-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmethods.py
executable file
·252 lines (224 loc) · 11.1 KB
/
methods.py
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
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
import numpy as np
import keras.models
from keras.models import Model
import matplotlib.pyplot as plt
import keras.backend as K
class Gradient:
def __init__(self, model):
mask0=K.sqrt(K.sum(K.square(K.relu(K.squeeze(K.gradients(model.layers[-2].output[0,0],model.layers[1].input),0))),axis=2))
mask1=K.sqrt(K.sum(K.square(K.relu(K.squeeze(K.gradients(model.layers[-2].output[0,1],model.layers[1].input),0))),axis=2))
mask0 = K.squeeze(mask0, 0)
mask1 = K.squeeze(mask1, 0)
getMasks=K.function([model.layers[0].input,model.layers[1].input],[mask0,mask1])
self.getMasks = getMasks
class EB:
def __init__(self, model):
pLamda4=self.ebDense(K.squeeze(model.layers[-3].output,0),model.layers[-2].weights[0],K.variable(np.array([1,0])))
pdense3=self.ebGAP(model.layers[-5].output,pLamda4)
pLambda3=self.ebMoleculeDense(model.layers[-6].output,model.layers[-5].weights[0],pdense3)
pdense2=self.ebMoleculeAdj(model.layers[-7].output,K.squeeze(model.layers[0].input,0),pLambda3)
pLambda2=self.ebMoleculeDense(model.layers[-8].output,model.layers[-7].weights[0],pdense2)
pdense1=self.ebMoleculeAdj(model.layers[-9].output,K.squeeze(model.layers[0].input,0),pLambda2)
pLambda1=self.ebMoleculeDense(model.layers[-10].output,model.layers[-9].weights[0],pdense1)
pin=self.ebMoleculeAdj(model.layers[-11].output,K.squeeze(model.layers[0].input,0),pLambda1)
mask0=K.squeeze(K.sum(pin,axis=2),0)
pLamda4=self.ebDense(K.squeeze(model.layers[-3].output,0),model.layers[-2].weights[0],K.variable(np.array([0,1])))
pdense3=self.ebGAP(model.layers[-5].output,pLamda4)
pLambda3=self.ebMoleculeDense(model.layers[-6].output,model.layers[-5].weights[0],pdense3)
pdense2=self.ebMoleculeAdj(model.layers[-7].output,K.squeeze(model.layers[0].input,0),pLambda3)
pLambda2=self.ebMoleculeDense(model.layers[-8].output,model.layers[-7].weights[0],pdense2)
pdense1=self.ebMoleculeAdj(model.layers[-9].output,K.squeeze(model.layers[0].input,0),pLambda2)
pLambda1=self.ebMoleculeDense(model.layers[-10].output,model.layers[-9].weights[0],pdense1)
pin=self.ebMoleculeAdj(model.layers[-11].output,K.squeeze(model.layers[0].input,0),pLambda1)
mask1=K.squeeze(K.sum(pin,axis=2),0)
getMasks=K.function([model.layers[0].input,model.layers[1].input],[mask0,mask1])
self.getMasks = getMasks
def ebDense(self, activations,W,bottomP):
'''
This function calculates eb for a dense layer
Input:
activations: d-dimensional vector
W: Weights dxk-dimensional matrix
bottomP: k-dimensional probability vector
Output:
p: the probability of activation d-dimensional vector
'''
Wrelu=K.relu(W)
pcond=K.tf.matmul(K.tf.diag(activations),Wrelu)
pcond=pcond/K.sum(pcond,axis=0)
return K.transpose(K.tf.matmul(pcond,K.expand_dims(bottomP,1)))
def ebMoleculeDense(self, activations,W,bottomP):
'''
This function calculates eb for a dense layer
Input:
activations: 1x?xK
W: Weights dxk-dimensional matrix KxL
bottomP: probability matrix 1x?xL
Output:
p: probability matrix 1x?xK
'''
k,l=W.shape.as_list()
Wrelu=K.relu(W)
pcond=K.tile(K.expand_dims(activations,3),(1,1,1,l))*Wrelu
p=K.mean(K.tile(K.expand_dims(bottomP,2),(1,1,k,1))*pcond,3)
return p
def ebGAP(self, activations,bottomP):
'''
This function calculates eb for GAP layer
Input:
activations: 1x?xK
bottomP: probability matrix 1xK
Output:
p: probability matrix 1x?xK
'''
epsilon=1e-5
pcond=activations/(epsilon+K.sum(activations,axis=1))
p=pcond*K.squeeze(bottomP,0)
p=p/(K.sum(p,axis=1)+epsilon)
return p
def ebMoleculeAdj(self,activations,A,bottomP):
'''
This function calculates eb for a Adj conv layer
Input:
activations: 1x?xK
A: Adjacency ?x?
