Initial commit

2026-02-06 20:31:14 +08:00
parent a0b51be095
commit c589bcb837
145 changed files with 28773 additions and 0 deletions
@@ -0,0 +1 @@
+"""Conformer layers for ROSVOT."""
@@ -0,0 +1,96 @@
+from torch import nn
+from .espnet_positional_embedding import RelPositionalEncoding, ScaledPositionalEncoding, PositionalEncoding
+from .espnet_transformer_attn import RelPositionMultiHeadedAttention, MultiHeadedAttention
+from .layers import Swish, ConvolutionModule, EncoderLayer, MultiLayeredConv1d
+from ..layers import Embedding
+
+
+class ConformerLayers(nn.Module):
+    def __init__(self, hidden_size, num_layers, kernel_size=9, dropout=0.0, num_heads=4,
+                 use_last_norm=True, save_hidden=False):
+        super().__init__()
+        self.use_last_norm = use_last_norm
+        self.layers = nn.ModuleList()
+        positionwise_layer = MultiLayeredConv1d
+        positionwise_layer_args = (hidden_size, hidden_size * 4, 1, dropout)
+        self.pos_embed = RelPositionalEncoding(hidden_size, dropout)
+        self.encoder_layers = nn.ModuleList([EncoderLayer(
+            hidden_size,
+            RelPositionMultiHeadedAttention(num_heads, hidden_size, 0.0),
+            positionwise_layer(*positionwise_layer_args),
+            positionwise_layer(*positionwise_layer_args),
+            ConvolutionModule(hidden_size, kernel_size, Swish()),
+            dropout,
+        ) for _ in range(num_layers)])
+        if self.use_last_norm:
+            self.layer_norm = nn.LayerNorm(hidden_size)
+        else:
+            self.layer_norm = nn.Linear(hidden_size, hidden_size)
+        self.save_hidden = save_hidden
+        if save_hidden:
+            self.hiddens = []
+
+    def forward(self, x, padding_mask=None):
+        """
+
+        :param x: [B, T, H]
+        :param padding_mask: [B, T]
+        :return: [B, T, H]
+        """
+        self.hiddens = []
+        nonpadding_mask = x.abs().sum(-1) > 0
+        x = self.pos_embed(x)
+        for l in self.encoder_layers:
+            x, mask = l(x, nonpadding_mask[:, None, :])
+            if self.save_hidden:
+                self.hiddens.append(x[0])
+        x = x[0]
+        x = self.layer_norm(x) * nonpadding_mask.float()[:, :, None]
+        return x
+
+class FastConformerLayers(ConformerLayers):
+    def __init__(self, hidden_size, num_layers, kernel_size=9, dropout=0.0, num_heads=4,
+                 use_last_norm=True, save_hidden=False):
+        super(ConformerLayers, self).__init__()
+        self.use_last_norm = use_last_norm
+        self.layers = nn.ModuleList()
+        positionwise_layer = MultiLayeredConv1d
+        positionwise_layer_args = (hidden_size, hidden_size * 4, 1, dropout)
+        self.pos_embed = PositionalEncoding(hidden_size, dropout)
+        self.encoder_layers = nn.ModuleList([EncoderLayer(
+            hidden_size,
+            MultiHeadedAttention(num_heads, hidden_size, 0.0, flash=True),
+            positionwise_layer(*positionwise_layer_args),
+            positionwise_layer(*positionwise_layer_args),
+            ConvolutionModule(hidden_size, kernel_size, Swish()),
+            dropout,
+        ) for _ in range(num_layers)])
+        if self.use_last_norm:
+            self.layer_norm = nn.LayerNorm(hidden_size)
+        else:
+            self.layer_norm = nn.Linear(hidden_size, hidden_size)
+        self.save_hidden = save_hidden
+        if save_hidden:
+            self.hiddens = []
+
+class ConformerEncoder(ConformerLayers):
+    def __init__(self, hidden_size, dict_size, num_layers=None):
+        conformer_enc_kernel_size = 9
+        super().__init__(hidden_size, num_layers, conformer_enc_kernel_size)
+        self.embed = Embedding(dict_size, hidden_size, padding_idx=0)
+
+    def forward(self, x):
+        """
+
+        :param src_tokens: [B, T]
+        :return: [B x T x C]
+        """
+        x = self.embed(x)  # [B, T, H]
+        x = super(ConformerEncoder, self).forward(x)
+        return x
+
+
+class ConformerDecoder(ConformerLayers):
+    def __init__(self, hidden_size, num_layers):
+        conformer_dec_kernel_size = 9
+        super().__init__(hidden_size, num_layers, conformer_dec_kernel_size)
@@ -0,0 +1,113 @@
+import math
+import torch
+
+
+class PositionalEncoding(torch.nn.Module):
+    """Positional encoding.
