-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathalexnet.py
165 lines (152 loc) · 6.42 KB
/
alexnet.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
import mxnet as mx
import time
from mxnet import nd, autograd
from mxnet import gluon
dshape = (128, 3, 224, 224)
def mxnet_symbol():
input_data = mx.symbol.Variable(name="data")
# stage 1
conv1 = mx.symbol.Convolution(
data=input_data, kernel=(11, 11), stride=(4, 4), num_filter=64)
relu1 = mx.symbol.Activation(data=conv1, act_type="relu")
pool1 = mx.symbol.Pooling(
data=relu1, pool_type="max", kernel=(3, 3), stride=(2,2))
# stage 2
conv2 = mx.symbol.Convolution(
data=pool1, kernel=(5, 5), pad=(2, 2), num_filter=192)
relu2 = mx.symbol.Activation(data=conv2, act_type="relu")
pool2 = mx.symbol.Pooling(data=relu2, kernel=(3, 3), stride=(2, 2),
pool_type="max")
# stage 3
conv3 = mx.symbol.Convolution(
data=pool2, kernel=(3, 3), pad=(1, 1), num_filter=384)
relu3 = mx.symbol.Activation(data=conv3, act_type="relu")
conv4 = mx.symbol.Convolution(
data=relu3, kernel=(3, 3), pad=(1, 1), num_filter=256)
relu4 = mx.symbol.Activation(data=conv4, act_type="relu")
conv5 = mx.symbol.Convolution(
data=relu4, kernel=(3, 3), pad=(1, 1), num_filter=256)
relu5 = mx.symbol.Activation(data=conv5, act_type="relu")
pool3 = mx.symbol.Pooling(data=relu5, kernel=(3, 3), stride=(2, 2),
pool_type="max")
# stage 4
flatten = mx.symbol.Flatten(data=pool3)
fc1 = mx.symbol.FullyConnected(data=flatten, num_hidden=4096)
relu6 = mx.symbol.Activation(data=fc1, act_type="relu")
# stage 5
fc2 = mx.symbol.FullyConnected(data=relu6, num_hidden=4096)
relu7 = mx.symbol.Activation(data=fc2, act_type="relu")
# stage 6
fc3 = mx.symbol.FullyConnected(data=relu7, num_hidden=10)
return input_data, fc3
def time_mxnet_symbol(n=100):
dbatch = mx.nd.random.uniform(-1, 1, shape=dshape,
dtype='float32', ctx=mx.gpu())
_, fc3 = mxnet_symbol()
sym_exec = fc3.simple_bind(ctx=mx.gpu(), data=dshape)
grad = mx.nd.ones((128, 10), ctx=mx.gpu())
tic = time.time()
for _ in range(n):
sym_exec.forward(is_train=True, data=dbatch)
sym_exec.backward([grad])
mx.nd.waitall()
toc = time.time()
return (toc - tic) / n
def gluon_symbolblock(n=100):
inputs, sym = mxnet_symbol()
model = gluon.SymbolBlock(sym, inputs)
model.collect_params().initialize(mx.init.One(), ctx=mx.gpu())
dbatch = mx.nd.random.uniform(-1, 1, shape=dshape,
dtype='float32', ctx=mx.gpu())
tic = time.time()
for _ in range(n):
with autograd.record():
out = model(dbatch)
out.backward()
mx.nd.waitall()
toc = time.time()
return (toc - tic) / n
def gluon_hybridblock(n=100, hybridize=True):
alex_net = gluon.nn.HybridSequential()
with alex_net.name_scope():
# First convolutional layer
alex_net.add(gluon.nn.Conv2D(channels=64, kernel_size=11,
strides=(4,4), activation='relu'))
alex_net.add(gluon.nn.MaxPool2D(pool_size=3, strides=2))
# Second convolutional layer
alex_net.add(gluon.nn.Conv2D(channels=192, kernel_size=5,
padding=(2, 2), activation='relu'))
alex_net.add(gluon.nn.MaxPool2D(pool_size=3, strides=(2,2)))
# Third convolutional layer
alex_net.add(gluon.nn.Conv2D(channels=384, kernel_size=3,
padding=(1, 1), activation='relu'))
# Fourth convolutional layer
alex_net.add(gluon.nn.Conv2D(channels=256, kernel_size=3,
padding=(1, 1), activation='relu'))
# Fifth convolutional layer
alex_net.add(gluon.nn.Conv2D(channels=256, kernel_size=3,
padding=(1, 1), activation='relu'))
alex_net.add(gluon.nn.MaxPool2D(pool_size=3, strides=2))
# Flatten and apply fullly connected layers
alex_net.add(gluon.nn.Flatten())
alex_net.add(gluon.nn.Dense(4096, activation="relu"))
alex_net.add(gluon.nn.Dense(4096, activation="relu"))
alex_net.add(gluon.nn.Dense(10))
if hybridize:
alex_net.hybridize()
alex_net.collect_params().initialize(mx.init.One(), ctx=mx.gpu())
dbatch = mx.nd.random.uniform(-1, 1, shape=dshape,
dtype='float32', ctx=mx.gpu())
tic = time.time()
for _ in range(n):
with autograd.record():
out = alex_net(dbatch)
out.backward()
mx.nd.waitall()
toc = time.time()
return (toc - tic) / n
def gluon_block(n=100):
alex_net = gluon.nn.Sequential()
with alex_net.name_scope():
# First convolutional layer
alex_net.add(gluon.nn.Conv2D(channels=64, kernel_size=11,
strides=(4,4), activation='relu'))
alex_net.add(gluon.nn.MaxPool2D(pool_size=3, strides=2))
# Second convolutional layer
alex_net.add(gluon.nn.Conv2D(channels=192, kernel_size=5,
padding=(2, 2), activation='relu'))
alex_net.add(gluon.nn.MaxPool2D(pool_size=3, strides=(2,2)))
# Third convolutional layer
alex_net.add(gluon.nn.Conv2D(channels=384, kernel_size=3,
padding=(1, 1), activation='relu'))
# Fourth convolutional layer
alex_net.add(gluon.nn.Conv2D(channels=256, kernel_size=3,
padding=(1, 1), activation='relu'))
# Fifth convolutional layer
alex_net.add(gluon.nn.Conv2D(channels=256, kernel_size=3,
padding=(1, 1), activation='relu'))
alex_net.add(gluon.nn.MaxPool2D(pool_size=3, strides=2))
# Flatten and apply fullly connected layers
alex_net.add(gluon.nn.Flatten())
alex_net.add(gluon.nn.Dense(4096, activation="relu"))
alex_net.add(gluon.nn.Dense(4096, activation="relu"))
alex_net.add(gluon.nn.Dense(10))
alex_net.collect_params().initialize(mx.init.One(), ctx=mx.gpu())
dbatch = mx.nd.random.uniform(-1, 1, shape=dshape,
dtype='float32', ctx=mx.gpu())
tic = time.time()
for _ in range(n):
with autograd.record():
out = alex_net(dbatch)
out.backward()
mx.nd.waitall()
toc = time.time()
return (toc - tic) / n
def main(n=100):
print('gluon SymbolBlock run time: {}'.format(gluon_symbolblock(n)))
print('gluon HybridBlock hyridized run time: {}'.format(gluon_hybridblock(n)))
print('gluon HybridBlock unhyridized run time: {}'.format(
gluon_hybridblock(n, hybridize=False)))
print('gluon Block run time: {}'.format(gluon_block(n)))
print('mxnet symbol run time: {}'.format(time_mxnet_symbol(n)))
main()