File size: 4,807 Bytes
be9690e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
# Copyright (c) 2019 Shigeki Karita
#               2020 Mobvoi Inc (Binbin Zhang)
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#   http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Decoder self-attention layer definition."""
from typing import Optional, Tuple

import torch
from torch import nn


class DecoderLayer(nn.Module):
    """Single decoder layer module.

    Args:
        size (int): Input dimension.
        self_attn (torch.nn.Module): Self-attention module instance.
            `MultiHeadedAttention` instance can be used as the argument.
        src_attn (torch.nn.Module): Inter-attention module instance.
            `MultiHeadedAttention` instance can be used as the argument.
            If `None` is passed, Inter-attention is not used, such as
            CIF, GPT, and other decoder only model.
        feed_forward (torch.nn.Module): Feed-forward module instance.
            `PositionwiseFeedForward` instance can be used as the argument.
        dropout_rate (float): Dropout rate.
        normalize_before (bool):
            True: use layer_norm before each sub-block.
            False: to use layer_norm after each sub-block.
    """

    def __init__(
        self,
        size: int,
        self_attn: nn.Module,
        src_attn: Optional[nn.Module],
        feed_forward: nn.Module,
        dropout_rate: float,
        normalize_before: bool = True,
    ):
        """Construct an DecoderLayer object."""
        super().__init__()
        self.size = size
        self.self_attn = self_attn
        self.src_attn = src_attn
        self.feed_forward = feed_forward
        self.norm1 = nn.LayerNorm(size, eps=1e-5)
        self.norm2 = nn.LayerNorm(size, eps=1e-5)
        self.norm3 = nn.LayerNorm(size, eps=1e-5)
        self.dropout = nn.Dropout(dropout_rate)
        self.normalize_before = normalize_before

    def forward(
        self,
        tgt: torch.Tensor,
        tgt_mask: torch.Tensor,
        memory: torch.Tensor,
        memory_mask: torch.Tensor,
        cache: Optional[torch.Tensor] = None
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
        """Compute decoded features.

        Args:
            tgt (torch.Tensor): Input tensor (#batch, maxlen_out, size).
            tgt_mask (torch.Tensor): Mask for input tensor
                (#batch, maxlen_out).
            memory (torch.Tensor): Encoded memory
                (#batch, maxlen_in, size).
            memory_mask (torch.Tensor): Encoded memory mask
                (#batch, maxlen_in).
            cache (torch.Tensor): cached tensors.
                (#batch, maxlen_out - 1, size).

        Returns:
            torch.Tensor: Output tensor (#batch, maxlen_out, size).
            torch.Tensor: Mask for output tensor (#batch, maxlen_out).
            torch.Tensor: Encoded memory (#batch, maxlen_in, size).
            torch.Tensor: Encoded memory mask (#batch, maxlen_in).

        """
        residual = tgt
        if self.normalize_before:
            tgt = self.norm1(tgt)

        if cache is None:
            tgt_q = tgt
            tgt_q_mask = tgt_mask
        else:
            # compute only the last frame query keeping dim: max_time_out -> 1
            assert cache.shape == (
                tgt.shape[0],
                tgt.shape[1] - 1,
                self.size,
            ), "{cache.shape} == {(tgt.shape[0], tgt.shape[1] - 1, self.size)}"
            tgt_q = tgt[:, -1:, :]
            residual = residual[:, -1:, :]
            tgt_q_mask = tgt_mask[:, -1:, :]

        x = residual + self.dropout(
            self.self_attn(tgt_q, tgt, tgt, tgt_q_mask)[0])
        if not self.normalize_before:
            x = self.norm1(x)

        if self.src_attn is not None:
            residual = x
            if self.normalize_before:
                x = self.norm2(x)
            x = residual + self.dropout(
                self.src_attn(x, memory, memory, memory_mask)[0])
            if not self.normalize_before:
                x = self.norm2(x)

        residual = x
        if self.normalize_before:
            x = self.norm3(x)
        x = residual + self.dropout(self.feed_forward(x))
        if not self.normalize_before:
            x = self.norm3(x)

        if cache is not None:
            x = torch.cat([cache, x], dim=1)

        return x, tgt_mask, memory, memory_mask