class Tacotron2(BaseTacotron):
"""Tacotron2 as in https://arxiv.org/abs/1712.05884
Check `TacotronConfig` for the arguments.
"""
def __init__(self, config: Coqpit):
super().__init__(config)
chars, self.config = self.get_characters(config)
config.num_chars = len(chars)
self.decoder_output_dim = config.out_channels
# pass all config fields to `self`
# for fewer code change
for key in config:
setattr(self, key, config[key])
# set speaker embedding channel size for determining `in_channels` for the connected layers.
# `init_multispeaker` needs to be called once more in training to initialize the speaker embedding layer based
# on the number of speakers infered from the dataset.
if self.use_speaker_embedding or self.use_d_vector_file:
self.init_multispeaker(config)
self.decoder_in_features += (
self.embedded_speaker_dim
) # add speaker embedding dim
if self.use_gst:
self.decoder_in_features += self.gst.gst_embedding_dim
# embedding layer
self.embedding = nn.Embedding(self.num_chars, 512, padding_idx=0)
# base model layers
self.encoder = Encoder(self.encoder_in_features)
self.decoder = Decoder(
self.decoder_in_features,
self.decoder_output_dim,
self.r,
self.attention_type,
self.attention_win,
self.attention_norm,
self.prenet_type,
self.prenet_dropout,
self.use_forward_attn,
self.transition_agent,
self.forward_attn_mask,
self.location_attn,
self.attention_heads,
self.separate_stopnet,
self.max_decoder_steps,
)
self.postnet = Postnet(self.out_channels)
# setup prenet dropout
self.decoder.prenet.dropout_at_inference = self.prenet_dropout_at_inference
# global style token layers
if self.gst and self.use_gst:
self.gst_layer = GST(
num_mel=self.decoder_output_dim,
num_heads=self.gst.gst_num_heads,
num_style_tokens=self.gst.gst_num_style_tokens,
gst_embedding_dim=self.gst.gst_embedding_dim,
)
# backward pass decoder
if self.bidirectional_decoder:
self._init_backward_decoder()
# setup DDC
if self.double_decoder_consistency:
self.coarse_decoder = Decoder(
self.decoder_in_features,
self.decoder_output_dim,
self.ddc_r,
self.attention_type,
self.attention_win,
self.attention_norm,
self.prenet_type,
self.prenet_dropout,
self.use_forward_attn,
self.transition_agent,
self.forward_attn_mask,
self.location_attn,
self.attention_heads,
self.separate_stopnet,
self.max_decoder_steps,
)
@staticmethod
def shape_outputs(mel_outputs, mel_outputs_postnet, alignments):
mel_outputs = mel_outputs.transpose(1, 2)
mel_outputs_postnet = mel_outputs_postnet.transpose(1, 2)
return mel_outputs, mel_outputs_postnet, alignments
def forward( # pylint: disable=dangerous-default-value
self,
text,
text_lengths,
mel_specs=None,
mel_lengths=None,
aux_input={"speaker_ids": None, "d_vectors": None},
):
"""
Shapes:
text: [B, T_in]
text_lengths: [B]
mel_specs: [B, T_out, C]
mel_lengths: [B]
aux_input: 'speaker_ids': [B, 1] and 'd_vectors':[B, C]
"""
aux_input = self._format_aux_input(aux_input)
outputs = {"alignments_backward": None, "decoder_outputs_backward": None}
# compute mask for padding
# B x T_in_max (boolean)
input_mask, output_mask = self.compute_masks(text_lengths, mel_lengths)
# B x D_embed x T_in_max
embedded_inputs = self.embedding(text).transpose(1, 2)
# B x T_in_max x D_en
encoder_outputs = self.encoder(embedded_inputs, text_lengths)
if self.gst and self.use_gst:
# B x gst_dim
encoder_outputs = self.compute_gst(encoder_outputs, mel_specs)
if self.use_speaker_embedding or self.use_d_vector_file:
if not self.use_d_vector_file:
# B x 1 x speaker_embed_dim
embedded_speakers = self.speaker_embedding(aux_input["speaker_ids"])[
:, None
]
else:
# B x 1 x speaker_embed_dim
embedded_speakers = torch.unsqueeze(aux_input["d_vectors"], 1)
encoder_outputs = self._concat_speaker_embedding(
encoder_outputs, embedded_speakers
)
encoder_outputs = encoder_outputs * input_mask.unsqueeze(2).expand_as(
encoder_outputs
)
# B x mel_dim x T_out -- B x T_out//r x T_in -- B x T_out//r
decoder_outputs, alignments, stop_tokens = self.decoder(
encoder_outputs, mel_specs, input_mask
)
# sequence masking
if mel_lengths is not None:
decoder_outputs = decoder_outputs * output_mask.unsqueeze(1).expand_as(
decoder_outputs
)
# B x mel_dim x T_out
postnet_outputs = self.postnet(decoder_outputs)
postnet_outputs = decoder_outputs + postnet_outputs
# sequence masking
if output_mask is not None:
postnet_outputs = postnet_outputs * output_mask.unsqueeze(1).expand_as(
