SelectorNet/SelectorNet.py at main · cshan-github/SelectorNet · 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 torch.nn as nn
import numpy as np

class SelectorBlock(nn.Module):
    def __init__(self, input_size):
        super(SelectorBlock, self).__init__()
        self.linear=nn.Linear(input_size,input_size)
    def forward(self, x,x_2):
        a=self.linear(x)
        a=nn.ReLU()(a)
        s=torch.zeros_like(a)
        s[a>0]=nn.Sigmoid()(a[a>0])
        x_1=x*s
        x=x+x_1
        x=x*x_2
        explain=x_2+a
        return x,explain
class ResBlock(nn.Module):
    def __init__(self, input_size):
        super(ResBlock, self).__init__()
        self.linear=nn.Linear(input_size,input_size,dtype=torch.float32)
        self.relu=nn.ReLU()
    def forward(self, x):
        a=x
        x=self.linear(x)
        x=self.relu(x)
        x=a+x
        return x

class FusionAttentionBlock(nn.Module):
    def __init__(self, input_size):
        super(FusionAttentionBlock, self).__init__()
        self.q_linear=nn.Linear(input_size,input_size)
        self.k_linear=nn.Linear(input_size,input_size)
        self.v1_linear=nn.Linear(input_size,input_size)
        self.v2_linear=nn.Linear(input_size,input_size)
        self.bn=nn.BatchNorm1d(input_size)
        self.relu=nn.ReLU()
    def forward(self, x1,x2):
        q=self.q_linear(x1)
        v1=q
        k=self.k_linear(x2)
        v2=k
        attn=(q.transpose(-2, -1) @ k).softmax(dim=-1)
        v1=(v1 @ attn)
        v2=(v2 @ attn)

        v1=self.v1_linear(v1)
        v2=self.v2_linear(v2)
        x1=x1+v1
        x2=x2+v2
        x1=self.bn(x1)
        x2=self.bn(x2)
        return x1,x2


class Step(nn.Module):
    def __init__(self, input_size):
        super(Step, self).__init__()
        self.resblock1=ResBlock(input_size)
        self.resblock2=ResBlock(input_size)
        self.FAB=FusionAttentionBlock(input_size)
        self.selector=SelectorBlock(input_size)
        self.MLP=torch.nn.ModuleList()
        self.MLP.append(ResBlock(input_size))
        self.MLP.append(ResBlock(input_size))
    def forward(self,input,a):
        a=self.resblock2(a)
        a=nn.Sigmoid()(a)
        x,explain=self.selector(input,a)
        x=self.resblock1(x)
        x_d,a=self.FAB(x,a)
        for i in range(len(self.MLP)):
            x_d=self.MLP[i](x_d)
        return x_d,a,explain
class Head(nn.Module):
    def __init__(self, input_size,fusion_dim):
        super(Head, self).__init__()
        self.downdim=nn.Linear(32,fusion_dim)
        self.final_linear=nn.Linear(input_size+fusion_dim,1)
    def forward(self,input,a):
        a=self.downdim(a)
        a=nn.ReLU()(a)
        input=torch.concat((input,a),dim=-1)
        x=self.final_linear(input)
        x=nn.Sigmoid()(x)
        return x

class FusionSelectorNet(nn.Module):
    def __init__(self, input_size,fusion_dim=10,n_weight=2):
        super(FusionSelectorNet, self).__init__()
        self.init_step=Step(input_size)
        self.steps=torch.nn.ModuleList()
        for i in range(n_weight):
            self.steps.append(Step(input_size))
        self.head=Head(input_size,fusion_dim)


    def forward(self,input,a):

        x_a=torch.ones(input.shape).to(input.device)
        x_a=self.init_step(input,x_a)[1]
        res_list=[]
        for i in range(len(self.steps)):
            x_d,x_a,w=self.steps[i](input,x_a)
            res_list.append(x_d)
        res = torch.sum(torch.stack(res_list, dim=0), dim=0)
        res=self.head(res,a)
        return res
    def explain(self, input,a):
        x_a=torch.ones(input.shape).to(input.device)
        x_a=self.init_step(input,x_a)[1]
        res_list=[]
        explain = torch.zeros(input.shape).to(input.device)
        for i in range(len(self.steps)):
            x_d,x_a,w=self.steps[i](input,x_a)
            res_list.append(x_d)
            step_importance = torch.sum(x_d, dim=1)
            explain += torch.mul(w, step_importance.unsqueeze(dim=1))
        explain=nn.ReLU()(explain)
        explain=explain.softmax(dim=-1)
        return explain