+    Args:
+        d_model (int): Embedding dimension.
+        dropout_rate (float): Dropout rate.
+        max_len (int): Maximum input length.
+        reverse (bool): Whether to reverse the input position.
+    """
+
+    def __init__(self, d_model, dropout_rate, max_len=5000, reverse=False):
+        """Construct an PositionalEncoding object."""
+        super(PositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.reverse = reverse
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x):
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                if self.pe.dtype != x.dtype or self.pe.device != x.device:
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model)
+        if self.reverse:
+            position = torch.arange(
+                x.size(1) - 1, -1, -1.0, dtype=torch.float32
+            ).unsqueeze(1)
+        else:
+            position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor):
+        """Add positional encoding.
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1)]
+        return self.dropout(x)
+
+
+class ScaledPositionalEncoding(PositionalEncoding):
+    """Scaled positional encoding module.
+    See Sec. 3.2  https://arxiv.org/abs/1809.08895
+    Args:
+        d_model (int): Embedding dimension.
+        dropout_rate (float): Dropout rate.
+        max_len (int): Maximum input length.
+    """
+
+    def __init__(self, d_model, dropout_rate, max_len=5000):
+        """Initialize class."""
+        super().__init__(d_model=d_model, dropout_rate=dropout_rate, max_len=max_len)
+        self.alpha = torch.nn.Parameter(torch.tensor(1.0))
+
+    def reset_parameters(self):
+        """Reset parameters."""
+        self.alpha.data = torch.tensor(1.0)
+
+    def forward(self, x):
+        """Add positional encoding.
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+        """
+        self.extend_pe(x)
+        x = x + self.alpha * self.pe[:, : x.size(1)]
+        return self.dropout(x)
+
+
+class RelPositionalEncoding(PositionalEncoding):
+    """Relative positional encoding module.
+    See : Appendix B in https://arxiv.org/abs/1901.02860
+    Args:
+        d_model (int): Embedding dimension.
+        dropout_rate (float): Dropout rate.
+        max_len (int): Maximum input length.
+    """
+
+    def __init__(self, d_model, dropout_rate, max_len=5000):
+        """Initialize class."""
+        super().__init__(d_model, dropout_rate, max_len, reverse=True)
+
+    def forward(self, x):
+        """Compute positional encoding.
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Positional embedding tensor (1, time, `*`).
+        """
+        self.extend_pe(x)
+        x = x * self.xscale
+        pos_emb = self.pe[:, : x.size(1)]
+        return self.dropout(x), self.dropout(pos_emb)
@@ -0,0 +1,198 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+# Copyright 2019 Shigeki Karita
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+"""Multi-Head Attention layer definition."""
+
+from packaging import version
+import math
+
+import numpy
+import torch
+from torch import nn
+
+
+class MultiHeadedAttention(nn.Module):
+    """Multi-Head Attention layer.
+    Args:
+        n_head (int): The number of heads.
+        n_feat (int): The number of features.
+        dropout_rate (float): Dropout rate.
+    """
+
+    def __init__(self, n_head, n_feat, dropout_rate, flash=False):
+        """Construct an MultiHeadedAttention object."""
