espnet2/tts/fastspeech2.py

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

"""Fastspeech2 related modules for ESPnet2."""

import logging

from typing import Dict
from typing import Sequence
from typing import Tuple

import torch
import torch.nn.functional as F

from typeguard import check_argument_types

from espnet.nets.pytorch_backend.conformer.encoder import (
    Encoder as ConformerEncoder,  # noqa: H301
)
from espnet.nets.pytorch_backend.fastspeech.duration_predictor import DurationPredictor
from espnet.nets.pytorch_backend.fastspeech.duration_predictor import (
    DurationPredictorLoss,  # noqa: H301
)
from espnet.nets.pytorch_backend.fastspeech.length_regulator import LengthRegulator
from espnet.nets.pytorch_backend.nets_utils import make_non_pad_mask
from espnet.nets.pytorch_backend.nets_utils import make_pad_mask
from espnet.nets.pytorch_backend.tacotron2.decoder import Postnet
from espnet.nets.pytorch_backend.transformer.embedding import PositionalEncoding
from espnet.nets.pytorch_backend.transformer.embedding import ScaledPositionalEncoding
from espnet.nets.pytorch_backend.transformer.encoder import (
    Encoder as TransformerEncoder,  # noqa: H301
)

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 espnet2.tts.variance_predictor import VariancePredictor


class FastSpeech2(AbsTTS):
    """FastSpeech2 module.

    This is a module of FastSpeech2 described in `FastSpeech 2: Fast and
    High-Quality End-to-End Text to Speech`_. Instead of quantized pitch and
    energy, we use token-averaged value introduced in `FastPitch: Parallel
    Text-to-speech with Pitch Prediction`_.

    .. _`FastSpeech 2: Fast and High-Quality End-to-End Text to Speech`:
        https://arxiv.org/abs/2006.04558
    .. _`FastPitch: Parallel Text-to-speech with Pitch Prediction`:
        https://arxiv.org/abs/2006.06873

    """

    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,
        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,
        reduction_factor: int = 1,
        encoder_type: str = "transformer",
        decoder_type: str = "transformer",
        # 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,
        zero_triu: bool = False,
        conformer_enc_kernel_size: int = 7,
        conformer_dec_kernel_size: int = 31,
        # duration predictor
        duration_predictor_layers: int = 2,
        duration_predictor_chans: int = 384,
        duration_predictor_kernel_size: int = 3,
        # energy predictor
        energy_predictor_layers: int = 2,
        energy_predictor_chans: int = 384,
        energy_predictor_kernel_size: int = 3,
        energy_predictor_dropout: float = 0.5,
        energy_embed_kernel_size: int = 9,
        energy_embed_dropout: float = 0.5,
        stop_gradient_from_energy_predictor: bool = False,
        # pitch predictor
        pitch_predictor_layers: int = 2,
        pitch_predictor_chans: int = 384,
        pitch_predictor_kernel_size: int = 3,
        pitch_predictor_dropout: float = 0.5,
        pitch_embed_kernel_size: int = 9,
        pitch_embed_dropout: float = 0.5,
        stop_gradient_from_pitch_predictor: bool = False,
        # pretrained spk emb
        spk_embed_dim: int = None,
        spk_embed_integration_type: str = "add",
        # GST
        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
        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,
        duration_predictor_dropout_rate: float = 0.1,
        postnet_dropout_rate: float = 0.5,
        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 FastSpeech2 module."""
        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.stop_gradient_from_pitch_predictor = stop_gradient_from_pitch_predictor
        self.stop_gradient_from_energy_predictor = stop_gradient_from_energy_predictor
        self.use_scaled_pos_enc = use_scaled_pos_enc
        self.use_gst = use_gst
        self.spk_embed_dim = spk_embed_dim
        if self.spk_embed_dim is not None:
            self.spk_embed_integration_type = spk_embed_integration_type

        # 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,
                zero_triu=zero_triu,
            )
        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 additional projection for speaker embedding
        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 pitch predictor
        self.pitch_predictor = VariancePredictor(
            idim=adim,
            n_layers=pitch_predictor_layers,
            n_chans=pitch_predictor_chans,
            kernel_size=pitch_predictor_kernel_size,
            dropout_rate=pitch_predictor_dropout,
        )
        # NOTE(kan-bayashi): We use continuous pitch + FastPitch style avg
        self.pitch_embed = torch.nn.Sequential(
            torch.nn.Conv1d(
                in_channels=1,
                out_channels=adim,
                kernel_size=pitch_embed_kernel_size,
                padding=(pitch_embed_kernel_size - 1) // 2,
            ),
            torch.nn.Dropout(pitch_embed_dropout),
        )

