-
Notifications
You must be signed in to change notification settings - Fork 8
/
distill.py
190 lines (144 loc) · 8.04 KB
/
distill.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
import tensorflow as tf
import numpy as np
from tensorflow.examples.tutorials.mnist import input_data
batch_size = 128
test_size = 1000
donor_name = 'conv_net_3x3-61'
acceptor_name = 'distil_comb_L2'
def init_weights(shape):
return tf.Variable(tf.random_normal(shape, stddev=0.01))
X = tf.placeholder("float", [None, 28, 28, 1])
Y = tf.placeholder("float", [None, 10])
p_keep_conv = tf.placeholder("float")
p_keep_hidden = tf.placeholder("float")
T = 1
# lenet
def lenet4_model(X, w, w2, w3, w4, w_o, p_keep_conv, p_keep_hidden):
l1a = tf.nn.relu(tf.nn.conv2d(X, w, # l1a shape=(?, 28, 28, 32)
strides=[1, 1, 1, 1], padding='SAME'))
l1 = tf.nn.max_pool(l1a, ksize=[1, 2, 2, 1], # l1 shape=(?, 14, 14, 32)
strides=[1, 2, 2, 1], padding='SAME')
l1 = tf.nn.dropout(l1, p_keep_conv)
l2a = tf.nn.relu(tf.nn.conv2d(l1, w2, # l2a shape=(?, 14, 14, 64)
strides=[1, 1, 1, 1], padding='SAME'))
l2 = tf.nn.max_pool(l2a, ksize=[1, 2, 2, 1], # l2 shape=(?, 7, 7, 64)
strides=[1, 2, 2, 1], padding='SAME')
l2 = tf.nn.dropout(l2, p_keep_conv)
l3a = tf.nn.relu(tf.nn.conv2d(l2, w3, # l3a shape=(?, 7, 7, 128)
strides=[1, 1, 1, 1], padding='SAME'))
l3 = tf.nn.max_pool(l3a, ksize=[1, 2, 2, 1], # l3 shape=(?, 4, 4, 128)
strides=[1, 2, 2, 1], padding='SAME')
l3 = tf.reshape(l3, [-1, w4.get_shape().as_list()[0]]) # reshape to (?, 2048)
l3 = tf.nn.dropout(l3, p_keep_conv)
l4 = tf.nn.relu(tf.matmul(l3, w4))
l4 = tf.nn.dropout(l4, p_keep_hidden)
pyx = tf.matmul(l4, w_o)
return pyx
def lenet4():
w = init_weights([3, 3, 1, 32]) # 3x3x1 conv, 32 outputs
w2 = init_weights([3, 3, 32, 64]) # 3x3x32 conv, 64 outputs
w3 = init_weights([3, 3, 64, 128]) # 3x3x32 conv, 128 outputs
w4 = init_weights([128 * 4 * 4, 625]) # FC 128 * 4 * 4 inputs, 625 outputs
w_o = init_weights([625, 10]) # FC 625 inputs, 10 outputs (labels)
return tf.nn.softmax((1.0/T) * lenet4_model(X, w, w2, w3, w4, w_o, p_keep_conv, p_keep_hidden)), [w,w2,w3,w4,w_o]
# fully-connected
def fc_layer(inp, size, name):
W_layer = tf.Variable(tf.truncated_normal([inp.get_shape()[1].value, size],stddev=0.1))
b_layer = tf.Variable(tf.constant(0, tf.float32, shape=[size]))
res = tf.matmul(inp, W_layer) + b_layer
tf.histogram_summary(name + "_weights", W_layer)
return res, [W_layer, b_layer]
def fully_connected():
L1, fc_params1 = fc_layer(tf.reshape(X, [-1,784]) , 100, "L1")
L1a = tf.nn.sigmoid(L1)
res, fc_params2 = fc_layer(L1a , 10, "L2")
return tf.nn.softmax((1.0/T) * res), fc_params1 + fc_params2
# donor & acceptor networks
y_donor, donor_params = lenet4()
y_acceptor, acceptor_params = fully_connected()
# distillation
distill_ent = tf.reduce_mean( - tf.reduce_sum(tf.stop_gradient(y_donor) * tf.log(y_acceptor), reduction_indices=1))
truth_ent = tf.reduce_mean( - tf.reduce_sum(Y * tf.log(y_acceptor), reduction_indices=1))
distil_step = tf.train.GradientDescentOptimizer(0.1).minimize(distill_ent + truth_ent)
# Summaries
def prec(y_pred):
return tf.reduce_mean(tf.cast(tf.equal(tf.argmax(y_pred, 1), tf.argmax(Y, 1)), tf.float32))
# MNIST data
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