postnet_outputs
)
# B x T_out x mel_dim -- B x T_out x mel_dim -- B x T_out//r x T_in
decoder_outputs, postnet_outputs, alignments = self.shape_outputs(
decoder_outputs, postnet_outputs, alignments
)
if self.bidirectional_decoder:
decoder_outputs_backward, alignments_backward = self._backward_pass(
mel_specs, encoder_outputs, input_mask
)
outputs["alignments_backward"] = alignments_backward
outputs["decoder_outputs_backward"] = decoder_outputs_backward
if self.double_decoder_consistency:
decoder_outputs_backward, alignments_backward = self._coarse_decoder_pass(
mel_specs, encoder_outputs, alignments, input_mask
)
outputs["alignments_backward"] = alignments_backward
outputs["decoder_outputs_backward"] = decoder_outputs_backward
outputs.update(
{
"model_outputs": postnet_outputs,
"decoder_outputs": decoder_outputs,
"alignments": alignments,
"stop_tokens": stop_tokens,
}
)
return outputs
@torch.no_grad()
def inference(self, text, aux_input=None):
aux_input = self._format_aux_input(aux_input)
embedded_inputs = self.embedding(text).transpose(1, 2)
encoder_outputs = self.encoder.inference(embedded_inputs)
if self.gst and self.use_gst:
# B x gst_dim
encoder_outputs = self.compute_gst(
encoder_outputs, aux_input["style_mel"], aux_input["d_vectors"]
)
if self.num_speakers > 1:
if not self.use_d_vector_file:
embedded_speakers = self.speaker_embedding(aux_input["speaker_ids"])[
None
]
# reshape embedded_speakers
if embedded_speakers.ndim == 1:
embedded_speakers = embedded_speakers[None, None, :]
elif embedded_speakers.ndim == 2:
embedded_speakers = embedded_speakers[None, :]
else:
embedded_speakers = aux_input["d_vectors"]
encoder_outputs = self._concat_speaker_embedding(
encoder_outputs, embedded_speakers
)
decoder_outputs, alignments, stop_tokens = self.decoder.inference(
encoder_outputs
)
postnet_outputs = self.postnet(decoder_outputs)
postnet_outputs = decoder_outputs + postnet_outputs
decoder_outputs, postnet_outputs, alignments = self.shape_outputs(
decoder_outputs, postnet_outputs, alignments
)
outputs = {
"model_outputs": postnet_outputs,
"decoder_outputs": decoder_outputs,
"alignments": alignments,
"stop_tokens": stop_tokens,
}
return outputs
def train_step(self, batch, criterion):
"""Perform a single training step by fetching the right set if samples from the batch.
Args:
batch ([type]): [description]
criterion ([type]): [description]
"""
text_input = batch["text_input"]
text_lengths = batch["text_lengths"]
mel_input = batch["mel_input"]
mel_lengths = batch["mel_lengths"]
linear_input = batch["linear_input"]
stop_targets = batch["stop_targets"]
stop_target_lengths = batch["stop_target_lengths"]
speaker_ids = batch["speaker_ids"]
d_vectors = batch["d_vectors"]
# forward pass model
outputs = self.forward(
text_input,
text_lengths,
mel_input,
mel_lengths,
aux_input={"speaker_ids": speaker_ids, "d_vectors": d_vectors},
)
# set the [alignment] lengths wrt reduction factor for guided attention
if mel_lengths.max() % self.decoder.r != 0:
alignment_lengths = (
mel_lengths + (self.decoder.r - (mel_lengths.max() % self.decoder.r))
) // self.decoder.r
else:
alignment_lengths = mel_lengths // self.decoder.r
aux_input = {"speaker_ids": speaker_ids, "d_vectors": d_vectors}
outputs = self.forward(
text_input, text_lengths, mel_input, mel_lengths, aux_input
)
# compute loss
loss_dict = criterion(
outputs["model_outputs"],
outputs["decoder_outputs"],
mel_input,
linear_input,
outputs["stop_tokens"],
stop_targets,
stop_target_lengths,
mel_lengths,
outputs["decoder_outputs_backward"],
outputs["alignments"],
alignment_lengths,
outputs["alignments_backward"],
text_lengths,
)
# compute alignment error (the lower the better )
align_error = 1 - alignment_diagonal_score(outputs["alignments"])
loss_dict["align_error"] = align_error
return outputs, loss_dict
def train_log(
self, ap: AudioProcessor, batch: dict, outputs: dict
) -> Tuple[Dict, Dict]:
raise NotImplementedError()
def eval_step(self, batch, criterion):
return self.train_step(batch, criterion)
def eval_log(self, ap, batch, outputs):
return self.train_log(ap, batch, outputs)
def __init__(self, config: Coqpit):
super().__init__(config)
chars, self.config = self.get_characters(config)
config.num_chars = len(chars)
self.decoder_output_dim = config.out_channels
# pass all config fields to `self`
# for fewer code change
for key in config:
setattr(self, key, config[key])
