File size: 6,178 Bytes
eb6d243
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
# Copyright (c) DP Technology.
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

from typing import Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
from . import TransformerDecoderLayer, LayerNorm
from .transformer_encoder import relative_position_bucket


def fill_with_neg_inf(t):
    return t.fill_(float("-inf"))


def bulid_future_mask(seq_len):
    return torch.triu(
        fill_with_neg_inf(torch.zeros([seq_len, seq_len])), 1
    )


class TransformerDecoder(nn.Module):
    def __init__(
        self,
        decoder_layers: int = 6,
        embed_dim: int = 768,
        ffn_embed_dim: int = 3072,
        attention_heads: int = 8,
        emb_dropout: float = 0.1,
        dropout: float = 0.1,
        attention_dropout: float = 0.1,
        activation_dropout: float = 0.0,
        max_seq_len: int = 256,
        activation_fn: str = "gelu",
        rel_pos: bool = True,
        rel_pos_bins: int = 32,
        max_rel_pos: int = 128,
        post_ln: bool = False,
        auto_regressive: bool = True,
    ) -> None:

        super().__init__()
        self.emb_dropout = emb_dropout
        self.max_seq_len = max_seq_len
        self.embed_dim = embed_dim
        self.attention_heads = attention_heads
        self.emb_layer_norm = LayerNorm(self.embed_dim)
        self.auto_regressive = auto_regressive
        if self.auto_regressive:
            self._future_mask = bulid_future_mask(self.max_seq_len)
        else:
            self._future_mask = None
        if not post_ln:
            self.final_layer_norm = LayerNorm(self.embed_dim)
        else:
            self.final_layer_norm = None

        self.layers = nn.ModuleList(
            [
                TransformerDecoderLayer(
                    embed_dim=self.embed_dim,
                    ffn_embed_dim=ffn_embed_dim,
                    attention_heads=attention_heads,
                    dropout=dropout,
                    attention_dropout=attention_dropout,
                    activation_dropout=activation_dropout,
                    activation_fn=activation_fn,
                    post_ln=post_ln,

                )
                for _ in range(decoder_layers)
            ]
        )

        self.rel_pos = rel_pos
        if self.rel_pos:
            assert rel_pos_bins % 2 == 0
            self.rel_pos_bins = rel_pos_bins
            self.max_rel_pos = max_rel_pos
            self.relative_attention_bias = nn.Embedding(
                self.rel_pos_bins, self.attention_heads)
            seq_len = self.max_seq_len
            context_position = torch.arange(seq_len, dtype=torch.long)[:, None]
            memory_position = torch.arange(seq_len, dtype=torch.long)[None, :]
            relative_position = memory_position - context_position
            self.rp_bucket = relative_position_bucket(
                relative_position,
                num_buckets=self.rel_pos_bins,
                max_distance=self.max_rel_pos
            )
            self.rp_bucket -= self.rp_bucket.min()

    def get_rel_pos_bias(self, x):
        # Assume the input is ordered. If your input token is permuted, you may need to update this accordingly
        if self.rp_bucket.device != x.device:
            self.rp_bucket = self.rp_bucket.to(x.device)
        seq_len = x.size(1)
        rp_bucket = self.rp_bucket[:seq_len, :seq_len]
        values = F.embedding(rp_bucket, self.relative_attention_bias.weight)
        values = values.permute([2, 0, 1])
        return values.contiguous()

    def get_future_mask(self, x, attn_mask):
        if not self.auto_regressive:
            return attn_mask
        if self._future_mask.device != x.device:
            self._future_mask = self._future_mask.to(x.device)
        if self._future_mask.dtype != x.dtype:
            self._future_mask = self._future_mask.type_as(x)
        if attn_mask is None:
            ret = self._future_mask[:x.size(1), :x.size(1)]
            ret = ret.contiguous().unsqueeze(0).repeat(
                x.size(0)*self.attention_heads, 1, 1)
            return ret
        else:
            assert list(attn_mask.size()) == [x.size(
                0) * self.attention_heads, x.size(1), x.size(1)]
            return attn_mask + self._future_mask[:x.size(1), :x.size(1)]

    def forward(
        self,
        emb,
        encoder_out: Optional[torch.Tensor] = None,
        padding_mask: Optional[torch.Tensor] = None,
        encoder_padding_mask: Optional[torch.Tensor] = None,
        attn_mask: Optional[torch.Tensor] = None,
        encoder_attn_mask: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:

        seq_len = emb.size(1)
        x = self.emb_layer_norm(emb)
        x = F.dropout(x, p=self.emb_dropout, training=self.training)

        # account for padding while computing the representation
        if padding_mask is not None:
            x = x * (1 - padding_mask.unsqueeze(-1).type_as(x))

        rel_pos_bias = self.get_rel_pos_bias(x).repeat(
            x.size(0), 1, 1) if self.rel_pos else None

        if attn_mask is None:
            attn_mask = rel_pos_bias
        elif rel_pos_bias is not None:
            attn_mask += rel_pos_bias

        if self.auto_regressive:
            attn_mask = self.get_future_mask(x, attn_mask)

        if attn_mask is not None and padding_mask is not None:
            # merge key_padding_mask and attn_mask
            attn_mask = attn_mask.view(x.size(0), -1, seq_len, seq_len)
            attn_mask.masked_fill_(
                padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool),
                float("-inf")
            )
            attn_mask = attn_mask.view(-1, seq_len, seq_len)
            padding_mask = None

        for layer in self.layers:
            x = layer(x, encoder_out=encoder_out, padding_mask=padding_mask, attn_bias=attn_mask,
                      encoder_padding_mask=encoder_padding_mask, encoder_attn_bias=encoder_attn_mask)

        if self.final_layer_norm is not None:
            x = self.final_layer_norm(x)

        return x