+        super(MultiHeadedAttention, self).__init__()
+        assert n_feat % n_head == 0
+        # We assume d_v always equals d_k
+        self.d_k = n_feat // n_head
+        self.h = n_head
+        self.linear_q = nn.Linear(n_feat, n_feat)
+        self.linear_k = nn.Linear(n_feat, n_feat)
+        self.linear_v = nn.Linear(n_feat, n_feat)
+        self.linear_out = nn.Linear(n_feat, n_feat)
+        self.attn = None
+        self.dropout = nn.Dropout(p=dropout_rate)
+        self.dropout_rate = dropout_rate
+        self.flash = flash
+
+    def forward_qkv(self, query, key, value):
+        """Transform query, key and value.
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+        Returns:
+            torch.Tensor: Transformed query tensor (#batch, n_head, time1, d_k).
+            torch.Tensor: Transformed key tensor (#batch, n_head, time2, d_k).
+            torch.Tensor: Transformed value tensor (#batch, n_head, time2, d_k).
+        """
+        n_batch = query.size(0)
+        q = self.linear_q(query).view(n_batch, -1, self.h, self.d_k)
+        k = self.linear_k(key).view(n_batch, -1, self.h, self.d_k)
+        v = self.linear_v(value).view(n_batch, -1, self.h, self.d_k)
+        q = q.transpose(1, 2)  # (batch, head, time1, d_k)
+        k = k.transpose(1, 2)  # (batch, head, time2, d_k)
+        v = v.transpose(1, 2)  # (batch, head, time2, d_k)
+
+        return q, k, v
+
+    def forward_attention(self, value, scores, mask):
+        """Compute attention context vector.
+        Args:
+            value (torch.Tensor): Transformed value (#batch, n_head, time2, d_k).
+            scores (torch.Tensor): Attention score (#batch, n_head, time1, time2).
+            mask (torch.Tensor): Mask (#batch, 1, time2) or (#batch, time1, time2).
+        Returns:
+            torch.Tensor: Transformed value (#batch, time1, d_model)
+                weighted by the attention score (#batch, time1, time2).
+        """
+        n_batch = value.size(0)
+        if mask is not None:
+            mask = mask.unsqueeze(1).eq(0)  # (batch, 1, *, time2)
+            min_value = float(
+                numpy.finfo(torch.tensor(0, dtype=scores.dtype).numpy().dtype).min
+            )
+            scores = scores.masked_fill(mask, min_value)
+            self.attn = torch.softmax(scores, dim=-1).masked_fill(
+                mask, 0.0
+            )  # (batch, head, time1, time2)
+        else:
+            self.attn = torch.softmax(scores, dim=-1)  # (batch, head, time1, time2)
+
+        p_attn = self.dropout(self.attn)
+        x = torch.matmul(p_attn, value)  # (batch, head, time1, d_k)
+        x = (
+            x.transpose(1, 2).contiguous().view(n_batch, -1, self.h * self.d_k)
+        )  # (batch, time1, d_model)
+
+        return self.linear_out(x)  # (batch, time1, d_model)
+
+    def forward(self, query, key, value, mask):
+        """Compute scaled dot product attention.
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+            mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
+                (#batch, time1, time2).
+        Returns:
+            torch.Tensor: Output tensor (#batch, time1, d_model).
+        """
+        q, k, v = self.forward_qkv(query, key, value)
+        if version.parse(torch.__version__) >= version.parse("2.0") and self.flash:
+            n_batch = value.size(0)
+            x = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v, attn_mask=mask.unsqueeze(1) if mask is not None else None, dropout_p=self.dropout_rate)
+            x = (
+                x.transpose(1, 2).contiguous().view(n_batch, -1, self.h * self.d_k)
+            )  # (batch, time1, d_model)
+            return self.linear_out(x)
+        else:
+            scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)
+            return self.forward_attention(v, scores, mask)
+
+
+class RelPositionMultiHeadedAttention(MultiHeadedAttention):
+    """Multi-Head Attention layer with relative position encoding.