# set speaker embedding channel size for determining `in_channels` for the connected layers.
# `init_multispeaker` needs to be called once more in training to initialize the speaker embedding layer based
# on the number of speakers infered from the dataset.
if self.use_speaker_embedding or self.use_d_vector_file:
self.init_multispeaker(config)
self.decoder_in_features += (
self.embedded_speaker_dim
) # add speaker embedding dim
if self.use_gst:
self.decoder_in_features += self.gst.gst_embedding_dim
# embedding layer
self.embedding = nn.Embedding(self.num_chars, 512, padding_idx=0)
# base model layers
self.encoder = Encoder(self.encoder_in_features)
self.decoder = Decoder(
self.decoder_in_features,
self.decoder_output_dim,
self.r,
self.attention_type,
self.attention_win,
self.attention_norm,
self.prenet_type,
self.prenet_dropout,
self.use_forward_attn,
self.transition_agent,
self.forward_attn_mask,
self.location_attn,
self.attention_heads,
self.separate_stopnet,
self.max_decoder_steps,
)
self.postnet = Postnet(self.out_channels)
# setup prenet dropout
self.decoder.prenet.dropout_at_inference = self.prenet_dropout_at_inference
# global style token layers
if self.gst and self.use_gst:
self.gst_layer = GST(
num_mel=self.decoder_output_dim,
num_heads=self.gst.gst_num_heads,
num_style_tokens=self.gst.gst_num_style_tokens,
gst_embedding_dim=self.gst.gst_embedding_dim,
)
# backward pass decoder
if self.bidirectional_decoder:
self._init_backward_decoder()
# setup DDC
if self.double_decoder_consistency:
self.coarse_decoder = Decoder(
self.decoder_in_features,
self.decoder_output_dim,
self.ddc_r,
self.attention_type,
self.attention_win,
self.attention_norm,
self.prenet_type,
self.prenet_dropout,
self.use_forward_attn,
self.transition_agent,
self.forward_attn_mask,
self.location_attn,
self.attention_heads,
self.separate_stopnet,
self.max_decoder_steps,
)
class Tacotron2(TacotronAbstract):
"""Tacotron2 as in https://arxiv.org/abs/1712.05884
It's an autoregressive encoder-attention-decoder-postnet architecture.
Args:
num_chars (int): number of input characters to define the size of embedding layer.
num_speakers (int): number of speakers in the dataset. >1 enables multi-speaker training and model learns speaker embeddings.
r (int): initial model reduction rate.
postnet_output_dim (int, optional): postnet output channels. Defaults to 80.
decoder_output_dim (int, optional): decoder output channels. Defaults to 80.
attn_type (str, optional): attention type. Check ```TTS.tts.layers.tacotron.common_layers.init_attn```. Defaults to 'original'.
attn_win (bool, optional): enable/disable attention windowing.
It especially useful at inference to keep attention alignment diagonal. Defaults to False.
attn_norm (str, optional): Attention normalization method. "sigmoid" or "softmax". Defaults to "softmax".
prenet_type (str, optional): prenet type for the decoder. Defaults to "original".
prenet_dropout (bool, optional): prenet dropout rate. Defaults to True.
prenet_dropout_at_inference (bool, optional): use dropout at inference time. This leads to a better quality for
some models. Defaults to False.
forward_attn (bool, optional): enable/disable forward attention.
It is only valid if ```attn_type``` is ```original```. Defaults to False.
trans_agent (bool, optional): enable/disable transition agent in forward attention. Defaults to False.
forward_attn_mask (bool, optional): enable/disable extra masking over forward attention. Defaults to False.
location_attn (bool, optional): enable/disable location sensitive attention.
It is only valid if ```attn_type``` is ```original```. Defaults to True.
attn_K (int, optional): Number of attention heads for GMM attention. Defaults to 5.
separate_stopnet (bool, optional): enable/disable separate stopnet training without only gradient
flow from stopnet to the rest of the model. Defaults to True.
bidirectional_decoder (bool, optional): enable/disable bidirectional decoding. Defaults to False.
double_decoder_consistency (bool, optional): enable/disable double decoder consistency. Defaults to False.
ddc_r (int, optional): reduction rate for the coarse decoder of double decoder consistency. Defaults to None.
encoder_in_features (int, optional): input channels for the encoder. Defaults to 512.
decoder_in_features (int, optional): input channels for the decoder. Defaults to 512.
speaker_embedding_dim (int, optional): external speaker conditioning vector channels. Defaults to None.
gst (bool, optional): enable/disable global style token learning. Defaults to False.
gst_embedding_dim (int, optional): size of channels for GST vectors. Defaults to 512.
gst_num_heads (int, optional): number of attention heads for GST. Defaults to 4.
gst_style_tokens (int, optional): number of GST tokens. Defaults to 10.
gst_use_speaker_embedding (bool, optional): enable/disable inputing speaker embedding to GST. Defaults to False.
"""
def __init__(
self,
num_chars,
num_speakers,
r,
postnet_output_dim=80,
decoder_output_dim=80,
attn_type="original",
attn_win=False,
attn_norm="softmax",
prenet_type="original",
prenet_dropout=True,
prenet_dropout_at_inference=False,
forward_attn=False,
trans_agent=False,
forward_attn_mask=False,
location_attn=True,
attn_K=5,
separate_stopnet=True,
bidirectional_decoder=False,
double_decoder_consistency=False,
ddc_r=None,
encoder_in_features=512,
decoder_in_features=512,
speaker_embedding_dim=None,
gst=False,
gst_embedding_dim=512,
gst_num_heads=4,
gst_style_tokens=10,
gst_use_speaker_embedding=False,
):
super().__init__(
num_chars,
num_speakers,
r,
postnet_output_dim,
decoder_output_dim,
attn_type,
attn_win,
attn_norm,
prenet_type,
prenet_dropout,
prenet_dropout_at_inference,