trX, trY, teX, teY = mnist.train.images, mnist.train.labels, mnist.test.images, mnist.test.labels
trX = trX.reshape(-1, 28, 28, 1) # 28x28x1 input img
teX = teX.reshape(-1, 28, 28, 1) # 28x28x1 input img
# distillate knowledge from donor to acceptor
def distillate(net_name):
acceptor_prec = prec(y_acceptor)
donor_prec = prec(y_donor)
distil_writer = tf.train.SummaryWriter("logs/" + net_name + "/train", flush_secs=5)
distil_test_writer = tf.train.SummaryWriter("logs/" + net_name + "/test", flush_secs=5)
tf.scalar_summary('accuracy', acceptor_prec)
summaries = tf.merge_all_summaries()
# Launch the graph in a session
with tf.Session() as sess:
# you need to initialize all variables
tf.initialize_all_variables().run()
# Donor loading
donor_saver = tf.train.Saver(donor_params)
donor_saver.restore(sess, 'checkpoints/' + donor_name)
acc_saver = tf.train.Saver(acceptor_params)
k = 1
for i in range(10):
training_batch = zip(range(0, len(trX), batch_size),
range(batch_size, len(trX)+1, batch_size))
for start, end in training_batch:
val_prec, val_donor_prec, log_summaries, val_distil_ent, val_truth_ent, _ = sess.run(
[acceptor_prec, donor_prec, summaries, distill_ent, truth_ent, distil_step],
feed_dict={X: trX[start:end],
Y: trY[start:end],
p_keep_conv: 1,
p_keep_hidden: 1 })
test_val_prec, test_log_summaries = sess.run([acceptor_prec, summaries],
feed_dict={X: teX[:test_size],
Y: teY[:test_size],
p_keep_conv: 1.0,
p_keep_hidden: 1.0})
distil_writer.add_summary(log_summaries, k)
distil_test_writer.add_summary(test_log_summaries, k)
print(i, k, 'distillation distil_ent:', val_distil_ent, 'truth_ent', val_truth_ent, 'donor_prec: ', val_donor_prec, '; train_prec', val_prec, '; test_prec', test_val_prec)
k = k + 1
acc_saver.save(sess, "checkpoints/" + net_name, global_step = i)
# train net using back-prop
def train_net(train_step, net_prec, net_params, net_name):
net_prec = prec(y_acceptor)
writer = tf.train.SummaryWriter("logs/" + net_name + "/train", flush_secs=5)
test_writer = tf.train.SummaryWriter("logs/" + net_name + "/test", flush_secs=5)
tf.scalar_summary('accuracy', net_prec)
summaries = tf.merge_all_summaries()
# Launch the graph in a session
with tf.Session() as sess:
# you need to initialize all variables
tf.initialize_all_variables().run()
net_saver = tf.train.Saver(net_params)
k = 1
for i in range(10):
training_batch = zip(range(0, len(trX), batch_size),
range(batch_size, len(trX)+1, batch_size))
for start, end in training_batch:
val_prec, log_summaries, _ = sess.run([net_prec, summaries, train_step],
feed_dict={X: trX[start:end],
Y: trY[start:end],
p_keep_conv: 1,
p_keep_hidden: 1 })
test_val_prec, test_log_summaries = sess.run([net_prec, summaries],
feed_dict={X: teX[:test_size],
Y: teY[:test_size],
p_keep_conv: 1,
p_keep_hidden: 1 })
writer.add_summary(log_summaries, k)
test_writer.add_summary(test_log_summaries, k)
print(i, k, 'train_prec', val_prec, '; test_prec', test_val_prec)
k = k + 1
net_saver.save(sess, "checkpoints/" + net_name, global_step = i)
train_cross_ent = tf.reduce_mean( - tf.reduce_sum(Y * tf.log(y_acceptor), reduction_indices=1))
train_prec = prec(y_acceptor)
train_step = tf.train.GradientDescentOptimizer(0.1).minimize(train_cross_ent)
distillate(acceptor_name)
#train_net(train_step, train_prec, acceptor_params, "L2")