Source code for espnet2.tts.fastspeech.fastspeech

# Copyright 2020 Nagoya University (Tomoki Hayashi)
#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)

"""Fastspeech related modules for ESPnet2."""

import logging
from typing import Dict, Optional, Sequence, Tuple

import torch
import torch.nn.functional as F
from typeguard import check_argument_types

from espnet2.torch_utils.device_funcs import force_gatherable
from espnet2.torch_utils.initialize import initialize
from espnet2.tts.abs_tts import AbsTTS
from espnet2.tts.gst.style_encoder import StyleEncoder
from espnet.nets.pytorch_backend.conformer.encoder import Encoder as ConformerEncoder
from espnet.nets.pytorch_backend.e2e_tts_fastspeech import (
    FeedForwardTransformerLoss as FastSpeechLoss,
)
from espnet.nets.pytorch_backend.fastspeech.duration_predictor import DurationPredictor
from espnet.nets.pytorch_backend.fastspeech.length_regulator import LengthRegulator
from espnet.nets.pytorch_backend.nets_utils import make_non_pad_mask, make_pad_mask
from espnet.nets.pytorch_backend.tacotron2.decoder import Postnet
from espnet.nets.pytorch_backend.transformer.embedding import (
    PositionalEncoding,
    ScaledPositionalEncoding,
)
from espnet.nets.pytorch_backend.transformer.encoder import (
    Encoder as TransformerEncoder,
)


[docs]class FastSpeech(AbsTTS): """FastSpeech module for end-to-end text-to-speech. This is a module of FastSpeech, feed-forward Transformer with duration predictor described in `FastSpeech: Fast, Robust and Controllable Text to Speech`_, which does not require any auto-regressive processing during inference, resulting in fast decoding compared with auto-regressive Transformer. .. _`FastSpeech: Fast, Robust and Controllable Text to Speech`: https://arxiv.org/pdf/1905.09263.pdf """ def __init__( self, # network structure related idim: int, odim: int, adim: int = 384, aheads: int = 4, elayers: int = 6, eunits: int = 1536, dlayers: int = 6, dunits: int = 1536, postnet_layers: int = 5, postnet_chans: int = 512, postnet_filts: int = 5, postnet_dropout_rate: float = 0.5, positionwise_layer_type: str = "conv1d", positionwise_conv_kernel_size: int = 1, use_scaled_pos_enc: bool = True, use_batch_norm: bool = True, encoder_normalize_before: bool = True, decoder_normalize_before: bool = True, encoder_concat_after: bool = False, decoder_concat_after: bool = False, duration_predictor_layers: int = 2, duration_predictor_chans: int = 384, duration_predictor_kernel_size: int = 3, duration_predictor_dropout_rate: float = 0.1, reduction_factor: int = 1, encoder_type: str = "transformer", decoder_type: str = "transformer", transformer_enc_dropout_rate: float = 0.1, transformer_enc_positional_dropout_rate: float = 0.1, transformer_enc_attn_dropout_rate: float = 0.1, transformer_dec_dropout_rate: float = 0.1, transformer_dec_positional_dropout_rate: float = 0.1, transformer_dec_attn_dropout_rate: float = 0.1, # only for conformer conformer_rel_pos_type: str = "legacy", conformer_pos_enc_layer_type: str = "rel_pos", conformer_self_attn_layer_type: str = "rel_selfattn", conformer_activation_type: str = "swish", use_macaron_style_in_conformer: bool = True, use_cnn_in_conformer: bool = True, conformer_enc_kernel_size: int = 7, conformer_dec_kernel_size: int = 31, zero_triu: bool = False, # extra embedding related spks: Optional[int] = None, langs: Optional[int] = None, spk_embed_dim: Optional[int] = None, spk_embed_integration_type: str = "add", use_gst: bool = False, gst_tokens: int = 10, gst_heads: int = 4, gst_conv_layers: int = 6, gst_conv_chans_list: Sequence[int] = (32, 32, 64, 64, 128, 128), gst_conv_kernel_size: int = 3, gst_conv_stride: int = 2, gst_gru_layers: int = 1, gst_gru_units: int = 128, # training related init_type: str = "xavier_uniform", init_enc_alpha: float = 