forward_attn,
trans_agent,
forward_attn_mask,
location_attn,
attn_K,
separate_stopnet,
bidirectional_decoder,
double_decoder_consistency,
ddc_r,
encoder_in_features,
decoder_in_features,
speaker_embedding_dim,
gst,
gst_embedding_dim,
gst_num_heads,
gst_style_tokens,
gst_use_speaker_embedding,
)
# speaker embedding layer
if self.num_speakers > 1:
if not self.embeddings_per_sample:
speaker_embedding_dim = 512
self.speaker_embedding = nn.Embedding(self.num_speakers, speaker_embedding_dim)
self.speaker_embedding.weight.data.normal_(0, 0.3)
# speaker and gst embeddings is concat in decoder input
if self.num_speakers > 1:
self.decoder_in_features += speaker_embedding_dim # add speaker embedding dim
# embedding layer
self.embedding = nn.Embedding(num_chars, 512, padding_idx=0)
# base model layers
self.encoder = Encoder(self.encoder_in_features)
self.decoder = Decoder(
self.decoder_in_features,
self.decoder_output_dim,
r,
attn_type,
attn_win,
attn_norm,
prenet_type,
prenet_dropout,
forward_attn,
trans_agent,
forward_attn_mask,
location_attn,
attn_K,
separate_stopnet,
)
self.postnet = Postnet(self.postnet_output_dim)
# setup prenet dropout
self.decoder.prenet.dropout_at_inference = prenet_dropout_at_inference
# global style token layers
if self.gst:
self.gst_layer = GST(
num_mel=80,
num_heads=self.gst_num_heads,
num_style_tokens=self.gst_style_tokens,
gst_embedding_dim=self.gst_embedding_dim,
speaker_embedding_dim=speaker_embedding_dim
if self.embeddings_per_sample and self.gst_use_speaker_embedding
else None,
)
# backward pass decoder
if self.bidirectional_decoder:
self._init_backward_decoder()
# setup DDC
if self.double_decoder_consistency:
self.coarse_decoder = Decoder(
self.decoder_in_features,
self.decoder_output_dim,
ddc_r,
attn_type,
attn_win,
attn_norm,
prenet_type,
prenet_dropout,
forward_attn,
trans_agent,
forward_attn_mask,
location_attn,
attn_K,
separate_stopnet,
)
@staticmethod
def shape_outputs(mel_outputs, mel_outputs_postnet, alignments):
mel_outputs = mel_outputs.transpose(1, 2)
mel_outputs_postnet = mel_outputs_postnet.transpose(1, 2)
return mel_outputs, mel_outputs_postnet, alignments
def forward(self, text, text_lengths, mel_specs=None, mel_lengths=None, speaker_ids=None, speaker_embeddings=None):
"""
Shapes:
text: [B, T_in]
text_lengths: [B]
mel_specs: [B, T_out, C]
mel_lengths: [B]
speaker_ids: [B, 1]
speaker_embeddings: [B, C]
"""
# compute mask for padding
# B x T_in_max (boolean)
input_mask, output_mask = self.compute_masks(text_lengths, mel_lengths)
# B x D_embed x T_in_max
embedded_inputs = self.embedding(text).transpose(1, 2)
# B x T_in_max x D_en
encoder_outputs = self.encoder(embedded_inputs, text_lengths)
if self.gst:
# B x gst_dim
encoder_outputs = self.compute_gst(
encoder_outputs, mel_specs, speaker_embeddings if self.gst_use_speaker_embedding else None
)
if self.num_speakers > 1:
if not self.embeddings_per_sample:
# B x 1 x speaker_embed_dim
speaker_embeddings = self.speaker_embedding(speaker_ids)[:, None]
else:
# B x 1 x speaker_embed_dim
speaker_embeddings = torch.unsqueeze(speaker_embeddings, 1)
encoder_outputs = self._concat_speaker_embedding(encoder_outputs, speaker_embeddings)
encoder_outputs = encoder_outputs * input_mask.unsqueeze(2).expand_as(encoder_outputs)
# B x mel_dim x T_out -- B x T_out//r x T_in -- B x T_out//r
decoder_outputs, alignments, stop_tokens = self.decoder(encoder_outputs, mel_specs, input_mask)
# sequence masking
if mel_lengths is not None:
decoder_outputs = decoder_outputs * output_mask.unsqueeze(1).expand_as(decoder_outputs)
# B x mel_dim x T_out
postnet_outputs = self.postnet(decoder_outputs)
postnet_outputs = decoder_outputs + postnet_outputs
# sequence masking
if output_mask is not None:
postnet_outputs = postnet_outputs * output_mask.unsqueeze(1).expand_as(postnet_outputs)
# B x T_out x mel_dim -- B x T_out x mel_dim -- B x T_out//r x T_in
decoder_outputs, postnet_outputs, alignments = self.shape_outputs(decoder_outputs, postnet_outputs, alignments)
if self.bidirectional_decoder:
decoder_outputs_backward, alignments_backward = self._backward_pass(mel_specs, encoder_outputs, input_mask)
return (
decoder_outputs,
postnet_outputs,
alignments,
stop_tokens,
decoder_outputs_backward,
alignments_backward,
)
if self.double_decoder_consistency:
decoder_outputs_backward, alignments_backward = self._coarse_decoder_pass(
mel_specs, encoder_outputs, alignments, input_mask
)
return (
decoder_outputs,
postnet_outputs,
alignments,
stop_tokens,
decoder_outputs_backward,
alignments_backward,
)
return decoder_outputs, postnet_outputs, alignments, stop_tokens
@torch.no_grad()
def inference(self, text, speaker_ids=None, style_mel=None, speaker_embeddings=None):
embedded_inputs = self.embedding(text).transpose(1, 2)
encoder_outputs = self.encoder.inference(embedded_inputs)
if self.gst:
# B x gst_dim
encoder_outputs = self.compute_gst(
encoder_outputs, style_mel, speaker_embeddings if self.gst_use_speaker_embedding else None
)
if self.num_speakers > 1:
if not self.embeddings_per_sample:
speaker_embeddings = self.speaker_embedding(speaker_ids)[:, None]
encoder_outputs = self._concat_speaker_embedding(encoder_outputs, speaker_embeddings)
decoder_outputs, alignments, stop_tokens = self.decoder.inference(encoder_outputs)
postnet_outputs = self.postnet(decoder_outputs)
postnet_outputs = decoder_outputs + postnet_outputs
decoder_outputs, postnet_outputs, alignments = self.shape_outputs(decoder_outputs, postnet_outputs, alignments)
return decoder_outputs, postnet_outputs, alignments, stop_tokens
def inference_truncated(self, text, speaker_ids=None, style_mel=None, speaker_embeddings=None):
"""
Preserve model states for continuous inference
"""
