File size: 4,877 Bytes
ad16788
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
134
135
#!/usr/bin/env python3
# -*- coding: utf-8 -*-

# Copyright 2019 Shigeki Karita
#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)

"""Decoder self-attention layer definition."""

import torch
from torch import nn

from espnet.nets.pytorch_backend.transformer.layer_norm import LayerNorm


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): Self-attention module instance.
            `MultiHeadedAttention` 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.
        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,
        src_attn,
        feed_forward,
        dropout_rate,
        normalize_before=True,
        concat_after=False,
    ):
        """Construct an DecoderLayer object."""
        super(DecoderLayer, self).__init__()
        self.size = size
        self.self_attn = self_attn
        self.src_attn = src_attn
        self.feed_forward = feed_forward
        self.norm1 = LayerNorm(size)
        self.norm2 = LayerNorm(size)
        self.norm3 = LayerNorm(size)
        self.dropout = nn.Dropout(dropout_rate)
        self.normalize_before = normalize_before
        self.concat_after = concat_after
        if self.concat_after:
            self.concat_linear1 = nn.Linear(size + size, size)
            self.concat_linear2 = nn.Linear(size + size, size)

    def forward(self, tgt, tgt_mask, memory, memory_mask, cache=None):
        """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, float32 (#batch, maxlen_in, size).
            memory_mask (torch.Tensor): Encoded memory mask (#batch, maxlen_in).
            cache (List[torch.Tensor]): List of cached tensors.
                Each tensor shape should be (#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,
            ), f"{cache.shape} == {(tgt.shape[0], tgt.shape[1] - 1, self.size)}"
            tgt_q = tgt[:, -1:, :]
            residual = residual[:, -1:, :]
            tgt_q_mask = None
            if tgt_mask is not None:
                tgt_q_mask = tgt_mask[:, -1:, :]

        if self.concat_after:
            tgt_concat = torch.cat(
                (tgt_q, self.self_attn(tgt_q, tgt, tgt, tgt_q_mask)), dim=-1
            )
            x = residual + self.concat_linear1(tgt_concat)
        else:
            x = residual + self.dropout(self.self_attn(tgt_q, tgt, tgt, tgt_q_mask))
        if not self.normalize_before:
            x = self.norm1(x)

        residual = x
        if self.normalize_before:
            x = self.norm2(x)
        if self.concat_after:
            x_concat = torch.cat(
                (x, self.src_attn(x, memory, memory, memory_mask)), dim=-1
            )
            x = residual + self.concat_linear2(x_concat)
        else:
            x = residual + self.dropout(self.src_attn(x, memory, memory, memory_mask))
        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