '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="UniGlow-22050Hz",
location=
"https://drive.google.com/file/d/18JO5heoz1pBicZnGGqJzAJYMpzxiDQDa/view?usp=sharing",
description=
"The model is trained on LJSpeech sampled at 22050Hz, and can be used as an universal vocoder",
)
list_of_models.append(model)
return list_of_models
def get_upsample_factor(self) -> int:
"""
As the MelSpectrogram upsampling is done using interpolation, the upsampling factor is determined
by the ratio of the MelSpectrogram length and the waveform length
Returns:
An integer representing the upsampling factor
"""
audio = torch.ones(1, self._cfg.train_ds.dataset.n_segments)
spec, spec_len = self.audio_to_melspec_precessor(
audio, torch.FloatTensor([len(audio)]))
spec = spec[:, :, :-1]
audio = audio.unfold(1, self._cfg.uniglow.n_group,
self._cfg.uniglow.n_group).permute(0, 2, 1)
upsample_factor = audio.shape[2] // spec.shape[2]
return upsample_factor
def output_types(self):
if not self.calculate_loss and not self.training:
return {
"spec_pred_dec":
NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
"spec_pred_postnet":
NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
"gate_pred":
NeuralType(('B', 'T'), LogitsType()),
"alignments":
NeuralType(('B', 'T', 'T'), SequenceToSequenceAlignmentType()),
"pred_length":
NeuralType(('B'), LengthsType()),
}
return {
"spec_pred_dec":
NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
"spec_pred_postnet":
NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
"gate_pred":
NeuralType(('B', 'T'), LogitsType()),
"spec_target":
NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
"spec_target_len":
NeuralType(('B'), LengthsType()),
"alignments":
NeuralType(('B', 'T', 'T'), SequenceToSequenceAlignmentType()),
}
def input_types(self):
return {
"decoder_input": NeuralType(('B', 'T', 'D'),
EncodedRepresentation()),
"lengths": NeuralType(('B'), LengthsType()),
}
class Tacotron2Model(SpectrogramGenerator):
"""Tacotron 2 Model that is used to generate mel spectrograms from text"""
def __init__(self, cfg: DictConfig, trainer: 'Trainer' = None):
# Convert to Hydra 1.0 compatible DictConfig
cfg = model_utils.convert_model_config_to_dict_config(cfg)
cfg = model_utils.maybe_update_config_version(cfg)
# setup normalizer
self.normalizer = None
self.text_normalizer_call = None
self.text_normalizer_call_kwargs = {}
self._setup_normalizer(cfg)
# setup tokenizer
self.tokenizer = None
if hasattr(cfg, 'text_tokenizer'):
self._setup_tokenizer(cfg)
self.num_tokens = len(self.tokenizer.tokens)
self.tokenizer_pad = self.tokenizer.pad
self.tokenizer_unk = self.tokenizer.oov
# assert self.tokenizer is not None
else:
self.num_tokens = len(cfg.labels) + 3
super().__init__(cfg=cfg, trainer=trainer)
schema = OmegaConf.structured(Tacotron2Config)
# 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
try:
OmegaConf.merge(cfg, schema)
self.pad_value = cfg.preprocessor.pad_value
except ConfigAttributeError:
self.pad_value = cfg.preprocessor.params.pad_value
logging.warning(
"Your config is using an old NeMo yaml configuration. Please ensure that the yaml matches the "
"current version in the main branch for future compatibility.")
self._parser = None
self.audio_to_melspec_precessor = instantiate(cfg.preprocessor)
self.text_embedding = nn.Embedding(self.num_tokens, 512)
self.encoder = instantiate(self._cfg.encoder)
self.decoder = instantiate(self._cfg.decoder)
self.postnet = instantiate(self._cfg.postnet)
self.loss = Tacotron2Loss()
self.calculate_loss = True
@property
def parser(self):
if self._parser is not None:
return self._parser
ds_class_name = self._cfg.train_ds.dataset._target_.split(".")[-1]
if ds_class_name == "TTSDataset":
self._parser = None
elif hasattr(self._cfg, "labels"):
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,
)
elif ds_class_name == "AudioToCharWithPriorAndPitchDataset":
self.parser = self.vocab.encode
else:
raise ValueError(
"Wanted to setup parser, but model does not have necessary paramaters"
)
return self._parser
def parse(self, text: str, normalize=True) -> torch.Tensor:
if self.training:
logging.warning("parse() is meant to be called in eval mode.")