        # define energy predictor
        self.energy_predictor = VariancePredictor(
            idim=adim,
            n_layers=energy_predictor_layers,
            n_chans=energy_predictor_chans,
            kernel_size=energy_predictor_kernel_size,
            dropout_rate=energy_predictor_dropout,
        )
        # NOTE(kan-bayashi): We use continuous enegy + FastPitch style avg
        self.energy_embed = torch.nn.Sequential(
            torch.nn.Conv1d(
                in_channels=1,
                out_channels=adim,
                kernel_size=energy_embed_kernel_size,
                padding=(energy_embed_kernel_size - 1) // 2,
            ),
            torch.nn.Dropout(energy_embed_dropout),
        )

        # 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 = FastSpeech2Loss(
            use_masking=use_masking, use_weighted_masking=use_weighted_masking
        )

    def forward(
        self,
        text: torch.Tensor,
        text_lengths: torch.Tensor,
        speech: torch.Tensor,
        speech_lengths: torch.Tensor,
        durations: torch.Tensor,
        durations_lengths: torch.Tensor,
        pitch: torch.Tensor,
        pitch_lengths: torch.Tensor,
        energy: torch.Tensor,
        energy_lengths: torch.Tensor,
        spembs: torch.Tensor = None,
    ) -> Tuple[torch.Tensor, Dict[str, torch.Tensor], torch.Tensor]:
        """Calculate forward propagation.

        Args:
            text (LongTensor): Batch of padded token ids (B, Tmax).
            text_lengths (LongTensor): Batch of lengths of each input (B,).
            speech (Tensor): Batch of padded target features (B, Lmax, odim).
            speech_lengths (LongTensor): Batch of the lengths of each target (B,).
            durations (LongTensor): Batch of padded durations (B, Tmax + 1).
            durations_lengths (LongTensor): Batch of duration lengths (B, Tmax + 1).
            pitch (Tensor): Batch of padded token-averaged pitch (B, Tmax + 1, 1).
            pitch_lengths (LongTensor): Batch of pitch lengths (B, Tmax + 1).
            energy (Tensor): Batch of padded token-averaged energy (B, Tmax + 1, 1).
            energy_lengths (LongTensor): Batch of energy lengths (B, Tmax + 1).
            spembs (Tensor, optional): Batch of speaker embeddings (B, spk_embed_dim).

        Returns:
            Tensor: Loss scalar value.
            Dict: Statistics to be monitored.
            Tensor: Weight value.

        """
        text = text[:, : text_lengths.max()]  # for data-parallel
        speech = speech[:, : speech_lengths.max()]  # for data-parallel
        durations = durations[:, : durations_lengths.max()]  # for data-parallel
        pitch = pitch[:, : pitch_lengths.max()]  # for data-parallel
        energy = energy[:, : energy_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, ps, es = speech, durations, pitch, energy
        olens = speech_lengths

        # forward propagation
        before_outs, after_outs, d_outs, p_outs, e_outs = self._forward(
            xs, ilens, ys, olens, ds, ps, es, spembs=spembs, is_inference=False
        )

        # modify 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

        # calculate loss
        l1_loss, duration_loss, pitch_loss, energy_loss = self.criterion(
            after_outs=after_outs,
            before_outs=before_outs,
            d_outs=d_outs,
            p_outs=p_outs,
            e_outs=e_outs,
            ys=ys,
            ds=ds,
            ps=ps,
            es=es,
            ilens=ilens,
            olens=olens,
        )
        loss = l1_loss + duration_loss + pitch_loss + energy_loss

        stats = dict(
            l1_loss=l1_loss.item(),
            duration_loss=duration_loss.item(),
            pitch_loss=pitch_loss.item(),
            energy_loss=energy_loss.item(),
            loss=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(),
            )

        loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
        return loss, stats, weight

    def _forward(
        self,
        xs: torch.Tensor,
        ilens: torch.Tensor,
        ys: torch.Tensor = None,
        olens: torch.Tensor = None,
        ds: torch.Tensor = None,
        ps: torch.Tensor = None,
        es: torch.Tensor = None,
        spembs: 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, Tmax, adim)

