-
Notifications
You must be signed in to change notification settings - Fork 2
/
net.py
102 lines (84 loc) · 4.96 KB
/
net.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
import numpy as np
import tensorflow as tf
from tensorflow.python.framework import ops
from tensorflow.python.ops import gen_nn_ops
class Net(object):
def __init__(self, config):
self.config = config
self.sess = tf.Session()
self.sess.__enter__()
self.graph = tf.get_default_graph()
# for the good of efficient guided-backprop
self.not_guided_flag = tf.placeholder(dtype=tf.float32, shape=[]) # 1 for normal, 0 for guided
self._add_guided_backprop()
self._build_net()
def _add_guided_backprop(self):
@ops.RegisterGradient("GuidedRelu")
def _GuidedReluGrad(op, grad):
return tf.where(0. < grad, gen_nn_ops.relu_grad(grad, op.outputs[0]),
gen_nn_ops.relu_grad(grad, op.outputs[0]) * self.not_guided_flag)
def _build_net(self):
with self.graph.gradient_override_map({'Relu': 'GuidedRelu'}):
self._vgg16()
self._fetch_shapes()
self._build_impact_gradient()
self._build_top_kmax_neuron_selection()
self._build_guided_backprop()
self._build_softmax_loss()
def _vgg16(self):
self.vgg16 = tf.keras.applications.vgg16
self.model = self.vgg16.VGG16()
self.feature_tensor = self.model.get_layer(self.config.vgg16.feature_layer).input
self.fc_features = self.model.get_layer('fc2').output
def _fetch_shapes(self):
self.features_shape = self.feature_tensor.get_shape().as_list()
self.classes_shape = self.model.output.get_shape().as_list()
self.features_shape[0] = self.classes_shape[0] = 1
def _build_impact_gradient(self):
self.category_indices = tf.argmax(self.model.output, axis=-1, name='max_scoring_categories')
fake_upstream_grad = tf.one_hot(self.category_indices, self.classes_shape[-1], axis=-1)
self.impact_grad = tf.gradients(self.model.output, self.feature_tensor,
grad_ys=[fake_upstream_grad], name='impact_gradients')[0]
def _build_top_kmax_neuron_selection(self):
# use DAM heuristic for selection
neurons_effect = self.impact_grad * self.feature_tensor
neurons_effect_flat_batch = tf.reshape(neurons_effect, (-1, np.prod(self.features_shape)))
self.batch_top_kmax_neuron_indices = tf.nn.top_k(neurons_effect_flat_batch, k=self.config.kmax)[1]
def _build_guided_backprop(self):
self.neuron_index = tf.placeholder(tf.int32, shape=[])
fake_upstream_grad = tf.one_hot(self.neuron_index, np.prod(self.features_shape), axis=-1)
fake_upstream_grad = tf.reshape(fake_upstream_grad, shape=self.features_shape)
self.guided_backprop = tf.gradients(self.feature_tensor, self.model.input,
grad_ys=[fake_upstream_grad], name='guided_backprop')
def _build_softmax_loss(self):
self.top_class_index_ph = tf.placeholder(tf.int32, shape=[])
self.top_class_batch_one_hot = tf.one_hot(tf.ones([self.config.kmax, ], dtype=tf.int32) * self.top_class_index_ph,
self.classes_shape[-1], axis=-1)
self.softmax_loss = tf.nn.softmax_cross_entropy_with_logits_v2(labels=self.top_class_batch_one_hot,
logits=self.model.output)
def get_top_kmax_neurons(self, images):
top_kmax_neurons_indices, max_scoring_indices = self.sess.run([self.batch_top_kmax_neuron_indices,
self.category_indices],
feed_dict={self.model.input: images,
self.not_guided_flag: 1.0})
return top_kmax_neurons_indices, max_scoring_indices
def get_guided_backprop(self, image, neuron_index):
numerical_guided_backprop = self.sess.run(self.guided_backprop,
feed_dict={
self.model.input: image,
self.neuron_index: neuron_index,
self.not_guided_flag: 0.0,
})
return numerical_guided_backprop[0]
def get_batch_loss(self, images, top_class):
assert images.shape[0] is self.config.kmax
batch_loss = self.sess.run(self.softmax_loss,
feed_dict={
self.top_class_index_ph: top_class,
self.model.input: images,
})
return batch_loss
def get_fc_features(self, images):
fc_features = self.sess.run(self.fc_features,
feed_dict={self.model.input: images})
return fc_features