-
Notifications
You must be signed in to change notification settings - Fork 1
/
test_relevance_propatation.py
211 lines (173 loc) · 8.01 KB
/
test_relevance_propatation.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
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import sys
import numpy as np
import tensorflow as tf
import relevance_propagation_vgg16 as rp
def assert_almost_equal(A,B, msg='', tol=1e-1, verbose=True):
assert A.shape==B.shape, "A.shape=%s B.shape=%s" % (A.shape, B.shape)
diff = np.sum(np.abs(A-B))
if verbose:
assert diff < tol, "FAIL %s: diff=%.4f\nA=%s\nB=%s" % (msg, diff,A,B)
else:
assert diff < tol, "FAIL %s: diff=%.4f" % (msg, diff)
class TestRelProp(object):
def __init__(self, sess):
self.sess=sess
self.naive = rp.Naive()
self.epsVariant = rp.EpsVariant(eps=1e-3)
self.betaVariant = rp.BetaVariant(beta=-.1)
class TestScalarFn(TestRelProp):
def __init__(self, sess, N=4):
TestRelProp.__init__(self, sess)
self.N=N
self.W=np.ones((self.N,4), dtype=np.float32)
self.W *= np.random.normal(size=(self.N,4))
self.X = tf.placeholder(tf.float32, (None, self.N))
self.Xarr = np.maximum(0,np.random.normal(size=(10,self.N)).astype(np.float32))
self.fn = 10.0;
self.test_naive(self.Xarr, np.maximum(.1, self.W), self.naive)
self.test_eps(self.Xarr, self.W, self.epsVariant)
self.test_beta(self.Xarr, self.W, self.betaVariant)
def test_naive(self, Xarr, Warr, rule):
tfR, tfZ_norm, tfZ = rp.relprop_from_scalar_fn(self.X[0], Warr[:,3],
self.fn, rule,
debug=True)
tfR_arr, tfZ_norm_arr, tfZ_arr =self.sess.run([tfR, tfZ_norm, tfZ],
feed_dict={self.X:Xarr})
Z = Xarr[0,:] * Warr[:,3]
Z_sum = np.sum(Z)
Rnaive = Z / Z_sum
Rnaive *= self.fn
assert_almost_equal(Z, tfZ_arr)
assert_almost_equal(Z_sum, tfZ_norm_arr)
assert_almost_equal(Rnaive, tfR_arr)
print("OK")
def test_eps(self, Xarr, Warr, rule):
tfR, tfZ_norm, tfZ = rp.relprop_from_scalar_fn(self.X[0], Warr[:,3], self.fn,
rule, debug=True)
tfR_arr, tfZ_norm_arr, tfZ_arr = self.sess.run([tfR, tfZ_norm, tfZ],
feed_dict={self.X:Xarr})
Z = Xarr[0,:] * Warr[:,3]
Z_sum = np.sum(Z)
Z_norm = (Z_sum + rule.eps * np.sign(Z_sum))
Reps = Z / Z_norm
Reps *= self.fn
assert_almost_equal(Z, tfZ_arr)
assert_almost_equal(Z_norm, tfZ_norm_arr, "eps: Z_norm != tfZ_norm_arr")
assert_almost_equal(Reps, tfR_arr)
def test_beta(self, Xarr, Warr, rule):
tfR, tfZneg_sum, tfZpos_sum, tfZneg, tfZpos, tfZ = \
rp.relprop_from_scalar_fn(self.X[0], Warr[:,3], self.fn, rule, debug=True)
tfR_arr, tfZneg_sum_arr, tfZpos_sum_arr, tfZneg_arr, tfZpos_arr, tfZ_arr = \
self.sess.run([tfR, tfZneg_sum, tfZpos_sum, tfZneg, tfZpos, tfZ],
feed_dict={self.X:Xarr})
Z = Xarr[0,:] * Warr[:,3]
Zpos = np.maximum(0,Z)
Zneg = np.minimum(0,Z)
Zpos_sum = np.sum(Zpos)
Zneg_sum = np.sum(Zneg)
Rbeta = np.zeros_like(Zpos)
if Zpos_sum > 1e-8:
Rbeta += Zpos * (rule.alpha/Zpos_sum)
if Zneg_sum < -1e-8:
Rbeta += Zneg * (rule.beta/Zneg_sum)
Rbeta *= self.fn
assert_almost_equal(Z, tfZ_arr)
assert_almost_equal(Rbeta, tfR_arr)
class TestDenseLayer(TestRelProp):
def __init__(self, sess, N=4, M=5):
TestRelProp.__init__(self, sess)
self.N=N
self.M=M
W=tf.ones((N,M), dtype=tf.float32)
W=tf.mul(W, np.random.normal(size=(N,M)).astype(np.float32))
RprevNaive = tf.ones((M,))
Rprev = tf.mul(tf.ones(M,), np.random.normal(size=(M,)).astype(np.float32))
self.X = tf.placeholder(tf.float32, (None, self.N))
Xarr = np.maximum(0.0,np.random.normal(size=(10,self.N)).astype(np.float32))
