-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathcifar100_conv.py
198 lines (142 loc) · 6.75 KB
/
cifar100_conv.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
import argparse
import os
import sys
##############################################
parser = argparse.ArgumentParser()
parser.add_argument('--epochs', type=int, default=100)
parser.add_argument('--batch_size', type=int, default=64)
parser.add_argument('--lr', type=float, default=1e-4)
parser.add_argument('--eps', type=float, default=1e-5)
parser.add_argument('--dropout', type=float, default=0.5)
parser.add_argument('--act', type=str, default='relu')
parser.add_argument('--bias', type=float, default=0.)
parser.add_argument('--gpu', type=int, default=0)
parser.add_argument('--dfa', type=int, default=0)
parser.add_argument('--sparse', type=int, default=0)
parser.add_argument('--rank', type=int, default=0)
parser.add_argument('--init', type=str, default="glorot_uniform")
parser.add_argument('--save', type=int, default=0)
parser.add_argument('--name', type=str, default="cifar100_conv")
parser.add_argument('--load', type=str, default=None)
args = parser.parse_args()
if args.gpu >= 0:
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"]=str(args.gpu)
##############################################
import tensorflow as tf
import keras
import numpy as np
from lib.Model import Model
from lib.Layer import Layer
from lib.ConvToFullyConnected import ConvToFullyConnected
from lib.FullyConnected import FullyConnected
from lib.Convolution import Convolution
from lib.MaxPool import MaxPool
from lib.Dropout import Dropout
from lib.FeedbackFC import FeedbackFC
from lib.FeedbackConv import FeedbackConv
from lib.Activation import Relu
from lib.Activation import Tanh
##############################################
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar100.load_data()
train_examples = 50000
test_examples = 10000
assert(np.shape(x_train) == (train_examples, 32, 32, 3))
y_train = keras.utils.to_categorical(y_train, 100)
assert(np.shape(x_test) == (test_examples, 32, 32, 3))
y_test = keras.utils.to_categorical(y_test, 100)
##############################################
if args.act == 'tanh':
act = Tanh()
elif args.act == 'relu':
act = Relu()
else:
assert(False)
##############################################
tf.set_random_seed(0)
tf.reset_default_graph()
batch_size = tf.placeholder(tf.int32, shape=())
dropout_rate = tf.placeholder(tf.float32, shape=())
lr = tf.placeholder(tf.float32, shape=())
X = tf.placeholder(tf.float32, [None, 32, 32, 3])
X = tf.map_fn(lambda frame: tf.image.per_image_standardization(frame), X)
Y = tf.placeholder(tf.float32, [None, 100])
l0 = Convolution(input_shape=[batch_size, 32, 32, 3], filter_sizes=[5, 5, 3, 96], init=args.init, activation=act, bias=args.bias, name='conv1')
l1 = MaxPool(size=[batch_size, 32, 32, 96], ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding="SAME")
l2 = FeedbackConv(size=[batch_size, 16, 16, 96], num_classes=100, sparse=args.sparse, rank=args.rank, name='conv1_fb')
l3 = Convolution(input_shape=[batch_size, 16, 16, 96], filter_sizes=[5, 5, 96, 128], init=args.init, activation=act, bias=args.bias, name='conv2')
l4 = MaxPool(size=[batch_size, 16, 16, 128], ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding="SAME")
l5 = FeedbackConv(size=[batch_size, 8, 8, 128], num_classes=100, sparse=args.sparse, rank=args.rank, name='conv2_fb')
l6 = Convolution(input_shape=[batch_size, 8, 8, 128], filter_sizes=[5, 5, 128, 256], init=args.init, activation=act, bias=args.bias, name='conv3')
l7 = MaxPool(size=[batch_size, 8, 8, 256], ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding="SAME")
l8 = FeedbackConv(size=[batch_size, 4, 4, 256], num_classes=100, sparse=args.sparse, rank=args.rank, name='conv3_fb')
l9 = ConvToFullyConnected(input_shape=[4, 4, 256])
l10 = FullyConnected(input_shape=4*4*256, size=2048, init=args.init, activation=act, bias=args.bias, name='fc1')
l11 = Dropout(rate=dropout_rate)
l12 = FeedbackFC(size=[4*4*256, 2048], num_classes=100, sparse=args.sparse, rank=args.rank, name='fc1_fb')
l13 = FullyConnected(input_shape=2048, size=2048, init=args.init, activation=act, bias=args.bias, name='fc2')
l14 = Dropout(rate=dropout_rate)
l15 = FeedbackFC(size=[2048, 2048], num_classes=100, sparse=args.sparse, rank=args.rank, name='fc2_fb')
l16 = FullyConnected(input_shape=2048, size=100, init=args.init, bias=args.bias, name='fc3')
##############################################
model = Model(layers=[l0, l1, l2, l3, l4, l5, l6, l7, l8, l9, l10, l11, l12, l13, l14, l15, l16])
predict = model.predict(X=X)
weights = model.get_weights()
if args.dfa:
grads_and_vars = model.dfa_gvs(X=X, Y=Y)
else:
grads_and_vars = model.gvs(X=X, Y=Y)
train = tf.train.AdamOptimizer(learning_rate=lr, epsilon=args.eps).apply_gradients(grads_and_vars=grads_and_vars)
correct = tf.equal(tf.argmax(predict,1), tf.argmax(Y,1))
total_correct = tf.reduce_sum(tf.cast(correct, tf.float32))
##############################################
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
tf.local_variables_initializer().run()
##############################################
filename = args.name + '.results'
f = open(filename, "w")
f.write(filename + "\n")
f.write("total params: " + str(model.num_params()) + "\n")
f.close()
##############################################
train_accs = []
test_accs = []
for ii in range(args.epochs):
#############################
_total_correct = 0
for jj in range(0, train_examples, args.batch_size):
s = jj
e = min(jj + args.batch_size, train_examples)
b = e - s
xs = x_train[s:e]
ys = y_train[s:e]
_correct, _ = sess.run([total_correct, train], feed_dict={batch_size: b, dropout_rate: args.dropout, lr: args.lr, X: xs, Y: ys})
_total_correct += _correct
train_acc = 1.0 * _total_correct / (train_examples - (train_examples % args.batch_size))
train_accs.append(train_acc)
#############################
_total_correct = 0
for jj in range(0, test_examples, args.batch_size):
s = jj
e = min(jj + args.batch_size, test_examples)
b = e - s
xs = x_test[s:e]
ys = y_test[s:e]
_correct = sess.run(total_correct, feed_dict={batch_size: b, dropout_rate: 0.0, lr: 0.0, X: xs, Y: ys})
_total_correct += _correct
test_acc = 1.0 * _total_correct / (test_examples - (test_examples % args.batch_size))
test_accs.append(test_acc)
#############################
p = "%d | train acc: %f | test acc: %f" % (ii, train_acc, test_acc)
print (p)
f = open(filename, "a")
f.write(p + "\n")
f.close()
##############################################
if args.save:
[w] = sess.run([weights], feed_dict={})
w['train_acc'] = train_accs
w['test_acc'] = test_accs
np.save(args.name, w)
##############################################