-
Notifications
You must be signed in to change notification settings - Fork 1
/
model_cifar.py
240 lines (199 loc) · 8.54 KB
/
model_cifar.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
#####
## MODIFIED BY: Edouard Oyallon
## Team DATA - ENS 2016
## Can be found on: https://github.com/bgshih/tf_resnet_cifar
#####
from __future__ import division
import math
import tensorflow as tf
import numpy as np
import joblib
import model_utils as mu
FLAGS = tf.app.flags.FLAGS
def one_hot_embedding(label, n_classes):
"""
One-hot embedding
Args:
label: int32 tensor [B]
n_classes: int32, number of classes
Return:
embedding: tensor [B x n_classes]
"""
embedding_params = np.eye(n_classes, dtype=np.float32)
with tf.device('/cpu:0'):
params = tf.constant(embedding_params)
embedding = tf.gather(params, label)
return embedding
def conv2d(x, n_in, n_out, k, s,p='SAME', bias=False, phase_train=True,scope='conv'):
with tf.variable_scope(scope):
kernel = tf.Variable(
tf.truncated_normal([k, k, n_in, n_out],
stddev=math.sqrt(8/(k*k*n_out))),
name='weight')
tf.add_to_collection('weights', kernel)
conv = tf.nn.conv2d(x, kernel, [1,s,s,1], padding=p)
return conv
def batch_norm(x, n_out, phase_train, scope='bn', affine=True):
"""
Batch normalization on convolutional maps.
Args:
x: Tensor, 4D BHWD input maps
n_out: integer, depth of input maps
phase_train: boolean tf.Variable, true indicates training phase
scope: string, variable scope
affine: whether to affine-transform outputs
Return:
normed: batch-normalized maps
"""
with tf.variable_scope(scope):
beta = tf.constant(0.0, shape=[n_out])
gamma = tf.constant(1.0, shape=[n_out])
batch_mean, batch_var = tf.nn.moments(x, [0,1,2], name='moments')
ema = tf.train.ExponentialMovingAverage(decay=0.99)
def mean_var_with_update():
ema_apply_op = ema.apply([batch_mean, batch_var])
with tf.control_dependencies([ema_apply_op]):
return tf.identity(batch_mean), tf.identity(batch_var)
mean, var = tf.cond(phase_train,
mean_var_with_update,
lambda: (ema.average(batch_mean), ema.average(batch_var)))
normed = tf.nn.batch_norm_with_global_normalization(x, mean, var,
beta, gamma, 1e-3, affine)
return normed
def block(x, n_in, n_out, subsample, alpha, phase_train, scope='res_block'):
with tf.variable_scope(scope):
n_out_1 = math.floor(alpha * n_out)
n_out_2 = n_out - n_out_1
if subsample:
y = conv2d(x, n_in, n_out, 3, 2, 'SAME', False,phase_train, scope='conv_1')
else:
y = conv2d(x, n_in, n_out, 3, 1, 'SAME', False,phase_train, scope='conv_1')
y = batch_norm(y, n_out, phase_train, scope='bn_1')
with tf.variable_scope('nonlinearity_1'):
y_1 = tf.slice(y,[0,0,0,0],[-1,-1,-1,n_out_1])
y_2 = tf.slice(y, [0,0,0,n_out_1],[-1,-1,-1,n_out_2])
y_1 = tf.nn.relu(y_1)
y = tf.concat(3,[y_1,y_2])
tf.histogram_summary('activations/' + y.op.name, y)
tf.scalar_summary('sparsity/' + y.op.name, tf.nn.zero_fraction(y))
y = conv2d(y, n_out, n_out, 3, 1, 'SAME', False, phase_train, scope='conv_2')
y = batch_norm(y, n_out, phase_train, scope='bn_2')
with tf.variable_scope('nonlinearity_2'):
y_1 = tf.slice(y,[0,0,0,0],[-1,-1,-1,n_out_1])
y_2 = tf.slice(y, [0,0,0,n_out_1],[-1,-1,-1,n_out_2])
y_1 = tf.nn.relu(y_1)
y = tf.concat(3,[y_1,y_2])
tf.histogram_summary('activations/' + y.op.name, y)
tf.scalar_summary('sparsity/' + y.op.name, tf.nn.zero_fraction(y))
return y
def group(x, n_in, n_out, n, first_subsample, alpha,phase_train, scope='group'):
with tf.variable_scope(scope):
y = block(x, n_in, n_out, first_subsample, alpha,phase_train, scope='block_1')
y = tf.cond(phase_train, lambda: tf.nn.dropout(y, 0.6), lambda: y)#0.6 : 89.5
for i in range(n - 1):
y = block(y, n_out, n_out, False, alpha,phase_train, scope='block_%d' % (i + 2))
y = tf.cond(phase_train, lambda: tf.nn.dropout(y, 0.6), lambda: y)
return y
def net(x, n_layer_per_block, n_classes, phase_train,alpha,number_channel, scope='deep_net'):
with tf.variable_scope(scope):
n1=number_channel
n2=number_channel
n3=number_channel
n4=number_channel
y = conv2d(x, 3, n1, 3, 1, 'SAME',False, phase_train,scope='conv_init')
y = batch_norm(y, n1, phase_train, scope='bn_init')
y = tf.nn.relu(y, name='relu_init')