self.test_naive(Xarr, 0.1 + tf.nn.relu(W), RprevNaive, self.naive)
self.test_eps(Xarr, W, Rprev, self.epsVariant)
self.test_beta(Xarr, W, Rprev, self.betaVariant)
def test_naive(self, Xarr, W, Rprev, rule):
tfR, tfZ_norm, tfZ = rp.relprop_from_fully_connected(self.X[0], W,
Rprev, rule, debug=True)
tfR_arr, tfZ_norm_arr, tfZ_arr = self.sess.run([tfR, tfZ_norm, tfZ], feed_dict={self.X:Xarr})
W_arr = W.eval()
Rprev_arr = Rprev.eval()
Z_arr = np.reshape(Xarr[0,:],(self.N,1)) * W_arr
Z_reduce_arr = np.sum(Z_arr, axis=0)
Z_reduce_arr = np.reshape(Z_reduce_arr, (1,self.M))
Z_norm_arr = Z_arr / Z_reduce_arr
R_arr = np.matmul(Z_norm_arr, Rprev_arr)
assert_almost_equal(Z_arr, tfZ_arr)
assert_almost_equal(R_arr, tfR_arr)
def test_eps(self, Xarr, W, Rprev, rule):
tfR, tfZ_norm, tfZ = rp.relprop_from_fully_connected(self.X[0], W, Rprev,
rule, debug=True)
tfR_arr, tfZ_norm_arr, tfZ_arr = self.sess.run([tfR, tfZ_norm, tfZ], feed_dict={self.X:Xarr})
W_arr = W.eval()
Rprev_arr = Rprev.eval()
Z_arr = np.reshape(Xarr[0,:],(self.N,1)) * W_arr
Z_reduce_arr = np.sum(Z_arr, axis=0)
Z_reduce_arr += rule.eps * np.sign(Z_reduce_arr)
Z_reduce_arr = np.reshape(Z_reduce_arr, (1,self.M))
Z_norm_arr = Z_arr / Z_reduce_arr
R_arr = np.matmul(Z_norm_arr, Rprev_arr)
assert_almost_equal(Z_arr, tfZ_arr)
assert_almost_equal(R_arr, tfR_arr)
def test_beta(self, Xarr, W, Rprev, rule):
tfR = rp.relprop_from_fully_connected(self.X[0], W, Rprev, rule)
tfR_arr = self.sess.run(tfR, feed_dict={self.X:Xarr})
Warr = W.eval()
Rprev_arr = Rprev.eval()
Z_arr = np.reshape(Xarr[0,:],(self.N,1)) * Warr
Z_pos = np.maximum(Z_arr, 0)
Z_neg = np.minimum(Z_arr, 0)
Z_pos_reduce = np.reshape(np.maximum(1e-10, np.sum(Z_pos,0)), (1,self.M))
Z_neg_reduce = np.reshape(np.minimum(-1e-10, np.sum(Z_neg,0)), (1,self.M))
Z_norm = Z_pos * rule.alpha / Z_pos_reduce
Z_norm += Z_neg * rule.beta / Z_neg_reduce
R_arr = np.matmul(Z_norm, Rprev_arr)
assert_almost_equal(R_arr, tfR_arr)
class TestConv(TestRelProp):
def __init__(self, sess, N=4, M=5, Ch_a=2, Ch_b=3):
TestRelProp.__init__(self, sess)
K=tf.ones((2,2,Ch_a,Ch_b))
K *= np.random.normal(size=(2,2,Ch_a,Ch_b))
strides=(1,1,1,1)
X = tf.placeholder(tf.float32, (None, N, M, Ch_a))
Y = tf.nn.conv2d(X, K, strides, padding='SAME')
Rprev = tf.ones_like(Y[0])
Rprev *= np.random.normal(size=(map(int,Rprev.get_shape())))
R = rp.relprop_from_conv(X,Y,Rprev, self.epsVariant, K, strides)
Xarr = np.random.normal(size=(1, N, M, Ch_a))
Rarr = self.sess.run(R, feed_dict={X:Xarr})
print(R)
print(Rarr)
def test_relprop_max(sess, B=3,N=6,M=8,C=2):
X = tf.placeholder(tf.float32, (None,N,M,C))
Y = tf.nn.max_pool(X, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')
Rprev = tf.ones((1,3,4,2), tf.float32)
Rprev *= np.random.normal(size=(1,3,4,2))
R = rp.relprop_from_max(X,Y,Rprev)
Xarr = np.maximum(0.0, np.random.normal(size=(1,N,M,C)))
Xarr[0,0,0,0]=.3
Xarr[0,0,1,0]=.5
Xarr[0,1,0,0]=.6
Xarr[0,1,1,0]=.4
Rarr = sess.run(R,feed_dict={X:Xarr})
assert_almost_equal(np.sum(Rarr), np.sum(Rprev.eval()))
assert Rarr[0,0,0,0]==0
assert Rarr[0,0,1,0]==0
assert Rarr[0,1,0,0]==Rprev.eval()[0,0,0,0], "Rarr[0,1,0,0]=%.3f != Rprev[0,0,0,0]=%.3f" % (Rarr[0,1,0,0], Rprev.eval()[0,0,0,0])
assert Rarr[0,1,1,0]==0
def test_relprop(sess):
# TestScalarFn(sess)
# TestDenseLayer(sess)
test_relprop_max(sess)
TestConv(sess)
if __name__ == '__main__':
pass