embedded_inputs = self.embedding(text).transpose(1, 2)
encoder_outputs = self.encoder.inference_truncated(embedded_inputs)
if self.gst:
# B x gst_dim
encoder_outputs = self.compute_gst(
encoder_outputs, style_mel, speaker_embeddings if self.gst_use_speaker_embedding else None
)
if self.num_speakers > 1:
if not self.embeddings_per_sample:
speaker_embeddings = self.speaker_embedding(speaker_ids)[:, None]
encoder_outputs = self._concat_speaker_embedding(encoder_outputs, speaker_embeddings)
mel_outputs, alignments, stop_tokens = self.decoder.inference_truncated(encoder_outputs)
mel_outputs_postnet = self.postnet(mel_outputs)
mel_outputs_postnet = mel_outputs + mel_outputs_postnet
mel_outputs, mel_outputs_postnet, alignments = self.shape_outputs(mel_outputs, mel_outputs_postnet, alignments)
return mel_outputs, mel_outputs_postnet, alignments, stop_tokens
def __init__(
self,
num_chars,
num_speakers,
r,
postnet_output_dim=80,
decoder_output_dim=80,
attn_type="original",
attn_win=False,
attn_norm="softmax",
prenet_type="original",
prenet_dropout=True,
prenet_dropout_at_inference=False,
forward_attn=False,
trans_agent=False,
forward_attn_mask=False,
location_attn=True,
attn_K=5,
separate_stopnet=True,
bidirectional_decoder=False,
double_decoder_consistency=False,
ddc_r=None,
encoder_in_features=512,
decoder_in_features=512,
speaker_embedding_dim=None,
gst=False,
gst_embedding_dim=512,
gst_num_heads=4,
gst_style_tokens=10,
gst_use_speaker_embedding=False,
):
super().__init__(
num_chars,
num_speakers,
r,
postnet_output_dim,
decoder_output_dim,
attn_type,
attn_win,
attn_norm,
prenet_type,
prenet_dropout,
prenet_dropout_at_inference,
forward_attn,
trans_agent,
forward_attn_mask,
location_attn,
attn_K,
separate_stopnet,
bidirectional_decoder,
double_decoder_consistency,
ddc_r,
encoder_in_features,
decoder_in_features,
speaker_embedding_dim,
gst,
gst_embedding_dim,
gst_num_heads,
gst_style_tokens,
gst_use_speaker_embedding,
)
# speaker embedding layer
if self.num_speakers > 1:
if not self.embeddings_per_sample:
speaker_embedding_dim = 512
self.speaker_embedding = nn.Embedding(self.num_speakers, speaker_embedding_dim)
self.speaker_embedding.weight.data.normal_(0, 0.3)
# speaker and gst embeddings is concat in decoder input
if self.num_speakers > 1:
self.decoder_in_features += speaker_embedding_dim # add speaker embedding dim
# embedding layer
self.embedding = nn.Embedding(num_chars, 512, padding_idx=0)
# base model layers
self.encoder = Encoder(self.encoder_in_features)
self.decoder = Decoder(
self.decoder_in_features,
self.decoder_output_dim,
r,
attn_type,
attn_win,
attn_norm,
prenet_type,
prenet_dropout,
forward_attn,
trans_agent,
forward_attn_mask,
location_attn,
attn_K,
separate_stopnet,
)
self.postnet = Postnet(self.postnet_output_dim)
# setup prenet dropout
self.decoder.prenet.dropout_at_inference = prenet_dropout_at_inference
# global style token layers
if self.gst:
self.gst_layer = GST(
num_mel=80,
num_heads=self.gst_num_heads,
num_style_tokens=self.gst_style_tokens,
gst_embedding_dim=self.gst_embedding_dim,
speaker_embedding_dim=speaker_embedding_dim
if self.embeddings_per_sample and self.gst_use_speaker_embedding
else None,
)
# backward pass decoder
if self.bidirectional_decoder:
self._init_backward_decoder()
# setup DDC
if self.double_decoder_consistency:
self.coarse_decoder = Decoder(
self.decoder_in_features,
self.decoder_output_dim,
ddc_r,
attn_type,
attn_win,
attn_norm,
prenet_type,
prenet_dropout,
forward_attn,
trans_agent,
forward_attn_mask,
location_attn,
attn_K,
separate_stopnet,
)
def __init__(self, config: Coqpit, speaker_manager: SpeakerManager = None):
super().__init__(config)
self.speaker_manager = speaker_manager
chars, self.config, _ = self.get_characters(config)
config.num_chars = len(chars)
self.decoder_output_dim = config.out_channels
# pass all config fields to `self`
# for fewer code change
for key in config:
setattr(self, key, config[key])
# init multi-speaker layers
if self.use_speaker_embedding or self.use_d_vector_file:
self.init_multispeaker(config)
self.decoder_in_features += self.embedded_speaker_dim # add speaker embedding dim
if self.use_gst:
self.decoder_in_features += self.gst.gst_embedding_dim
# embedding layer
self.embedding = nn.Embedding(self.num_chars, 512, padding_idx=0)
# base model layers
self.encoder = Encoder(self.encoder_in_features)
self.decoder = Decoder(
self.decoder_in_features,
self.decoder_output_dim,
self.r,
self.attention_type,
self.attention_win,
self.attention_norm,
self.prenet_type,
self.prenet_dropout,
self.use_forward_attn,
self.transition_agent,
self.forward_attn_mask,
self.location_attn,
self.attention_heads,
self.separate_stopnet,
self.max_decoder_steps,
)
self.postnet = Postnet(self.out_channels)
# setup prenet dropout
self.decoder.prenet.dropout_at_inference = self.prenet_dropout_at_inference
# global style token layers
if self.gst and self.use_gst:
self.gst_layer = GST(
num_mel=self.decoder_output_dim,
num_heads=self.gst.gst_num_heads,
num_style_tokens=self.gst.gst_num_style_tokens,
gst_embedding_dim=self.gst.gst_embedding_dim,
)
# backward pass decoder
if self.bidirectional_decoder:
self._init_backward_decoder()
# setup DDC
if self.double_decoder_consistency:
self.coarse_decoder = Decoder(
self.decoder_in_features,
self.decoder_output_dim,
self.ddc_r,
self.attention_type,
self.attention_win,
self.attention_norm,
self.prenet_type,
self.prenet_dropout,
self.use_forward_attn,
self.transition_agent,
self.forward_attn_mask,
self.location_attn,
self.attention_heads,
self.separate_stopnet,
self.max_decoder_steps,
)
class Tacotron2(BaseTacotron):
"""Tacotron2 model implementation inherited from :class:`TTS.tts.models.base_tacotron.BaseTacotron`.
