File size: 7,096 Bytes
6775edf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
from torch.nn import BCEWithLogitsLoss,CrossEntropyLoss, Dropout, Softmax, Linear, Conv2d, LayerNorm
import models.configs as configs
import math

class Attention(nn.Module):
    def __init__(self, config, vis, mm=True):
        super(Attention, self).__init__()
        self.vis = vis
        self.num_attention_heads = config.transformer["num_heads"]
        self.attention_head_size = int(config.hidden_size / self.num_attention_heads)
        self.all_head_size = self.num_attention_heads * self.attention_head_size

        self.query = Linear(config.hidden_size, self.all_head_size)
        self.key = Linear(config.hidden_size, self.all_head_size)
        self.value = Linear(config.hidden_size, self.all_head_size)

        if mm:
            self.query_text = Linear(config.hidden_size, self.all_head_size)
            self.key_text = Linear(config.hidden_size, self.all_head_size)
            self.value_text = Linear(config.hidden_size, self.all_head_size)
            self.out_text = Linear(config.hidden_size, config.hidden_size)
            self.attn_dropout_text = Dropout(config.transformer["attention_dropout_rate"])
            self.attn_dropout_it = Dropout(config.transformer["attention_dropout_rate"])
            self.attn_dropout_ti = Dropout(config.transformer["attention_dropout_rate"])
            self.proj_dropout_text = Dropout(config.transformer["attention_dropout_rate"])

        self.out = Linear(config.hidden_size, config.hidden_size)
        self.attn_dropout = Dropout(config.transformer["attention_dropout_rate"])
        self.proj_dropout = Dropout(config.transformer["attention_dropout_rate"])

        self.softmax = Softmax(dim=-1)

    def transpose_for_scores(self, x):
        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size) 
        x = x.view(*new_x_shape)
        return x.permute(0, 2, 1, 3)

    def forward(self, hidden_states, text=None):
        mixed_query_layer = self.query(hidden_states)
        mixed_key_layer = self.key(hidden_states)
        mixed_value_layer = self.value(hidden_states)
        if text is not None:
            text_q = self.query_text(text)      
            text_k = self.key_text(text)      
            text_v = self.value_text(text)      
        query_layer = self.transpose_for_scores(mixed_query_layer)
        key_layer = self.transpose_for_scores(mixed_key_layer)
        value_layer = self.transpose_for_scores(mixed_value_layer)
        if text is not None:
            query_layer_img = query_layer
            key_layer_img = key_layer
            value_layer_img = value_layer
            query_layer_text = self.transpose_for_scores(text_q)
            key_layer_text = self.transpose_for_scores(text_k)
            value_layer_text = self.transpose_for_scores(text_v)

        if text is None:
            attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
            attention_scores = attention_scores / math.sqrt(self.attention_head_size)
            attention_probs = self.softmax(attention_scores)
            weights = attention_probs if self.vis else None
            attention_probs = self.attn_dropout(attention_probs)

            context_layer = torch.matmul(attention_probs, value_layer)
            context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
            new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
            context_layer = context_layer.view(*new_context_layer_shape)
            attention_output = self.out(context_layer)
            attention_output = self.proj_dropout(attention_output)

            return attention_output,None, weights

        else:
            attention_scores_img = torch.matmul(query_layer, key_layer.transpose(-1, -2))
            attention_scores_text = torch.matmul(query_layer_text, key_layer_text.transpose(-1, -2))
            attention_scores_it = torch.matmul(query_layer_img, key_layer_text.transpose(-1, -2))
            attention_scores_ti = torch.matmul(query_layer_text, key_layer_img.transpose(-1, -2))
            attention_scores_img = attention_scores_img / math.sqrt(self.attention_head_size)
            
            attention_probs_img = self.softmax(attention_scores_img)
            weights_img = attention_probs_img if self.vis else None

            attention_probs_img = self.attn_dropout(attention_probs_img)

            attention_scores_text = attention_scores_text / math.sqrt(self.attention_head_size)
            attention_probs_text = self.softmax(attention_scores_text)

            text_per_weights = attention_probs_text.mean(dim=-1)  

            text_per_weights = self.softmax(text_per_weights)

            weights_text = attention_probs_text if self.vis else None

            attention_probs_text = self.attn_dropout_text(attention_probs_text)

            attention_scores_it = attention_scores_it / math.sqrt(self.attention_head_size)
            attention_probs_it = self.softmax(attention_scores_it)

            attention_probs_it = self.attn_dropout_it(attention_probs_it)

            attention_scores_ti = attention_scores_ti / math.sqrt(self.attention_head_size)
            attention_probs_ti = self.softmax(attention_scores_ti)
            attention_probs_ti = self.attn_dropout_ti(attention_probs_ti)

            context_layer_img = torch.matmul(attention_probs_img, value_layer_img)
            context_layer_img = context_layer_img.permute(0, 2, 1, 3).contiguous()
            context_layer_text = torch.matmul(attention_probs_text, value_layer_text)
            context_layer_text = context_layer_text.permute(0, 2, 1, 3).contiguous()
            context_layer_it = torch.matmul(attention_probs_it, value_layer_text)
            context_layer_it = context_layer_it.permute(0, 2, 1, 3).contiguous()
            context_layer_ti = torch.matmul(attention_probs_ti, value_layer_img)
            context_layer_ti = context_layer_ti.permute(0, 2, 1, 3).contiguous()
            new_context_layer_shape = context_layer_img.size()[:-2] + (self.all_head_size,)
            context_layer_img = context_layer_img.view(*new_context_layer_shape)
            new_context_layer_shape = context_layer_text.size()[:-2] + (self.all_head_size,)
            context_layer_text = context_layer_text.view(*new_context_layer_shape)
            new_context_layer_shape = context_layer_it.size()[:-2] + (self.all_head_size,)
            context_layer_it = context_layer_it.view(*new_context_layer_shape)
            new_context_layer_shape = context_layer_ti.size()[:-2] + (self.all_head_size,)
            context_layer_ti = context_layer_ti.view(*new_context_layer_shape)
            attention_output_img = self.out((context_layer_img + context_layer_it)/2)
            attention_output_text = self.out((context_layer_text + context_layer_ti)/2)
            attention_output_img = self.proj_dropout(attention_output_img)
            attention_output_text = self.proj_dropout_text(attention_output_text)
           
            return attention_output_img, attention_output_text