        # integrate with GST
        if self.use_gst:
            style_embs = self.gst(ys)
            hs = hs + style_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 variance predictors
        d_masks = make_pad_mask(ilens).to(xs.device)

        if self.stop_gradient_from_pitch_predictor:
            p_outs = self.pitch_predictor(hs.detach(), d_masks.unsqueeze(-1))
        else:
            p_outs = self.pitch_predictor(hs, d_masks.unsqueeze(-1))
        if self.stop_gradient_from_energy_predictor:
            e_outs = self.energy_predictor(hs.detach(), d_masks.unsqueeze(-1))
        else:
            e_outs = self.energy_predictor(hs, d_masks.unsqueeze(-1))

        if is_inference:
            d_outs = self.duration_predictor.inference(hs, d_masks)  # (B, Tmax)
            # use prediction in inference
            p_embs = self.pitch_embed(p_outs.transpose(1, 2)).transpose(1, 2)
            e_embs = self.energy_embed(e_outs.transpose(1, 2)).transpose(1, 2)
            hs = hs + e_embs + p_embs
            hs = self.length_regulator(hs, d_outs, alpha)  # (B, Lmax, adim)
        else:
            d_outs = self.duration_predictor(hs, d_masks)
            # use groundtruth in training
            p_embs = self.pitch_embed(ps.transpose(1, 2)).transpose(1, 2)
            e_embs = self.energy_embed(es.transpose(1, 2)).transpose(1, 2)
            hs = hs + e_embs + p_embs
            hs = self.length_regulator(hs, ds)  # (B, Lmax, 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, Lmax, adim)
        before_outs = self.feat_out(zs).view(
            zs.size(0), -1, self.odim
        )  # (B, Lmax, odim)

        # postnet -> (B, Lmax//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, p_outs, e_outs

    def inference(
        self,
        text: torch.Tensor,
        speech: torch.Tensor = None,
        spembs: torch.Tensor = None,
        durations: torch.Tensor = None,
        pitch: torch.Tensor = None,
        energy: torch.Tensor = None,
        alpha: float = 1.0,
        use_teacher_forcing: bool = False,
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """Generate the sequence of features given the sequences of characters.

        Args:
            text (LongTensor): Input sequence of characters (T,).
            speech (Tensor, optional): Feature sequence to extract style (N, idim).
            spembs (Tensor, optional): Speaker embedding vector (spk_embed_dim,).
            durations (LongTensor, optional): Groundtruth of duration (T + 1,).
            pitch (Tensor, optional): Groundtruth of token-averaged pitch (T + 1, 1).
            energy (Tensor, optional): Groundtruth of token-averaged energy (T + 1, 1).
            alpha (float, optional): Alpha to control the speed.
            use_teacher_forcing (bool, optional): Whether to use teacher forcing.
                If true, groundtruth of duration, pitch and energy will be used.

        Returns:
            Tensor: Output sequence of features (L, odim).
            None: Dummy for compatibility.
            None: Dummy for compatibility.

        """
        x, y = text, speech
        spemb, d, p, e = spembs, durations, pitch, energy

        # 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, pitch, and energy
            ds, ps, es = d.unsqueeze(0), p.unsqueeze(0), e.unsqueeze(0)
            _, outs, *_ = self._forward(
                xs,
                ilens,
                ys,
                ds=ds,
                ps=ps,
                es=es,
                spembs=spembs,
            )  # (1, L, odim)
        else:
            _, outs, *_ = self._forward(
                xs,
                ilens,
                ys,
                spembs=spembs,
                is_inference=True,
                alpha=alpha,
            )  # (1, L, odim)

        return outs[0], None, None

    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, Tmax, adim).
            spembs (Tensor): Batch of speaker embeddings (B, spk_embed_dim).

        Returns:
            Tensor: Batch of integrated hidden state sequences (B, Tmax, 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)


class FastSpeech2Loss(torch.nn.Module):
    """Loss function module for FastSpeech2."""

    def __init__(self, use_masking: bool = True, use_weighted_masking: bool = False):
        """Initialize feed-forward Transformer loss module.

        Args:
            use_masking (bool):
                Whether to apply masking for padded part in loss calculation.
            use_weighted_masking (bool):
                Whether to weighted masking in loss calculation.