Paper::
https://arxiv.org/abs/1712.05884
Paper abstract::
This paper describes Tacotron 2, a neural network architecture for speech synthesis directly from text.
The system is composed of a recurrent sequence-to-sequence feature prediction network that maps character
embeddings to mel-scale spectrograms, followed by a modified WaveNet model acting as a vocoder to synthesize
timedomain waveforms from those spectrograms. Our model achieves a mean opinion score (MOS) of 4.53 comparable
to a MOS of 4.58 for professionally recorded speech. To validate our design choices, we present ablation
studies of key components of our system and evaluate the impact of using mel spectrograms as the input to
WaveNet instead of linguistic, duration, and F0 features. We further demonstrate that using a compact acoustic
intermediate representation enables significant simplification of the WaveNet architecture.
Check :class:`TTS.tts.configs.tacotron2_config.Tacotron2Config` for model arguments.
Args:
config (TacotronConfig):
Configuration for the Tacotron2 model.
speaker_manager (SpeakerManager):
Speaker manager for multi-speaker training. Uuse only for multi-speaker training. Defaults to None.
"""
def __init__(self, config: Coqpit, speaker_manager: SpeakerManager = None):
super().__init__(config)
self.speaker_manager = speaker_manager
chars, self.config, _ = self.get_characters(config)
config.num_chars = len(chars)
self.decoder_output_dim = config.out_channels
# pass all config fields to `self`
# for fewer code change
for key in config:
setattr(self, key, config[key])
# init multi-speaker layers
if self.use_speaker_embedding or self.use_d_vector_file:
self.init_multispeaker(config)
self.decoder_in_features += self.embedded_speaker_dim # add speaker embedding dim
if self.use_gst:
self.decoder_in_features += self.gst.gst_embedding_dim
# embedding layer
self.embedding = nn.Embedding(self.num_chars, 512, padding_idx=0)
# base model layers
self.encoder = Encoder(self.encoder_in_features)
self.decoder = Decoder(
self.decoder_in_features,
self.decoder_output_dim,
self.r,
self.attention_type,
self.attention_win,
self.attention_norm,
self.prenet_type,
self.prenet_dropout,
self.use_forward_attn,
self.transition_agent,
self.forward_attn_mask,
self.location_attn,
self.attention_heads,
self.separate_stopnet,
self.max_decoder_steps,
)
self.postnet = Postnet(self.out_channels)
# setup prenet dropout
self.decoder.prenet.dropout_at_inference = self.prenet_dropout_at_inference
# global style token layers
if self.gst and self.use_gst:
self.gst_layer = GST(
num_mel=self.decoder_output_dim,
num_heads=self.gst.gst_num_heads,
num_style_tokens=self.gst.gst_num_style_tokens,
gst_embedding_dim=self.gst.gst_embedding_dim,
)
# backward pass decoder
if self.bidirectional_decoder:
self._init_backward_decoder()
# setup DDC
if self.double_decoder_consistency:
self.coarse_decoder = Decoder(
self.decoder_in_features,
self.decoder_output_dim,
self.ddc_r,
self.attention_type,
self.attention_win,
self.attention_norm,
self.prenet_type,
self.prenet_dropout,
self.use_forward_attn,
self.transition_agent,
self.forward_attn_mask,
self.location_attn,
self.attention_heads,
self.separate_stopnet,
self.max_decoder_steps,
)
@staticmethod
def shape_outputs(mel_outputs, mel_outputs_postnet, alignments):
"""Final reshape of the model output tensors."""
mel_outputs = mel_outputs.transpose(1, 2)
mel_outputs_postnet = mel_outputs_postnet.transpose(1, 2)
return mel_outputs, mel_outputs_postnet, alignments
def forward( # pylint: disable=dangerous-default-value
self, text, text_lengths, mel_specs=None, mel_lengths=None, aux_input={"speaker_ids": None, "d_vectors": None}
):
"""Forward pass for training with Teacher Forcing.