+    Paper: https://arxiv.org/abs/1901.02860
+    Args:
+        n_head (int): The number of heads.
+        n_feat (int): The number of features.
+        dropout_rate (float): Dropout rate.
+    """
+
+    def __init__(self, n_head, n_feat, dropout_rate):
+        """Construct an RelPositionMultiHeadedAttention object."""
+        super().__init__(n_head, n_feat, dropout_rate)
+        # linear transformation for positional ecoding
+        self.linear_pos = nn.Linear(n_feat, n_feat, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in https://arxiv.org/abs/1901.02860 Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(self.h, self.d_k))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(self.h, self.d_k))
+        torch.nn.init.xavier_uniform_(self.pos_bias_u)
+        torch.nn.init.xavier_uniform_(self.pos_bias_v)
+
+    def rel_shift(self, x, zero_triu=False):
+        """Compute relative positinal encoding.
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, size).
+            zero_triu (bool): If true, return the lower triangular part of the matrix.
+        Returns:
+            torch.Tensor: Output tensor.
+        """
+        zero_pad = torch.zeros((*x.size()[:3], 1), device=x.device, dtype=x.dtype)
+        x_padded = torch.cat([zero_pad, x], dim=-1)
+
+        x_padded = x_padded.view(*x.size()[:2], x.size(3) + 1, x.size(2))
+        x = x_padded[:, :, 1:].view_as(x)
+
+        if zero_triu:
+            ones = torch.ones((x.size(2), x.size(3)))
+            x = x * torch.tril(ones, x.size(3) - x.size(2))[None, None, :, :]
+
+        return x
+
+    def forward(self, query, key, value, pos_emb, mask):
+        """Compute 'Scaled Dot Product Attention' with rel. positional encoding.
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+            pos_emb (torch.Tensor): Positional embedding tensor (#batch, time2, size).
+            mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
+                (#batch, time1, time2).
+        Returns:
+            torch.Tensor: Output tensor (#batch, time1, d_model).
+        """
+        q, k, v = self.forward_qkv(query, key, value)
+        q = q.transpose(1, 2)  # (batch, time1, head, d_k)
+
+        n_batch_pos = pos_emb.size(0)
+        p = self.linear_pos(pos_emb).view(n_batch_pos, -1, self.h, self.d_k)
+        p = p.transpose(1, 2)  # (batch, head, time1, d_k)
+
+        # (batch, head, time1, d_k)
+        q_with_bias_u = (q + self.pos_bias_u).transpose(1, 2)
+        # (batch, head, time1, d_k)
+        q_with_bias_v = (q + self.pos_bias_v).transpose(1, 2)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in https://arxiv.org/abs/1901.02860 Section 3.3
+        # (batch, head, time1, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k.transpose(-2, -1))
+
+        # compute matrix b and matrix d
+        # (batch, head, time1, time2)
+        matrix_bd = torch.matmul(q_with_bias_v, p.transpose(-2, -1))
+        matrix_bd = self.rel_shift(matrix_bd)
+
+        scores = (matrix_ac + matrix_bd) / math.sqrt(
+            self.d_k
+        )  # (batch, head, time1, time2)
+
+        return self.forward_attention(v, scores, mask)
@@ -0,0 +1,260 @@
+from torch import nn
+import torch
+
+from ..layers import LayerNorm
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model.
+    Args:
+        channels (int): The number of channels of conv layers.
+        kernel_size (int): Kernerl size of conv layers.
+    """
+
+    def __init__(self, channels, kernel_size, activation=nn.ReLU(), bias=True):
+        """Construct an ConvolutionModule object."""
+        super(ConvolutionModule, self).__init__()
+        # kernerl_size should be a odd number for 'SAME' padding
+        assert (kernel_size - 1) % 2 == 0
+
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.depthwise_conv = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+            groups=channels,
+            bias=bias,
+        )
+        self.norm = nn.BatchNorm1d(channels)
+        self.pointwise_conv2 = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.activation = activation
+
+    def forward(self, x):
+        """Compute convolution module.