1.0, init_dec_alpha: float = 1.0, use_masking: bool = False, use_weighted_masking: bool = False, ): """Initialize FastSpeech module. Args: idim (int): Dimension of the inputs. odim (int): Dimension of the outputs. elayers (int): Number of encoder layers. eunits (int): Number of encoder hidden units. dlayers (int): Number of decoder layers. dunits (int): Number of decoder hidden units. postnet_layers (int): Number of postnet layers. postnet_chans (int): Number of postnet channels. postnet_filts (int): Kernel size of postnet. postnet_dropout_rate (float): Dropout rate in postnet. use_scaled_pos_enc (bool): Whether to use trainable scaled pos encoding. use_batch_norm (bool): Whether to use batch normalization in encoder prenet. encoder_normalize_before (bool): Whether to apply layernorm layer before encoder block. decoder_normalize_before (bool): Whether to apply layernorm layer before decoder block. encoder_concat_after (bool): Whether to concatenate attention layer's input and output in encoder. decoder_concat_after (bool): Whether to concatenate attention layer's input and output in decoder. duration_predictor_layers (int): Number of duration predictor layers. duration_predictor_chans (int): Number of duration predictor channels. duration_predictor_kernel_size (int): Kernel size of duration predictor. duration_predictor_dropout_rate (float): Dropout rate in duration predictor. reduction_factor (int): Reduction factor. encoder_type (str): Encoder type ("transformer" or "conformer"). decoder_type (str): Decoder type ("transformer" or "conformer"). transformer_enc_dropout_rate (float): Dropout rate in encoder except attention and positional encoding. transformer_enc_positional_dropout_rate (float): Dropout rate after encoder positional encoding. transformer_enc_attn_dropout_rate (float): Dropout rate in encoder self-attention module. transformer_dec_dropout_rate (float): Dropout rate in decoder except attention & positional encoding. transformer_dec_positional_dropout_rate (float): Dropout rate after decoder positional encoding. transformer_dec_attn_dropout_rate (float): Dropout rate in decoder self-attention module. conformer_rel_pos_type (str): Relative pos encoding type in conformer. conformer_pos_enc_layer_type (str): Pos encoding layer type in conformer. conformer_self_attn_layer_type (str): Self-attention layer type in conformer conformer_activation_type (str): Activation function type in conformer. use_macaron_style_in_conformer: Whether to use macaron style FFN. use_cnn_in_conformer: Whether to use CNN in conformer. conformer_enc_kernel_size: Kernel size of encoder conformer. conformer_dec_kernel_size: Kernel size of decoder conformer. zero_triu: Whether to use zero triu in relative self-attention module. spks (Optional[int]): Number of speakers. If set to > 1, assume that the sids will be provided as the input and use sid embedding layer. langs (Optional[int]): Number of languages. If set to > 1, assume that the lids will be provided as the input and use sid embedding layer. spk_embed_dim (Optional[int]): Speaker embedding dimension. If set to > 0, assume that spembs will be provided as the input. spk_embed_integration_type: How to integrate speaker embedding. use_gst (str): Whether to use global style token. gst_tokens (int): The number of GST embeddings. gst_heads (int): The number of heads in GST multihead attention. gst_conv_layers (int): The number of conv layers in GST. gst_conv_chans_list: (Sequence[int]): List of the number of channels of conv layers in GST. gst_conv_kernel_size (int): Kernel size of conv layers in GST. gst_conv_stride (int): Stride size of conv layers in GST. gst_gru_layers (int): The number of GRU layers in GST. gst_gru_units (int): The number of GRU units in GST. init_type (str): How to initialize transformer parameters. init_enc_alpha (float): Initial value of alpha in scaled pos encoding of the encoder. init_dec_alpha (float): Initial value of alpha in scaled pos encoding of the decoder. use_masking (bool): Whether to apply masking for padded part in loss calculation. use_weighted_masking (bool): Whether to apply weighted masking in loss calculation. """ assert check_argument_types() super().