Shapes:
text: :math:`[B, T_in]`
text_lengths: :math:`[B]`
mel_specs: :math:`[B, T_out, C]`
mel_lengths: :math:`[B]`
aux_input: 'speaker_ids': :math:`[B, 1]` and 'd_vectors': :math:`[B, C]`
"""
aux_input = self._format_aux_input(aux_input)
outputs = {"alignments_backward": None, "decoder_outputs_backward": None}
# compute mask for padding
# B x T_in_max (boolean)
input_mask, output_mask = self.compute_masks(text_lengths, mel_lengths)
# B x D_embed x T_in_max
embedded_inputs = self.embedding(text).transpose(1, 2)
# B x T_in_max x D_en
encoder_outputs = self.encoder(embedded_inputs, text_lengths)
if self.gst and self.use_gst:
# B x gst_dim
encoder_outputs = self.compute_gst(encoder_outputs, mel_specs)
if self.use_speaker_embedding or self.use_d_vector_file:
if not self.use_d_vector_file:
# B x 1 x speaker_embed_dim
embedded_speakers = self.speaker_embedding(aux_input["speaker_ids"])[:, None]
else:
# B x 1 x speaker_embed_dim
embedded_speakers = torch.unsqueeze(aux_input["d_vectors"], 1)
encoder_outputs = self._concat_speaker_embedding(encoder_outputs, embedded_speakers)
encoder_outputs = encoder_outputs * input_mask.unsqueeze(2).expand_as(encoder_outputs)
# B x mel_dim x T_out -- B x T_out//r x T_in -- B x T_out//r
decoder_outputs, alignments, stop_tokens = self.decoder(encoder_outputs, mel_specs, input_mask)
# sequence masking
if mel_lengths is not None:
decoder_outputs = decoder_outputs * output_mask.unsqueeze(1).expand_as(decoder_outputs)
# B x mel_dim x T_out
postnet_outputs = self.postnet(decoder_outputs)
postnet_outputs = decoder_outputs + postnet_outputs
# sequence masking
if output_mask is not None:
postnet_outputs = postnet_outputs * output_mask.unsqueeze(1).expand_as(postnet_outputs)
# B x T_out x mel_dim -- B x T_out x mel_dim -- B x T_out//r x T_in
decoder_outputs, postnet_outputs, alignments = self.shape_outputs(decoder_outputs, postnet_outputs, alignments)
if self.bidirectional_decoder:
decoder_outputs_backward, alignments_backward = self._backward_pass(mel_specs, encoder_outputs, input_mask)
outputs["alignments_backward"] = alignments_backward
outputs["decoder_outputs_backward"] = decoder_outputs_backward
if self.double_decoder_consistency:
decoder_outputs_backward, alignments_backward = self._coarse_decoder_pass(
mel_specs, encoder_outputs, alignments, input_mask
)
outputs["alignments_backward"] = alignments_backward
outputs["decoder_outputs_backward"] = decoder_outputs_backward
outputs.update(
{
"model_outputs": postnet_outputs,
"decoder_outputs": decoder_outputs,
"alignments": alignments,
"stop_tokens": stop_tokens,
}
)
return outputs
@torch.no_grad()
def inference(self, text, aux_input=None):
"""Forward pass for inference with no Teacher-Forcing.
Shapes:
text: :math:`[B, T_in]`
text_lengths: :math:`[B]`
"""
aux_input = self._format_aux_input(aux_input)
embedded_inputs = self.embedding(text).transpose(1, 2)
encoder_outputs = self.encoder.inference(embedded_inputs)
if self.gst and self.use_gst:
# B x gst_dim
encoder_outputs = self.compute_gst(encoder_outputs, aux_input["style_mel"], aux_input["d_vectors"])
if self.num_speakers > 1:
if not self.use_d_vector_file:
embedded_speakers = self.speaker_embedding(aux_input["speaker_ids"])[None]
# reshape embedded_speakers
if embedded_speakers.ndim == 1:
embedded_speakers = embedded_speakers[None, None, :]
elif embedded_speakers.ndim == 2:
embedded_speakers = embedded_speakers[None, :]
else:
embedded_speakers = aux_input["d_vectors"]
encoder_outputs = self._concat_speaker_embedding(encoder_outputs, embedded_speakers)
decoder_outputs, alignments, stop_tokens = self.decoder.inference(encoder_outputs)
postnet_outputs = self.postnet(decoder_outputs)
postnet_outputs = decoder_outputs + postnet_outputs
decoder_outputs, postnet_outputs, alignments = self.shape_outputs(decoder_outputs, postnet_outputs, alignments)
outputs = {
"model_outputs": postnet_outputs,
"decoder_outputs": decoder_outputs,
"alignments": alignments,
"stop_tokens": stop_tokens,
}
return outputs
def train_step(self, batch: Dict, criterion: torch.nn.Module):
"""A single training step. Forward pass and loss computation.
Args:
batch ([Dict]): A dictionary of input tensors.
criterion ([type]): Callable criterion to compute model loss.
"""
text_input = batch["text_input"]
text_lengths = batch["text_lengths"]
mel_input = batch["mel_input"]
mel_lengths = batch["mel_lengths"]
stop_targets = batch["stop_targets"]
stop_target_lengths = batch["stop_target_lengths"]
speaker_ids = batch["speaker_ids"]
d_vectors = batch["d_vectors"]
# forward pass model
outputs = self.forward(
text_input,
text_lengths,
mel_input,
mel_lengths,
aux_input={"speaker_ids": speaker_ids, "d_vectors": d_vectors},
)
# set the [alignment] lengths wrt reduction factor for guided attention
if mel_lengths.max() % self.decoder.r != 0:
alignment_lengths = (
mel_lengths + (self.decoder.r - (mel_lengths.max() % self.decoder.r))
) // self.decoder.r
else:
alignment_lengths = mel_lengths // self.decoder.r
aux_input = {"speaker_ids": speaker_ids, "d_vectors": d_vectors}
outputs = self.forward(text_input, text_lengths, mel_input, mel_lengths, aux_input)
# compute loss
with autocast(enabled=False): # use float32 for the criterion
loss_dict = criterion(
outputs["model_outputs"].float(),
outputs["decoder_outputs"].float(),
mel_input.float(),
None,
outputs["stop_tokens"].float(),
stop_targets.float(),
stop_target_lengths,
mel_lengths,
None if outputs["decoder_outputs_backward"] is None else outputs["decoder_outputs_backward"].float(),
outputs["alignments"].float(),
alignment_lengths,
None if outputs["alignments_backward"] is None else outputs["alignments_backward"].float(),
text_lengths,
)
# compute alignment error (the lower the better )
align_error = 1 - alignment_diagonal_score(outputs["alignments"])
loss_dict["align_error"] = align_error
return outputs, loss_dict
def _create_logs(self, batch, outputs, ap):
"""Create dashboard log information."""
