MultiDTA/transformer.py at master · dengjiejin/MultiDTA · GitHub

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
import torch
import math
import torch.nn as nn


class NaiveTransformerLayer(nn.Module):
    def __init__(self):
        super(NaiveTransformerLayer, self).__init__()
        self.dim = 512
        self.att_drop_rate = 0.1
        self.state_drop_rate = 0.5
        self.Wq = nn.Linear(self.dim, self.dim, bias=False)
        self.Wk = nn.Linear(self.dim, self.dim, bias=False)
        self.Wv = nn.Linear(self.dim, self.dim, bias=False)
        self.lm = nn.LayerNorm(self.dim)
        self.ffn1 = nn.Linear(self.dim, self.dim*4)
        self.ffn2 = nn.Linear(self.dim*4, self.dim)
        self.act = nn.GELU()
        self.lm_ffn = nn.LayerNorm(self.dim)
        self.att_drop = nn.Dropout(self.att_drop_rate)
        self.state_drop = nn.Dropout(self.state_drop_rate)

    def SelfAttention(self, x):
        """
        :param x: b*n*d
        :return: b*n*d
        """
        Q = self.Wq(x)
        K = self.Wk(x)
        V = self.Wv(x)
        attention_score = torch.bmm(Q, K.transpose(1, 2))/math.sqrt(self.dim)
        attention_score = nn.Softmax(dim=2)(attention_score)
        attention_score = self.att_drop(attention_score)
        output = torch.bmm(attention_score, V)
        output = self.state_drop(output)
        output = self.lm(x + output)
        return output

    def FFN(self, x):
        hidden = self.act(self.ffn1(x))
        output = self.ffn2(hidden)
        output = self.state_drop(output)
        output = self.lm_ffn(x+output)
        return output

    def forward(self, x):
        """
        :param x: b*n*d
        :return: b*n*d
        """
        x = self.SelfAttention(x)
        x = self.FNN(x)
        return x


class MultiTransformerLayer(nn.Module):
    def __init__(self):
        super(MultiTransformerLayer, self).__init__()
        self.dim = 512
        self.att_drop_rate = 0.1
        self.state_drop_rate = 0.5
        self.num_heads = 12
        self.size_per_head = self.dim // self.num_head   #64
        self.Wq = nn.Linear(self.dim, self.num_heads * self.size_per_head, bias=False)
        self.Wk = nn.Linear(self.dim, self.num_heads * self.size_per_head, bias=False)
        self.Wv = nn.Linear(self.dim, self.num_heads * self.size_per_head, bias=False)
        self.W = nn.Linear(self.num_heads * self.size_per_head, self.dim)
        self.lm = nn.LayerNorm(self.dim)
        self.ffn1 = nn.Linear(self.dim, self.dim*4)
        self.ffn2 = nn.Linear(self.dim*4, self.dim)
        self.act = nn.GELU()
        self.lm_ffn = nn.LayerNorm(self.dim)
        self.att_drop = nn.Dropout(self.att_drop_rate)
        self.state_drop = nn.Dropout(self.state_drop_rate)

    def calc_mask_score(self, attention_mask):
        """
        :param attention_mask: b x n
        :return: b x h x n x n
        """
        mask_score = torch.zeros(attention_mask.size(0), self.num_heads, attention_mask.size(1), attention_mask.size(1))
        mask_score = mask_score + attention_mask[:, None, None, :]
        mask_score = (1.0 - mask_score) * -10000.
        return mask_score

    def SelfAttention(self, x, attention_mask):
        """
        :param x: b*n*d
        Q,K,V  b x n x (h x s)-->b x n x h x s--> b x h x n x s
        attention_mask # b x n
            1 normal token
            0 masked token
        :return: b*n*d
        """

        new_size = x.size()[:-1] + (self.num_heads, self.size_per_head)  # b, n , h, s
        Q = self.Wq(x).view(*new_size).permute(0, 2, 1, 3)
        K = self.Wk(x).view(*new_size).permute(0, 2, 1, 3)
        V = self.Wv(x).view(*new_size).permute(0, 2, 1, 3)
        attention_score = torch.matmul(Q, K.transpose(2, 3))/math.sqrt(self.dim)
        # attention mask here
        attention_score = attention_score + self.calc_mask_score(attention_score)
        attention_score = nn.Softmax(dim=3)(attention_score)
        attention_score = self.att_drop(attention_score)
        output = torch.mm(attention_score, V)
        output = self.W(output.permute(0, 2, 1, 3))  # b x n x d
        output = self.state_drop(output)
        output = self.lm(x + output)
        return output

    def FFN(self, x):
        hidden = self.act(self.ffn1(x))
        output = self.ffn2(hidden)
        output = self.state_drop(output)
        output = self.lm_ffn(x+output)
        return output

    def forward(self, x):
        """
        :param x: b*n*d
        :return: b*n*d
        """
        x = self.SelfAttention(x)
        x = self.FNN(x)
        return x