__init__() # store hyperparameters self.idim = idim self.odim = odim self.eos = idim - 1 self.reduction_factor = reduction_factor self.encoder_type = encoder_type self.decoder_type = decoder_type self.use_scaled_pos_enc = use_scaled_pos_enc self.use_gst = use_gst # use idx 0 as padding idx self.padding_idx = 0 # get positional encoding class pos_enc_class = ( ScaledPositionalEncoding if self.use_scaled_pos_enc else PositionalEncoding ) # check relative positional encoding compatibility if "conformer" in [encoder_type, decoder_type]: if conformer_rel_pos_type == "legacy": if conformer_pos_enc_layer_type == "rel_pos": conformer_pos_enc_layer_type = "legacy_rel_pos" logging.warning( "Fallback to conformer_pos_enc_layer_type = 'legacy_rel_pos' " "due to the compatibility. If you want to use the new one, " "please use conformer_pos_enc_layer_type = 'latest'." ) if conformer_self_attn_layer_type == "rel_selfattn": conformer_self_attn_layer_type = "legacy_rel_selfattn" logging.warning( "Fallback to " "conformer_self_attn_layer_type = 'legacy_rel_selfattn' " "due to the compatibility. If you want to use the new one, " "please use conformer_pos_enc_layer_type = 'latest'." ) elif conformer_rel_pos_type == "latest": assert conformer_pos_enc_layer_type != "legacy_rel_pos" assert conformer_self_attn_layer_type != "legacy_rel_selfattn" else: raise ValueError(f"Unknown rel_pos_type: {conformer_rel_pos_type}") # define encoder encoder_input_layer = torch.nn.Embedding( num_embeddings=idim, embedding_dim=adim, padding_idx=self.padding_idx ) if encoder_type == "transformer": self.encoder = TransformerEncoder( idim=idim, attention_dim=adim, attention_heads=aheads, linear_units=eunits, num_blocks=elayers, input_layer=encoder_input_layer, dropout_rate=transformer_enc_dropout_rate, positional_dropout_rate=transformer_enc_positional_dropout_rate, attention_dropout_rate=transformer_enc_attn_dropout_rate, pos_enc_class=pos_enc_class, normalize_before=encoder_normalize_before, concat_after=encoder_concat_after, positionwise_layer_type=positionwise_layer_type, positionwise_conv_kernel_size=positionwise_conv_kernel_size, ) elif encoder_type == "conformer": self.encoder = ConformerEncoder( idim=idim, attention_dim=adim, attention_heads=aheads, linear_units=eunits, num_blocks=elayers, input_layer=encoder_input_layer, dropout_rate=transformer_enc_dropout_rate, positional_dropout_rate=transformer_enc_positional_dropout_rate, attention_dropout_rate=transformer_enc_attn_dropout_rate, normalize_before=encoder_normalize_before, concat_after=encoder_concat_after, positionwise_layer_type=positionwise_layer_type, positionwise_conv_kernel_size=positionwise_conv_kernel_size, macaron_style=use_macaron_style_in_conformer, pos_enc_layer_type=conformer_pos_enc_layer_type, selfattention_layer_type=conformer_self_attn_layer_type, activation_type=conformer_activation_type, use_cnn_module=use_cnn_in_conformer, cnn_module_kernel=conformer_enc_kernel_size, ) else: raise ValueError(f"{encoder_type} is not supported.") # define GST if self.use_gst: self.gst = StyleEncoder( idim=odim, # the input is mel-spectrogram gst_tokens=gst_tokens, gst_token_dim=adim, gst_heads=gst_heads, conv_layers=gst_conv_layers, conv_chans_list=gst_conv_chans_list, conv_kernel_size=gst_conv_kernel_size, conv_stride=gst_conv_stride, gru_layers=gst_gru_layers, gru_units=gst_gru_units, ) # define spk and lang embedding self.spks = None if spks is not