bottomP: probability matrix 1x?xK
Output:
p: probability matrix 1x?xK
'''
pcond=K.expand_dims(K.tf.matmul(A,K.squeeze(activations,0)),0)
p=pcond*bottomP
return p
class cEB:
def __init__(self, model):
pLamda04=self.ebDense(K.squeeze(model.layers[-3].output,0),model.layers[-2].weights[0],K.variable(np.array([1,0])))
pdense03=self.ebGAP(model.layers[-5].output,pLamda04)
pLambda3=self.ebMoleculeDense(model.layers[-6].output,model.layers[-5].weights[0],pdense03)
pdense2=self.ebMoleculeAdj(model.layers[-7].output,K.squeeze(model.layers[0].input,0),pLambda3)
pLambda2=self.ebMoleculeDense(model.layers[-8].output,model.layers[-7].weights[0],pdense2)
pdense1=self.ebMoleculeAdj(model.layers[-9].output,K.squeeze(model.layers[0].input,0),pLambda2)
pLambda1=self.ebMoleculeDense(model.layers[-10].output,model.layers[-9].weights[0],pdense1)
pin=self.ebMoleculeAdj(model.layers[-11].output,K.squeeze(model.layers[0].input,0),pLambda1)
mask0=K.squeeze(K.sum(pin,axis=2),0)
pLamda14=self.ebDense(K.squeeze(model.layers[-3].output,0),model.layers[-2].weights[0],K.variable(np.array([0,1])))
pdense13=self.ebGAP(model.layers[-5].output,pLamda14)
pLambda3=self.ebMoleculeDense(model.layers[-6].output,model.layers[-5].weights[0],pdense13)
pdense2=self.ebMoleculeAdj(model.layers[-7].output,K.squeeze(model.layers[0].input,0),pLambda3)
pLambda2=self.ebMoleculeDense(model.layers[-8].output,model.layers[-7].weights[0],pdense2)
pdense1=self.ebMoleculeAdj(model.layers[-9].output,K.squeeze(model.layers[0].input,0),pLambda2)
pLambda1=self.ebMoleculeDense(model.layers[-10].output,model.layers[-9].weights[0],pdense1)
pin=self.ebMoleculeAdj(model.layers[-11].output,K.squeeze(model.layers[0].input,0),pLambda1)
mask1=K.squeeze(K.sum(pin,axis=2),0)
mask0=K.relu(mask0-mask1)
mask1=K.relu(mask1-mask0)
self.getMasks=K.function([model.layers[0].input,model.layers[1].input],[mask0,mask1])
def ebDense(self,activations,W,bottomP):
'''
This function calculates eb for a dense layer
Input:
activations: d-dimensional vector
W: Weights dxk-dimensional matrix
bottomP: k-dimensional probability vector
Output:
p: the probability of activation d-dimensional vector
'''
Wrelu=K.relu(W)
pcond=K.tf.matmul(K.tf.diag(activations),Wrelu)
pcond=pcond/K.sum(pcond,axis=0)
return K.transpose(K.tf.matmul(pcond,K.expand_dims(bottomP,1)))
def ebMoleculeDense(self,activations,W,bottomP):
'''
This function calculates eb for a dense layer
Input:
activations: 1x?xK
W: Weights dxk-dimensional matrix KxL
bottomP: probability matrix 1x?xL
Output:
p: probability matrix 1x?xK
'''
k,l=W.shape.as_list()
Wrelu=K.relu(W)
pcond=K.tile(K.expand_dims(activations,3),(1,1,1,l))*Wrelu
p=K.mean(K.tile(K.expand_dims(bottomP,2),(1,1,k,1))*pcond,3)
return p/K.sum(p)
def ebGAP(self,activations,bottomP):
'''
This function calculates eb for GAP layer
Input:
activations: 1x?xK
bottomP: probability matrix 1xK
Output:
p: probability matrix 1x?xK
'''
epsilon=1e-5
pcond=activations/(epsilon+K.sum(activations,axis=1))
p=pcond*K.squeeze(bottomP,0)
return p/K.sum(p)
def ebMoleculeAdj(self,activations,A,bottomP):
'''
This function calculates eb for a Adj conv layer
Input:
activations: 1x?xK
A: Adjacency ?x?
