# coding: utf-8
import numpy as np
import torch
from torch import nn
from TTS.tts.utils.measures import alignment_diagonal_score
from TTS.tts.utils.visual import plot_alignment, plot_spectrogram
from TTS.tts.layers.tacotron.gst_layers import GST
from TTS.tts.layers.tacotron.tacotron2 import Decoder, Encoder, Postnet
from TTS.tts.models.tacotron_abstract import TacotronAbstract
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.
use_gst (bool, optional): enable/disable Global style token module.
gst (Coqpit, optional): Coqpit to initialize the GST module. If `None`, GST is disabled. Defaults to None.
gradual_trainin (List): Gradual training schedule. If None or `[]`, no gradual training is used.
Defaults to `[]`.
"""
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,
use_gst=False,
gst=None,
gradual_training=[]):
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, use_gst, gst, gradual_training)
# 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 and use_gst:
self.gst_layer = GST(
num_mel=decoder_output_dim,
speaker_embedding_dim=speaker_embedding_dim,
num_heads=gst.gst_num_heads,
num_style_tokens=gst.gst_num_style_tokens,
gst_embedding_dim=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,
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,
cond_input=None):
"""
Shapes:
text: [B, T_in]
text_lengths: [B]
mel_specs: [B, T_out, C]
mel_lengths: [B]
cond_input: 'speaker_ids': [B, 1] and 'x_vectors':[B, C]
"""
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,
cond_input['x_vectors'])
if self.num_speakers > 1:
if not self.embeddings_per_sample:
# B x 1 x speaker_embed_dim
speaker_embeddings = self.speaker_embedding(cond_input['speaker_ids'])[:,
None]
else:
# B x 1 x speaker_embed_dim
speaker_embeddings = torch.unsqueeze(cond_input['x_vectors'], 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)
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({
'postnet_outputs': postnet_outputs,
'decoder_outputs': decoder_outputs,
'alignments': alignments,
'stop_tokens': stop_tokens
})
return outputs
@torch.no_grad()
def inference(self, text, cond_input=None):
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, cond_input['style_mel'],
cond_input['x_vectors'])
if self.num_speakers > 1:
if not self.embeddings_per_sample:
x_vector = self.speaker_embedding(cond_input['speaker_ids'])[:, None]
x_vector = torch.unsqueeze(x_vector, 0).transpose(1, 2)
else:
x_vector = cond_input
encoder_outputs = self._concat_speaker_embedding(
encoder_outputs, x_vector)
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 = {
'postnet_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']
speaker_ids = batch['speaker_ids']
x_vectors = batch['x_vectors']
# forward pass model
outputs = self.forward(text_input,
text_lengths,
mel_input,
mel_lengths,
cond_input={
'speaker_ids': speaker_ids,
'x_vectors': x_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
cond_input = {'speaker_ids': speaker_ids, 'x_vectors': x_vectors}
outputs = self.forward(text_input, text_lengths, mel_input,
mel_lengths, cond_input)
# compute loss
loss_dict = criterion(
outputs['model_outputs'],
outputs['decoder_outputs'],
mel_input,
linear_input,
outputs['stop_tokens'],
stop_targets,
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, batch, outputs):
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
train_audio = ap.inv_melspectrogram(pred_spec.T)
return figures, train_audio
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)