None and spks > 1: self.spks = spks self.sid_emb = torch.nn.Embedding(spks, adim) self.langs = None if langs is not None and langs > 1: self.langs = langs self.lid_emb = torch.nn.Embedding(langs, adim) # define additional projection for speaker embedding self.spk_embed_dim = None if spk_embed_dim is not None and spk_embed_dim > 0: self.spk_embed_dim = spk_embed_dim self.spk_embed_integration_type = spk_embed_integration_type if self.spk_embed_dim is not None: if self.spk_embed_integration_type == "add": self.projection = torch.nn.Linear(self.spk_embed_dim, adim) else: self.projection = torch.nn.Linear(adim + self.spk_embed_dim, adim) # define duration predictor self.duration_predictor = DurationPredictor( idim=adim, n_layers=duration_predictor_layers, n_chans=duration_predictor_chans, kernel_size=duration_predictor_kernel_size, dropout_rate=duration_predictor_dropout_rate, ) # define length regulator self.length_regulator = LengthRegulator() # define decoder # NOTE: we use encoder as decoder # because fastspeech's decoder is the same as encoder if decoder_type == "transformer": self.decoder = TransformerEncoder( idim=0, attention_dim=adim, attention_heads=aheads, linear_units=dunits, num_blocks=dlayers, input_layer=None, dropout_rate=transformer_dec_dropout_rate, positional_dropout_rate=transformer_dec_positional_dropout_rate, attention_dropout_rate=transformer_dec_attn_dropout_rate, pos_enc_class=pos_enc_class, normalize_before=decoder_normalize_before, concat_after=decoder_concat_after, positionwise_layer_type=positionwise_layer_type, positionwise_conv_kernel_size=positionwise_conv_kernel_size, ) elif decoder_type == "conformer": self.decoder = ConformerEncoder( idim=0, attention_dim=adim, attention_heads=aheads, linear_units=dunits, num_blocks=dlayers, input_layer=None, dropout_rate=transformer_dec_dropout_rate, positional_dropout_rate=transformer_dec_positional_dropout_rate, attention_dropout_rate=transformer_dec_attn_dropout_rate, normalize_before=decoder_normalize_before, concat_after=decoder_concat_after, positionwise_layer_type=positionwise_layer_type, positionwise_conv_kernel_size=positionwise_conv_kernel_size, macaron_style=use_macaron_style_in_conformer, pos_enc_layer_type=conformer_pos_enc_layer_type, selfattention_layer_type=conformer_self_attn_layer_type, activation_type=conformer_activation_type, use_cnn_module=use_cnn_in_conformer, cnn_module_kernel=conformer_dec_kernel_size, ) else: raise ValueError(f"{decoder_type} is not supported.") # define final projection self.feat_out = torch.nn.Linear(adim, odim * reduction_factor) # define postnet self.postnet = ( None if postnet_layers == 0 else Postnet( idim=idim, odim=odim, n_layers=postnet_layers, n_chans=postnet_chans, n_filts=postnet_filts, use_batch_norm=use_batch_norm, dropout_rate=postnet_dropout_rate, ) ) # initialize parameters self._reset_parameters( init_type=init_type, init_enc_alpha=init_enc_alpha, init_dec_alpha=init_dec_alpha, ) # define criterions self.criterion = FastSpeechLoss( use_masking=use_masking, use_weighted_masking=use_weighted_masking ) def _forward( self, xs: torch.Tensor, ilens: torch.Tensor, ys: Optional[torch.Tensor] = None, olens: Optional[torch.Tensor] = None, ds: Optional[torch.Tensor] = None, spembs: Optional[torch.Tensor] = None, sids: Optional[torch.Tensor] = None, lids: Optional[torch.Tensor] = None, is_inference: bool = False, alpha: float = 1.0, ) -> Sequence[torch.Tensor]: # forward encoder x_masks = self._source_mask(ilens) hs, _ = self.encoder(xs, x_masks) # (B, T_text, adim) # integrate with GST if self.use_gst: style_embs = self.gst(ys) hs = hs + style_embs.unsqueeze(1) # integrate with SID and LID embeddings if self.spks is not None: sid_embs = self.sid_emb(sids.view(-1)) hs = hs + sid_embs.unsqueeze(1) if self.langs is not None: lid_embs = self.lid_emb(lids.view(-1)) hs = hs + lid_embs.unsqueeze(1) # integrate speaker embedding if self.spk_embed_dim is not None: hs = self._integrate_with_spk_embed(hs, spembs) # forward duration predictor and length regulator d_masks = make_pad_mask(ilens).to(xs.device) if is_inference: d_outs = self.duration_predictor.inference(hs, d_masks) # (B, T_text) hs = self.length_regulator(hs, d_outs, alpha) # (B, T_feats, adim) else: d_outs = self.duration_predictor(hs, d_masks) # (B, T_text) hs = self.length_regulator(hs, ds) # (B, T_feats, adim) # forward decoder if olens is not None and not is_inference: if self.reduction_factor > 1: olens_in = olens.new([olen // self.reduction_factor for olen in olens]) else: olens_in = olens h_masks = self._source_mask(olens_in) else: h_masks = None zs, _ = self.decoder(hs, h_masks) # (B, T_feats, adim) before_outs = self.feat_out(zs).view( zs.size(0), -1, self.odim ) # (B, T_feats, odim) # postnet -> (B, T_feats//r * r, odim) if self.postnet is None: after_outs = before_outs else: after_outs = before_outs + self.postnet( before_outs.transpose(1, 2) ).transpose(1, 2) return before_outs, after_outs, d_outs
[docs] def forward( self, text: torch.Tensor, text_lengths: torch.Tensor, feats: torch.Tensor, feats_lengths: torch.Tensor, durations: torch.Tensor, durations_lengths: torch.Tensor, spembs: Optional[torch.Tensor] = None, sids: Optional[torch.Tensor] = None, lids: Optional[torch.Tensor] = None, joint_training: bool = False, ) -> Tuple[torch.Tensor, Dict[str, torch.Tensor], torch.Tensor]: """Calculate forward propagation. Args: text (LongTensor): Batch of padded character ids (B, T_text). text_lengths (LongTensor): Batch of lengths of each input (B,). feats (Tensor): Batch of padded target features (B, T_feats, odim). feats_lengths (LongTensor): Batch of the lengths of each target (B,). durations (LongTensor): Batch of padded durations (B, T_text + 1). durations_lengths (LongTensor): Batch of duration lengths (B, T_text + 1). spembs (Optional[Tensor]): Batch of speaker embeddings (B, spk_embed_dim). sids (Optional[Tensor]): Batch of speaker IDs (B, 1). lids (Optional[Tensor]): Batch of language IDs (B, 1). joint_training (bool): Whether to perform joint training with vocoder. Returns: Tensor: Loss scalar value. Dict: Statistics to be monitored. Tensor: Weight value if not joint training else model outputs. """ text = text[:, : text_lengths.max()] # for data-parallel feats = feats[:, : feats_lengths.max()] # for data-parallel durations = durations[:, : durations_lengths.max()] # for data-parallel batch_size = text.size(0) # Add eos at the last of sequence xs = F.pad(text, [0, 1], "constant", self.padding_idx) for i, l in enumerate(text_lengths): xs[i, l] = self.eos ilens = text_lengths + 1 ys, ds = feats, durations olens = feats_lengths # forward propagation before_outs, after_outs, d_outs = self._forward( xs, ilens, ys, olens, ds, spembs=spembs, sids=sids, lids=lids, is_inference=False, ) # modifiy mod part of groundtruth if self.reduction_factor > 1: olens = olens.new([olen - olen % self.reduction_factor for olen in olens]) max_olen = max(olens) ys = ys[:, :max_olen] # calculate loss if self.postnet is None: after_outs = None l1_loss, duration_loss = self.criterion( after_outs, before_outs, d_outs, ys, ds, ilens, olens ) loss = l1_loss + duration_loss stats = dict( l1_loss=l1_loss.item(), duration_loss=duration_loss.item(), ) # report extra information if self.encoder_type == "transformer" and self.use_scaled_pos_enc: stats.update( encoder_alpha=self.encoder.embed[-1].alpha.data.item(), ) if self.decoder_type == "transformer" and self.use_scaled_pos_enc: stats.update( decoder_alpha=self.decoder.embed[-1].alpha.data.item(), ) if not joint_training: stats.update(loss=loss.item()) loss, stats, weight = force_gatherable( (loss, stats, batch_size), loss.device ) return loss, stats, weight else: return loss, stats, after_outs if after_outs is not None else before_outs
