File size: 6,038 Bytes
01bb3bb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
171
172
173
174
175
176
177
# Scene Text Recognition Model Hub
# Copyright 2022 Darwin Bautista
#
# 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
#
#     https://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.

import math
from typing import Optional

import torch
from torch import Tensor, nn as nn
from torch.nn import functional as F
from torch.nn.modules import transformer

from timm.models.vision_transformer import PatchEmbed, VisionTransformer


class DecoderLayer(nn.Module):
    """A Transformer decoder layer supporting two-stream attention (XLNet)
    This implements a pre-LN decoder, as opposed to the post-LN default in PyTorch."""

    def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1, activation='gelu', layer_norm_eps=1e-5):
        super().__init__()
        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout, batch_first=True)
        self.cross_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout, batch_first=True)
        # Implementation of Feedforward model
        self.linear1 = nn.Linear(d_model, dim_feedforward)
        self.dropout = nn.Dropout(dropout)
        self.linear2 = nn.Linear(dim_feedforward, d_model)

        self.norm1 = nn.LayerNorm(d_model, eps=layer_norm_eps)
        self.norm2 = nn.LayerNorm(d_model, eps=layer_norm_eps)
        self.norm_q = nn.LayerNorm(d_model, eps=layer_norm_eps)
        self.norm_c = nn.LayerNorm(d_model, eps=layer_norm_eps)
        self.dropout1 = nn.Dropout(dropout)
        self.dropout2 = nn.Dropout(dropout)
        self.dropout3 = nn.Dropout(dropout)

        self.activation = transformer._get_activation_fn(activation)

    def __setstate__(self, state):
        if 'activation' not in state:
            state['activation'] = F.gelu
        super().__setstate__(state)

    def forward_stream(
        self,
        tgt: Tensor,
        tgt_norm: Tensor,
        tgt_kv: Tensor,
        memory: Tensor,
        tgt_mask: Optional[Tensor],
        tgt_key_padding_mask: Optional[Tensor],
    ):
        """Forward pass for a single stream (i.e. content or query)
        tgt_norm is just a LayerNorm'd tgt. Added as a separate parameter for efficiency.
        Both tgt_kv and memory are expected to be LayerNorm'd too.
        memory is LayerNorm'd by ViT.
        """
        tgt2, sa_weights = self.self_attn(
            tgt_norm, tgt_kv, tgt_kv, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask
        )
        tgt = tgt + self.dropout1(tgt2)

        tgt2, ca_weights = self.cross_attn(self.norm1(tgt), memory, memory)
        tgt = tgt + self.dropout2(tgt2)

        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(self.norm2(tgt)))))
        tgt = tgt + self.dropout3(tgt2)
        return tgt, sa_weights, ca_weights

    def forward(
        self,
        query,
        content,
        memory,
        query_mask: Optional[Tensor] = None,
        content_mask: Optional[Tensor] = None,
        content_key_padding_mask: Optional[Tensor] = None,
        update_content: bool = True,
    ):
        query_norm = self.norm_q(query)
        content_norm = self.norm_c(content)
        query = self.forward_stream(query, query_norm, content_norm, memory, query_mask, content_key_padding_mask)[0]
        if update_content:
            content = self.forward_stream(
                content, content_norm, content_norm, memory, content_mask, content_key_padding_mask
            )[0]
        return query, content


class Decoder(nn.Module):
    __constants__ = ['norm']

    def __init__(self, decoder_layer, num_layers, norm):
        super().__init__()
        self.layers = transformer._get_clones(decoder_layer, num_layers)
        self.num_layers = num_layers
        self.norm = norm

    def forward(
        self,
        query,
        content,
        memory,
        query_mask: Optional[Tensor] = None,
        content_mask: Optional[Tensor] = None,
        content_key_padding_mask: Optional[Tensor] = None,
    ):
        for i, mod in enumerate(self.layers):
            last = i == len(self.layers) - 1
            query, content = mod(
                query, content, memory, query_mask, content_mask, content_key_padding_mask, update_content=not last
            )
        query = self.norm(query)
        return query


class Encoder(VisionTransformer):

    def __init__(
        self,
        img_size=224,
        patch_size=16,
        in_chans=3,
        embed_dim=768,
        depth=12,
        num_heads=12,
        mlp_ratio=4.0,
        qkv_bias=True,
        drop_rate=0.0,
        attn_drop_rate=0.0,
        drop_path_rate=0.0,
        embed_layer=PatchEmbed,
    ):
        super().__init__(
            img_size,
            patch_size,
            in_chans,
            embed_dim=embed_dim,
            depth=depth,
            num_heads=num_heads,
            mlp_ratio=mlp_ratio,
            qkv_bias=qkv_bias,
            drop_rate=drop_rate,
            attn_drop_rate=attn_drop_rate,
            drop_path_rate=drop_path_rate,
            embed_layer=embed_layer,
            num_classes=0,  # These
            global_pool='',  # disable the
            class_token=False,  # classifier head.
        )

    def forward(self, x):
        # Return all tokens
        return self.forward_features(x)


class TokenEmbedding(nn.Module):

    def __init__(self, charset_size: int, embed_dim: int):
        super().__init__()
        self.embedding = nn.Embedding(charset_size, embed_dim)
        self.embed_dim = embed_dim

    def forward(self, tokens: torch.Tensor):
        return math.sqrt(self.embed_dim) * self.embedding(tokens)