y = group(y, n1, n2, n_layer_per_block, False, alpha,phase_train, scope='group_1')
y = group(y, n2, n3, n_layer_per_block, True,alpha, phase_train, scope='group_2')
y = group(y, n3, n4, n_layer_per_block, True,alpha, phase_train, scope='group_3')
y = tf.nn.avg_pool(y, [1, 8, 8, 1], [1, 1, 1, 1], 'VALID', name='avg_pool')
y = tf.squeeze(y, squeeze_dims=[1, 2])
w = tf.get_variable('DW', [n4, n_classes],initializer=tf.uniform_unit_scaling_initializer(factor=1.0))
tf.add_to_collection('weights', w)
bias = tf.get_variable('bias', [n_classes], initializer=tf.constant_initializer(0.0))
y=tf.nn.xw_plus_b(y, w, bias)
return y
def loss(logits, labels,n_class, scope='loss'):
with tf.variable_scope(scope):
# entropy loss
targets = one_hot_embedding(labels, n_class)
entropy_loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits, targets),
name='entropy_loss')
tf.add_to_collection('losses', entropy_loss)
# weight l2 decay loss
weight_l2_losses = [tf.nn.l2_loss(o) for o in tf.get_collection('weights')]
weight_decay_loss = tf.mul(FLAGS.weight_decay, tf.add_n(weight_l2_losses),
name='weight_decay_loss')
tf.add_to_collection('losses', weight_decay_loss)
for var in tf.get_collection('losses'):
tf.scalar_summary('losses/' + var.op.name, var)
# total loss
return tf.add_n(tf.get_collection('losses'), name='total_loss')
def accuracy(logits, gt_label, scope='accuracy'):
with tf.variable_scope(scope):
pred_label = tf.argmax(logits, 1)
acc = 1.0 - tf.nn.zero_fraction(
tf.cast(tf.equal(pred_label, gt_label), tf.int32))
return acc
def train_op(loss, global_step, learning_rate):
params = tf.trainable_variables()
gradients = tf.gradients(loss, params, name='gradients')
optim = tf.train.MomentumOptimizer(learning_rate, 0.9)
update = optim.apply_gradients(zip(gradients, params))
with tf.control_dependencies([update]):
train_op = tf.no_op(name='train_op')
return train_op
def cifar_input_stream(records_path):
reader = tf.TFRecordReader()
filename_queue = tf.train.string_input_producer([records_path], None)
_, record_value = reader.read(filename_queue)
features = tf.parse_single_example(record_value,
{
'image_raw': tf.FixedLenFeature([], tf.string),
'label': tf.FixedLenFeature([], tf.int64),
})
image = tf.decode_raw(features['image_raw'], tf.uint8)
image = tf.reshape(image, [32,32,3])
image = tf.cast(image, tf.float32)
label = tf.cast(features['label'], tf.int64)
return image, label
def normalize_image(image):
meanstd = joblib.load(FLAGS.mean_std_path)
mean, std = meanstd['mean'], meanstd['std']
normed_image = (image - mean) / std
return normed_image
def random_distort_image(image):
image = tf.image.resize_image_with_crop_or_pad(image, 36, 36)
image = tf.random_crop(image, [32, 32, 3])
image = tf.image.random_flip_left_right(image)
return image
def make_train_batch(train_records_path, batch_size):
with tf.variable_scope('train_batch'):
with tf.device('/cpu:0'):
train_image, train_label = cifar_input_stream(train_records_path)
train_image = normalize_image(train_image)
train_image = random_distort_image(train_image)
train_image_batch, train_label_batch = tf.train.shuffle_batch(
[train_image, train_label], batch_size=batch_size, num_threads=4,
capacity=50000,
min_after_dequeue=1000)
return train_image_batch, train_label_batch
def make_validation_train_batch(train_records_path, batch_size):
with tf.variable_scope('eval_train_batch'):
with tf.device('/cpu:0'):
train_image, train_label = cifar_input_stream(train_records_path)
train_image = normalize_image(train_image)
train_image_batch, train_label_batch = tf.train.batch(
[train_image, train_label], batch_size=batch_size, num_threads=1,
capacity=50000)
return train_image_batch, train_label_batch
def make_validation_batch(test_records_path, batch_size):
with tf.variable_scope('evaluate_batch'):
with tf.device('/cpu:0'):
test_image, test_label = cifar_input_stream(test_records_path)
test_image = normalize_image(test_image)
test_image_batch, test_label_batch = tf.train.batch(
[test_image, test_label], batch_size=batch_size, num_threads=1,
capacity=10000)
return test_image_batch, test_label_batch