if normalize and self.text_normalizer_call is not None:
text = self.text_normalizer_call(
text, **self.text_normalizer_call_kwargs)
eval_phon_mode = contextlib.nullcontext()
if hasattr(self.tokenizer, "set_phone_prob"):
eval_phon_mode = self.tokenizer.set_phone_prob(prob=1.0)
with eval_phon_mode:
if self.tokenizer is not None:
tokens = self.tokenizer.encode(text)
else:
tokens = self.parser(text)
# Old parser doesn't add bos and eos ids, so maunally add it
tokens = [len(self._cfg.labels)
] + tokens + [len(self._cfg.labels) + 1]
tokens_tensor = torch.tensor(tokens).unsqueeze_(0).to(self.device)
return tokens_tensor
@property
def input_types(self):
if self.training:
return {
"tokens": NeuralType(('B', 'T'), EmbeddedTextType()),
"token_len": NeuralType(('B'), LengthsType()),
"audio": NeuralType(('B', 'T'), AudioSignal()),
"audio_len": NeuralType(('B'), LengthsType()),
}
else:
return {
"tokens": NeuralType(('B', 'T'), EmbeddedTextType()),
"token_len": NeuralType(('B'), LengthsType()),
"audio": NeuralType(('B', 'T'), AudioSignal(), optional=True),
"audio_len": NeuralType(('B'), LengthsType(), optional=True),
}
@property
def output_types(self):
if not self.calculate_loss and not self.training:
return {
"spec_pred_dec":
NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
"spec_pred_postnet":
NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
"gate_pred":
NeuralType(('B', 'T'), LogitsType()),
"alignments":
NeuralType(('B', 'T', 'T'), SequenceToSequenceAlignmentType()),
"pred_length":
NeuralType(('B'), LengthsType()),
}
return {
"spec_pred_dec":
NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
"spec_pred_postnet":
NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
"gate_pred":
NeuralType(('B', 'T'), LogitsType()),
"spec_target":
NeuralType(('B', 'D', 'T'), MelSpectrogramType()),
"spec_target_len":
NeuralType(('B'), LengthsType()),
"alignments":
NeuralType(('B', 'T', 'T'), SequenceToSequenceAlignmentType()),
}
@typecheck()
def forward(self, *, tokens, token_len, audio=None, audio_len=None):
if audio is not None and audio_len is not None:
spec_target, spec_target_len = self.audio_to_melspec_precessor(
audio, audio_len)
token_embedding = self.text_embedding(tokens).transpose(1, 2)
encoder_embedding = self.encoder(token_embedding=token_embedding,
token_len=token_len)
if self.training:
spec_pred_dec, gate_pred, alignments = self.decoder(
memory=encoder_embedding,
decoder_inputs=spec_target,
memory_lengths=token_len)
else:
spec_pred_dec, gate_pred, alignments, pred_length = self.decoder(
memory=encoder_embedding, memory_lengths=token_len)
spec_pred_postnet = self.postnet(mel_spec=spec_pred_dec)
if not self.calculate_loss:
return spec_pred_dec, spec_pred_postnet, gate_pred, alignments, pred_length
return spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, alignments
@typecheck(
input_types={"tokens": NeuralType(('B', 'T'), EmbeddedTextType())},
output_types={
"spec": NeuralType(('B', 'D', 'T'), MelSpectrogramType())
},
)
def generate_spectrogram(self, *, tokens):
self.eval()
self.calculate_loss = False
token_len = torch.tensor([len(i) for i in tokens]).to(self.device)
tensors = self(tokens=tokens, token_len=token_len)
spectrogram_pred = tensors[1]
if spectrogram_pred.shape[0] > 1:
# Silence all frames past the predicted end
mask = ~get_mask_from_lengths(tensors[-1])
mask = mask.expand(spectrogram_pred.shape[1], mask.size(0),
mask.size(1))
mask = mask.permute(1, 0, 2)
spectrogram_pred.data.masked_fill_(mask, self.pad_value)
return spectrogram_pred
def training_step(self, batch, batch_idx):
audio, audio_len, tokens, token_len = batch
spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, _ = self.forward(
audio=audio,
audio_len=audio_len,
tokens=tokens,
token_len=token_len)
loss, _ = self.loss(
spec_pred_dec=spec_pred_dec,
spec_pred_postnet=spec_pred_postnet,
gate_pred=gate_pred,
spec_target=spec_target,
spec_target_len=spec_target_len,
pad_value=self.pad_value,
)
output = {
'loss': loss,
'progress_bar': {
'training_loss': loss
},
'log': {
'loss': loss
},
}
return output
def validation_step(self, batch, batch_idx):
audio, audio_len, tokens, token_len = batch
spec_pred_dec, spec_pred_postnet, gate_pred, spec_target, spec_target_len, alignments = self.forward(
audio=audio,
audio_len=audio_len,
tokens=tokens,
token_len=token_len)
loss, gate_target = self.loss(
spec_pred_dec=spec_pred_dec,
spec_pred_postnet=spec_pred_postnet,
gate_pred=gate_pred,
spec_target=spec_target,
spec_target_len=spec_target_len,
pad_value=self.pad_value,
)
return {
"val_loss": loss,
"mel_target": spec_target,
"mel_postnet": spec_pred_postnet,
"gate": gate_pred,
"gate_target": gate_target,
"alignments": alignments,
}
def validation_epoch_end(self, outputs):
if self.logger is not None and self.logger.experiment is not None:
logger = self.logger.experiment
if isinstance(self.logger, LoggerCollection):
for logger in self.logger:
if isinstance(logger, TensorBoardLogger):
logger = logger.experiment
break
if isinstance(logger, TensorBoardLogger):
tacotron2_log_to_tb_func(
logger,
outputs[0].values(),
self.global_step,
tag="val",
log_images=True,
add_audio=False,
)
elif isinstance(logger, WandbLogger):
tacotron2_log_to_wandb_func(
logger,
outputs[0].values(),
self.global_step,
tag="val",
log_images=True,
add_audio=False,
)
avg_loss = torch.stack([
x['val_loss'] for x in outputs
]).mean() # This reduces across batches, not workers!