        """
        assert check_argument_types()
        super().__init__()

        assert (use_masking != use_weighted_masking) or not use_masking
        self.use_masking = use_masking
        self.use_weighted_masking = use_weighted_masking

        # define criterions
        reduction = "none" if self.use_weighted_masking else "mean"
        self.l1_criterion = torch.nn.L1Loss(reduction=reduction)
        self.mse_criterion = torch.nn.MSELoss(reduction=reduction)
        self.duration_criterion = DurationPredictorLoss(reduction=reduction)

    def forward(
        self,
        after_outs: torch.Tensor,
        before_outs: torch.Tensor,
        d_outs: torch.Tensor,
        p_outs: torch.Tensor,
        e_outs: torch.Tensor,
        ys: torch.Tensor,
        ds: torch.Tensor,
        ps: torch.Tensor,
        es: torch.Tensor,
        ilens: torch.Tensor,
        olens: torch.Tensor,
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
        """Calculate forward propagation.

        Args:
            after_outs (Tensor): Batch of outputs after postnets (B, Lmax, odim).
            before_outs (Tensor): Batch of outputs before postnets (B, Lmax, odim).
            d_outs (LongTensor): Batch of outputs of duration predictor (B, Tmax).
            p_outs (Tensor): Batch of outputs of pitch predictor (B, Tmax, 1).
            e_outs (Tensor): Batch of outputs of energy predictor (B, Tmax, 1).
            ys (Tensor): Batch of target features (B, Lmax, odim).
            ds (LongTensor): Batch of durations (B, Tmax).
            ps (Tensor): Batch of target token-averaged pitch (B, Tmax, 1).
            es (Tensor): Batch of target token-averaged energy (B, Tmax, 1).
            ilens (LongTensor): Batch of the lengths of each input (B,).
            olens (LongTensor): Batch of the lengths of each target (B,).

        Returns:
            Tensor: L1 loss value.
            Tensor: Duration predictor loss value.
            Tensor: Pitch predictor loss value.
            Tensor: Energy predictor loss value.

        """
        # apply mask to remove padded part
        if self.use_masking:
            out_masks = make_non_pad_mask(olens).unsqueeze(-1).to(ys.device)
            before_outs = before_outs.masked_select(out_masks)
            if after_outs is not None:
                after_outs = after_outs.masked_select(out_masks)
            ys = ys.masked_select(out_masks)
            duration_masks = make_non_pad_mask(ilens).to(ys.device)
            d_outs = d_outs.masked_select(duration_masks)
            ds = ds.masked_select(duration_masks)
            pitch_masks = make_non_pad_mask(ilens).unsqueeze(-1).to(ys.device)
            p_outs = p_outs.masked_select(pitch_masks)
            e_outs = e_outs.masked_select(pitch_masks)
            ps = ps.masked_select(pitch_masks)
            es = es.masked_select(pitch_masks)

        # calculate loss
        l1_loss = self.l1_criterion(before_outs, ys)
        if after_outs is not None:
            l1_loss += self.l1_criterion(after_outs, ys)
        duration_loss = self.duration_criterion(d_outs, ds)
        pitch_loss = self.mse_criterion(p_outs, ps)
        energy_loss = self.mse_criterion(e_outs, es)

        # make weighted mask and apply it
        if self.use_weighted_masking:
            out_masks = make_non_pad_mask(olens).unsqueeze(-1).to(ys.device)
            out_weights = out_masks.float() / out_masks.sum(dim=1, keepdim=True).float()
            out_weights /= ys.size(0) * ys.size(2)
            duration_masks = make_non_pad_mask(ilens).to(ys.device)
            duration_weights = (
                duration_masks.float() / duration_masks.sum(dim=1, keepdim=True).float()
            )
            duration_weights /= ds.size(0)

            # apply weight
            l1_loss = l1_loss.mul(out_weights).masked_select(out_masks).sum()
            duration_loss = (
                duration_loss.mul(duration_weights).masked_select(duration_masks).sum()
            )
            pitch_masks = duration_masks.unsqueeze(-1)
            pitch_weights = duration_weights.unsqueeze(-1)
            pitch_loss = pitch_loss.mul(pitch_weights).masked_select(pitch_masks).sum()
            energy_loss = (
                energy_loss.mul(pitch_weights).masked_select(pitch_masks).sum()
            )

        return l1_loss, duration_loss, pitch_loss, energy_loss