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, channels).
+        Returns:
+            torch.Tensor: Output tensor (#batch, time, channels).
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.transpose(1, 2)
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channel, dim)
+        x = nn.functional.glu(x, dim=1)  # (batch, channel, dim)
+
+        # 1D Depthwise Conv
+        x = self.depthwise_conv(x)
+        x = self.activation(self.norm(x))
+
+        x = self.pointwise_conv2(x)
+
+        return x.transpose(1, 2)
+
+
+class MultiLayeredConv1d(torch.nn.Module):
+    """Multi-layered conv1d for Transformer block.
+    This is a module of multi-leyered conv1d designed
+    to replace positionwise feed-forward network
+    in Transforner block, which is introduced in
+    `FastSpeech: Fast, Robust and Controllable Text to Speech`_.
+    .. _`FastSpeech: Fast, Robust and Controllable Text to Speech`:
+        https://arxiv.org/pdf/1905.09263.pdf
+    """
+
+    def __init__(self, in_chans, hidden_chans, kernel_size, dropout_rate):
+        """Initialize MultiLayeredConv1d module.
+        Args:
+            in_chans (int): Number of input channels.
+            hidden_chans (int): Number of hidden channels.
+            kernel_size (int): Kernel size of conv1d.
+            dropout_rate (float): Dropout rate.
+        """
+        super(MultiLayeredConv1d, self).__init__()
+        self.w_1 = torch.nn.Conv1d(
+            in_chans,
+            hidden_chans,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+        )
+        self.w_2 = torch.nn.Conv1d(
+            hidden_chans,
+            in_chans,
+            kernel_size,
+            stride=1,
+            padding=(kernel_size - 1) // 2,
+        )
+        self.dropout = torch.nn.Dropout(dropout_rate)
+
+    def forward(self, x):
+        """Calculate forward propagation.
+        Args:
+            x (torch.Tensor): Batch of input tensors (B, T, in_chans).
+        Returns:
+            torch.Tensor: Batch of output tensors (B, T, hidden_chans).
+        """
+        x = torch.relu(self.w_1(x.transpose(-1, 1))).transpose(-1, 1)
+        return self.w_2(self.dropout(x).transpose(-1, 1)).transpose(-1, 1)
+
+
+class Swish(torch.nn.Module):
+    """Construct an Swish object."""
+
+    def forward(self, x):
+        """Return Swich activation function."""
+        return x * torch.sigmoid(x)
+
+
+class EncoderLayer(nn.Module):
+    """Encoder layer module.
+    Args:
+        size (int): Input dimension.
+        self_attn (torch.nn.Module): Self-attention module instance.
+            `MultiHeadedAttention` or `RelPositionMultiHeadedAttention` instance
+            can be used as the argument.
+        feed_forward (torch.nn.Module): Feed-forward module instance.
+            `PositionwiseFeedForward`, `MultiLayeredConv1d`, or `Conv1dLinear` instance
+            can be used as the argument.
+        feed_forward_macaron (torch.nn.Module): Additional feed-forward module instance.
+            `PositionwiseFeedForward`, `MultiLayeredConv1d`, or `Conv1dLinear` instance
+            can be used as the argument.
+        conv_module (torch.nn.Module): Convolution module instance.
+            `ConvlutionModule` instance can be used as the argument.
+        dropout_rate (float): Dropout rate.
+        normalize_before (bool): Whether to use layer_norm before the first block.
+        concat_after (bool): Whether to concat attention layer's input and output.
+            if True, additional linear will be applied.
+            i.e. x -> x + linear(concat(x, att(x)))
+            if False, no additional linear will be applied. i.e. x -> x + att(x)
+    """
+
+    def __init__(
+            self,
+            size,
+            self_attn,
+            feed_forward,
+            feed_forward_macaron,
+            conv_module,
+            dropout_rate,
+            normalize_before=True,
+            concat_after=False,
+    ):
+        """Construct an EncoderLayer object."""