[docs] def inference( self, text: torch.Tensor, feats: Optional[torch.Tensor] = None, durations: Optional[torch.Tensor] = None, spembs: Optional[torch.Tensor] = None, sids: Optional[torch.Tensor] = None, lids: Optional[torch.Tensor] = None, alpha: float = 1.0, use_teacher_forcing: bool = False, ) -> Dict[str, torch.Tensor]: """Generate the sequence of features given the sequences of characters. Args: text (LongTensor): Input sequence of characters (T_text,). feats (Optional[Tensor]): Feature sequence to extract style (N, idim). durations (Optional[LongTensor]): Groundtruth of duration (T_text + 1,). spembs (Optional[Tensor]): Speaker embedding (spk_embed_dim,). sids (Optional[Tensor]): Speaker ID (1,). lids (Optional[Tensor]): Language ID (1,). alpha (float): Alpha to control the speed. use_teacher_forcing (bool): Whether to use teacher forcing. If true, groundtruth of duration, pitch and energy will be used. Returns: Dict[str, Tensor]: Output dict including the following items: * feat_gen (Tensor): Output sequence of features (T_feats, odim). * duration (Tensor): Duration sequence (T_text + 1,). """ x, y = text, feats spemb, d = spembs, durations # add eos at the last of sequence x = F.pad(x, [0, 1], "constant", self.eos) # setup batch axis ilens = torch.tensor([x.shape[0]], dtype=torch.long, device=x.device) xs, ys = x.unsqueeze(0), None if y is not None: ys = y.unsqueeze(0) if spemb is not None: spembs = spemb.unsqueeze(0) if use_teacher_forcing: # use groundtruth of duration ds = d.unsqueeze(0) _, outs, d_outs = self._forward( xs, ilens, ys, ds=ds, spembs=spembs, sids=sids, lids=lids, ) # (1, T_feats, odim) else: # inference _, outs, d_outs = self._forward( xs, ilens, ys, spembs=spembs, sids=sids, lids=lids, is_inference=True, alpha=alpha, ) # (1, T_feats, odim) return dict(feat_gen=outs[0], duration=d_outs[0])
def _integrate_with_spk_embed( self, hs: torch.Tensor, spembs: torch.Tensor ) -> torch.Tensor: """Integrate speaker embedding with hidden states. Args: hs (Tensor): Batch of hidden state sequences (B, T_text, adim). spembs (Tensor): Batch of speaker embeddings (B, spk_embed_dim). Returns: Tensor: Batch of integrated hidden state sequences (B, T_text, adim). """ if self.spk_embed_integration_type == "add": # apply projection and then add to hidden states spembs = self.projection(F.normalize(spembs)) hs = hs + spembs.unsqueeze(1) elif self.spk_embed_integration_type == "concat": # concat hidden states with spk embeds and then apply projection spembs = F.normalize(spembs).unsqueeze(1).expand(-1, hs.size(1), -1) hs = self.projection(torch.cat([hs, spembs], dim=-1)) else: raise NotImplementedError("support only add or concat.") return hs def _source_mask(self, ilens: torch.Tensor) -> torch.Tensor: """Make masks for self-attention. Args: ilens (LongTensor): Batch of lengths (B,). Returns: Tensor: Mask tensor for self-attention. dtype=torch.uint8 in PyTorch 1.2- dtype=torch.bool in PyTorch 1.2+ (including 1.2) Examples: >>> ilens = [5, 3] >>> self._source_mask(ilens) tensor([[[1, 1, 1, 1, 1], [1, 1, 1, 0, 0]]], dtype=torch.uint8) """ x_masks = make_non_pad_mask(ilens).to(next(self.parameters()).device) return x_masks.unsqueeze(-2) def _reset_parameters( self, init_type: str, init_enc_alpha: float, init_dec_alpha: float ): # initialize parameters if init_type != "pytorch": initialize(self, init_type) # initialize alpha in scaled positional encoding if self.encoder_type == "transformer" and self.use_scaled_pos_enc: self.encoder.embed[-1].alpha.data = torch.tensor(init_enc_alpha) if self.decoder_type == "transformer" and self.use_scaled_pos_enc: self.decoder.embed[-1].alpha.data = torch.tensor(init_dec_alpha)