self.log('val_loss', avg_loss)
def _setup_normalizer(self, cfg):
if "text_normalizer" in cfg:
normalizer_kwargs = {}
if "whitelist" in cfg.text_normalizer:
normalizer_kwargs["whitelist"] = self.register_artifact(
'text_normalizer.whitelist', cfg.text_normalizer.whitelist)
self.normalizer = instantiate(cfg.text_normalizer,
**normalizer_kwargs)
self.text_normalizer_call = self.normalizer.normalize
if "text_normalizer_call_kwargs" in cfg:
self.text_normalizer_call_kwargs = cfg.text_normalizer_call_kwargs
def _setup_tokenizer(self, cfg):
text_tokenizer_kwargs = {}
if "g2p" in cfg.text_tokenizer and cfg.text_tokenizer.g2p is not None:
g2p_kwargs = {}
if "phoneme_dict" in cfg.text_tokenizer.g2p:
g2p_kwargs["phoneme_dict"] = self.register_artifact(
'text_tokenizer.g2p.phoneme_dict',
cfg.text_tokenizer.g2p.phoneme_dict,
)
if "heteronyms" in cfg.text_tokenizer.g2p:
g2p_kwargs["heteronyms"] = self.register_artifact(
'text_tokenizer.g2p.heteronyms',
cfg.text_tokenizer.g2p.heteronyms,
)
text_tokenizer_kwargs["g2p"] = instantiate(cfg.text_tokenizer.g2p,
**g2p_kwargs)
self.tokenizer = instantiate(cfg.text_tokenizer,
**text_tokenizer_kwargs)
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,
text_normalizer=self.normalizer,
text_normalizer_call_kwargs=self.text_normalizer_call_kwargs,
text_tokenizer=self.tokenizer,
)
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) -> '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_tacotron2",
location=
"https://api.ngc.nvidia.com/v2/models/nvidia/nemo/tts_en_tacotron2/versions/1.0.0/files/tts_en_tacotron2.nemo",
description=
"This model is trained on LJSpeech sampled at 22050Hz, and can be used to generate female English voices with an American accent.",
class_=cls,
aliases=["Tacotron2-22050Hz"],
)
list_of_models.append(model)
return list_of_models
def input_types(self): # phonemes
return {
"text": NeuralType(('B', 'T'), TokenIndex()),
"text_length": NeuralType(('B'), LengthsType())
}
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.loss = BaseFastPitchLoss()
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', 'T'), MelSpectrogramType()),
"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 = None
if splice:
output = []
splices = []
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(gen_in.transpose(1, 2))
return output, 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(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 = 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, pitch_loss = self.loss(
log_durs_predicted=log_dur_preds,
pitch_predicted=pitch_preds,
durs_tgt=durs,
dur_lens=text_lens,
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(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
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="",
# location="",
# description="",
# 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()
durations = torch.sum(
torch.clamp(
torch.exp(log_dur_pred) - 1, 0, self.max_token_duration), 1)
audio_list = []
for i, sample in enumerate(audio):
audio_list.append(sample[:durations[i] * self.hop_size])
return audio_list
def input_types(self):
return {
"x_mag": NeuralType(('B', 'T', 'D'), SpectrogramType()),
"y_mag": NeuralType(('B', 'T', 'D'), SpectrogramType()),
"input_lengths": NeuralType(('B'), LengthsType(), optional=True),
}
def input_types(self):
return {
"y": NeuralType(('B', 'S', 'T'), AudioSignal()),
"y_hat": NeuralType(('B', 'S', 'T'), AudioSignal()),
}