+        super(EncoderLayer, self).__init__()
+        self.self_attn = self_attn
+        self.feed_forward = feed_forward
+        self.feed_forward_macaron = feed_forward_macaron
+        self.conv_module = conv_module
+        self.norm_ff = LayerNorm(size)  # for the FNN module
+        self.norm_mha = LayerNorm(size)  # for the MHA module
+        if feed_forward_macaron is not None:
+            self.norm_ff_macaron = LayerNorm(size)
+            self.ff_scale = 0.5
+        else:
+            self.ff_scale = 1.0
+        if self.conv_module is not None:
+            self.norm_conv = LayerNorm(size)  # for the CNN module
+            self.norm_final = LayerNorm(size)  # for the final output of the block
+        self.dropout = nn.Dropout(dropout_rate)
+        self.size = size
+        self.normalize_before = normalize_before
+        self.concat_after = concat_after
+        if self.concat_after:
+            self.concat_linear = nn.Linear(size + size, size)
+
+    def forward(self, x_input, mask, cache=None):
+        """Compute encoded features.
+        Args:
+            x_input (Union[Tuple, torch.Tensor]): Input tensor w/ or w/o pos emb.
+                - w/ pos emb: Tuple of tensors [(#batch, time, size), (1, time, size)].
+                - w/o pos emb: Tensor (#batch, time, size).
+            mask (torch.Tensor): Mask tensor for the input (#batch, time).
+            cache (torch.Tensor): Cache tensor of the input (#batch, time - 1, size).
+        Returns:
+            torch.Tensor: Output tensor (#batch, time, size).
+            torch.Tensor: Mask tensor (#batch, time).
+        """
+        if isinstance(x_input, tuple):
+            x, pos_emb = x_input[0], x_input[1]
+        else:
+            x, pos_emb = x_input, None
+
+        # whether to use macaron style
+        if self.feed_forward_macaron is not None:
+            residual = x
+            if self.normalize_before:
+                x = self.norm_ff_macaron(x)
+            x = residual + self.ff_scale * self.dropout(self.feed_forward_macaron(x))
+            if not self.normalize_before:
+                x = self.norm_ff_macaron(x)
+
+        # multi-headed self-attention module
+        residual = x
+        if self.normalize_before:
+            x = self.norm_mha(x)
+
+        if cache is None:
+            x_q = x
+        else:
+            assert cache.shape == (x.shape[0], x.shape[1] - 1, self.size)
+            x_q = x[:, -1:, :]
+            residual = residual[:, -1:, :]
+            mask = None if mask is None else mask[:, -1:, :]
+
+        if pos_emb is not None:
+            x_att = self.self_attn(x_q, x, x, pos_emb, mask)
+        else:
+            x_att = self.self_attn(x_q, x, x, mask)
+
+        if self.concat_after:
+            x_concat = torch.cat((x, x_att), dim=-1)
+            x = residual + self.concat_linear(x_concat)
+        else:
+            x = residual + self.dropout(x_att)
+        if not self.normalize_before:
+            x = self.norm_mha(x)
+
+        # convolution module
+        if self.conv_module is not None:
+            residual = x
+            if self.normalize_before:
+                x = self.norm_conv(x)
+            x = residual + self.dropout(self.conv_module(x))
+            if not self.normalize_before:
+                x = self.norm_conv(x)
+
+        # feed forward module
+        residual = x
+        if self.normalize_before:
+            x = self.norm_ff(x)
+        x = residual + self.ff_scale * self.dropout(self.feed_forward(x))
+        if not self.normalize_before:
+            x = self.norm_ff(x)
+
+        if self.conv_module is not None:
+            x = self.norm_final(x)
+
+        if cache is not None:
+            x = torch.cat([cache, x], dim=1)
+
+        if pos_emb is not None:
+            return (x, pos_emb), mask
+
+        return x, mask