postnet_outputs = outputs["model_outputs"]
alignments = outputs["alignments"]
alignments_backward = outputs["alignments_backward"]
mel_input = batch["mel_input"]
pred_spec = postnet_outputs[0].data.cpu().numpy()
gt_spec = mel_input[0].data.cpu().numpy()
align_img = alignments[0].data.cpu().numpy()
figures = {
"prediction": plot_spectrogram(pred_spec, ap, output_fig=False),
"ground_truth": plot_spectrogram(gt_spec, ap, output_fig=False),
"alignment": plot_alignment(align_img, output_fig=False),
}
if self.bidirectional_decoder or self.double_decoder_consistency:
figures["alignment_backward"] = plot_alignment(alignments_backward[0].data.cpu().numpy(), output_fig=False)
# Sample audio
audio = ap.inv_melspectrogram(pred_spec.T)
return figures, {"audio": audio}
def train_log(
self, batch: dict, outputs: dict, logger: "Logger", assets: dict, steps: int
) -> None: # pylint: disable=no-self-use
"""Log training progress."""
ap = assets["audio_processor"]
figures, audios = self._create_logs(batch, outputs, ap)
logger.train_figures(steps, figures)
logger.train_audios(steps, audios, ap.sample_rate)
def eval_step(self, batch: dict, criterion: nn.Module):
return self.train_step(batch, criterion)
def eval_log(self, batch: dict, outputs: dict, logger: "Logger", assets: dict, steps: int) -> None:
ap = assets["audio_processor"]
figures, audios = self._create_logs(batch, outputs, ap)
logger.eval_figures(steps, figures)
logger.eval_audios(steps, audios, ap.sample_rate)
def __init__(
self,
config: "Tacotron2Config",
ap: "AudioProcessor" = None,
tokenizer: "TTSTokenizer" = None,
speaker_manager: SpeakerManager = None,
):
super().__init__(config, ap, tokenizer, speaker_manager)
self.decoder_output_dim = config.out_channels
# pass all config fields to `self`
# for fewer code change
for key in config:
setattr(self, key, config[key])
# init multi-speaker layers
if self.use_speaker_embedding or self.use_d_vector_file:
self.init_multispeaker(config)
self.decoder_in_features += self.embedded_speaker_dim # add speaker embedding dim
if self.use_gst:
self.decoder_in_features += self.gst.gst_embedding_dim
if self.use_capacitron_vae:
self.decoder_in_features += self.capacitron_vae.capacitron_VAE_embedding_dim
# embedding layer
self.embedding = nn.Embedding(self.num_chars, 512, padding_idx=0)
# base model layers
self.encoder = Encoder(self.encoder_in_features)
self.decoder = Decoder(
self.decoder_in_features,
self.decoder_output_dim,
self.r,
self.attention_type,
self.attention_win,
self.attention_norm,
self.prenet_type,
self.prenet_dropout,
self.use_forward_attn,
self.transition_agent,
self.forward_attn_mask,
self.location_attn,
self.attention_heads,
self.separate_stopnet,
self.max_decoder_steps,
)
self.postnet = Postnet(self.out_channels)
# setup prenet dropout
self.decoder.prenet.dropout_at_inference = self.prenet_dropout_at_inference
# global style token layers
if self.gst and self.use_gst:
self.gst_layer = GST(
num_mel=self.decoder_output_dim,
num_heads=self.gst.gst_num_heads,
num_style_tokens=self.gst.gst_num_style_tokens,
gst_embedding_dim=self.gst.gst_embedding_dim,
)
# Capacitron VAE Layers
if self.capacitron_vae and self.use_capacitron_vae:
self.capacitron_vae_layer = CapacitronVAE(
num_mel=self.decoder_output_dim,
encoder_output_dim=self.encoder_in_features,
capacitron_VAE_embedding_dim=self.capacitron_vae.
capacitron_VAE_embedding_dim,
speaker_embedding_dim=self.embedded_speaker_dim if
self.capacitron_vae.capacitron_use_speaker_embedding else None,
text_summary_embedding_dim=self.capacitron_vae.
capacitron_text_summary_embedding_dim
if self.capacitron_vae.capacitron_use_text_summary_embeddings
else None,
)
# backward pass decoder
if self.bidirectional_decoder:
self._init_backward_decoder()
# setup DDC
if self.double_decoder_consistency:
self.coarse_decoder = Decoder(
self.decoder_in_features,
self.decoder_output_dim,
self.ddc_r,
self.attention_type,
self.attention_win,
self.attention_norm,
self.prenet_type,
self.prenet_dropout,
self.use_forward_attn,
self.transition_agent,
self.forward_attn_mask,
self.location_attn,
self.attention_heads,
self.separate_stopnet,
self.max_decoder_steps,
)