bottomP: probability matrix 1x?xK
Output:
p: probability matrix 1x?xK
'''
pcond=K.expand_dims(K.tf.matmul(A,K.squeeze(activations,0)),0)
p=pcond*bottomP
return p/K.sum(p)
class CAM:
def __init__(self, model):
self.weights=K.eval(model.layers[-2].weights[0])
self.bias=K.eval(model.layers[-2].weights[1])
self.tempModel=Model(model.input,model.layers[-5].output)
self.getMasks = self.getCAM
def getCAM(self, XY):
temp = np.matmul(self.tempModel.predict(XY), self.weights).squeeze() + self.bias
return (1*(temp>0)*temp).T
class GradCAM:
def __init__(self, model):
maskh0=self.getGradCamMask(model.layers[-2].output[0,0],model.layers[-5].output)
maskh1=self.getGradCamMask(model.layers[-2].output[0,1],model.layers[-5].output)
getMasks=K.function([model.layers[0].input,model.layers[1].input],[maskh0,maskh1])
self.getMasks = getMasks
def getGradCamMask(self,output,activation):
'''
This function calculates the importance weight reported in GradCam
Input:
ouput: The class output
activation: activation that we will take gradient with respect to
'''
grad=K.gradients(output,activation)[0]
alpha=K.squeeze(K.mean(grad,axis=1),0)
mask=K.squeeze(K.relu(K.sum(activation*alpha,axis=2)),0)
return mask
class GradCAMAvg:
def __init__(self, model):
maskInput0=self.getGradCamMask(model.layers[-2].output[0,0],model.layers[1].input)
maskInput1=self.getGradCamMask(model.layers[-2].output[0,1],model.layers[1].input)
tempMax0=K.max(K.stack([maskInput0,maskInput1]))
mask0h1=self.getGradCamMask(model.layers[-2].output[0,0],model.layers[-9].output)
mask1h1=self.getGradCamMask(model.layers[-2].output[0,1],model.layers[-9].output)
tempMax1=K.max(K.stack([mask0h1,mask1h1]))
mask0h2=self.getGradCamMask(model.layers[-2].output[0,0],model.layers[-7].output)
mask1h2=self.getGradCamMask(model.layers[-2].output[0,1],model.layers[-7].output)
tempMax2=K.max(K.stack([mask0h2,mask1h2]))
mask0h3=self.getGradCamMask(model.layers[-2].output[0,0],model.layers[-5].output)
mask1h3=self.getGradCamMask(model.layers[-2].output[0,1],model.layers[-5].output)
tempMax3=K.max(K.stack([mask0h3,mask1h3]))
getMasks=K.function([model.layers[0].input,model.layers[1].input],[maskInput0/tempMax0+mask0h1/tempMax1+mask0h2/tempMax2
+mask0h3/tempMax3,maskInput1/tempMax0+mask1h1/tempMax1+
mask1h2/tempMax2+mask1h3/tempMax3])
self.getMasks = getMasks
def getGradCamMask(self,output,activation):
'''
This function calculates the importance weight reported in GradCam
Input:
ouput: The class output
activation: activation that we will take gradient with respect to
'''
grad=K.gradients(output,activation)[0]
alpha=K.squeeze(K.mean(grad,axis=1),0)
mask=K.squeeze(K.relu(K.sum(activation*alpha,axis=2)),0)
return mask