-
Notifications
You must be signed in to change notification settings - Fork 44
/
Copy pathsymbol_densenet.py
176 lines (168 loc) · 8.85 KB
/
symbol_densenet.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
"""
Adapted from https://github.com/tornadomeet/ResNet/blob/master/symbol_resnet.py
Original author Wei Wu
Referenced https://github.com/bamos/densenet.pytorch/blob/master/densenet.py
Original author bamos
Referenced https://github.com/andreasveit/densenet-pytorch/blob/master/densenet.py
Original author andreasveit
Referenced https://github.com/Nicatio/Densenet/blob/master/mxnet/symbol_densenet.py
Original author Nicatio
Implemented the following paper: DenseNet-BC
Gao Huang, Zhuang Liu, Kilian Q. Weinberger, Laurens van der Maaten. "Densely Connected Convolutional Networks"
Coded by Lin Xiong Mar-1, 2017
"""
import mxnet as mx
import math
def BasicBlock(data, growth_rate, stride, name, bottle_neck=True, drop_out=0.0, bn_mom=0.9, workspace=512):
"""Return BaiscBlock Unit symbol for building DenseBlock
Parameters
----------
data : str
Input data
growth_rate : int
Number of output channels
stride : tupe
Stride used in convolution
drop_out : float
Probability of an element to be zeroed. Default = 0.2
name : str
Base name of the operators
workspace : int
Workspace used in convolution operator
"""
# import pdb
# pdb.set_trace()
if bottle_neck:
# the same as https://github.com/facebook/fb.resnet.torch#notes, a bit difference with origin paper
bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
act1 = mx.sym.Activation(data=bn1, act_type='relu', name=name + '_relu1')
conv1 = mx.sym.Convolution(data=act1, num_filter=int(growth_rate*4), kernel=(1,1), stride=(1,1), pad=(0,0),
no_bias=True, workspace=workspace, name=name + '_conv1')
if drop_out > 0:
conv1 = mx.symbol.Dropout(data=conv1, p=drop_out, name=name + '_dp1')
bn2 = mx.sym.BatchNorm(data=conv1, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn2')
act2 = mx.sym.Activation(data=bn2, act_type='relu', name=name + '_relu2')
conv2 = mx.sym.Convolution(data=act2, num_filter=int(growth_rate), kernel=(3,3), stride=stride, pad=(1,1),
no_bias=True, workspace=workspace, name=name + '_conv2')
if drop_out > 0:
conv2 = mx.symbol.Dropout(data=conv2, p=drop_out, name=name + '_dp2')
#return mx.symbol.Concat(data, conv2, name=name + '_concat0')
return conv2
else:
bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
act1 = mx.sym.Activation(data=bn1, act_type='relu', name=name + '_relu1')
conv1 = mx.sym.Convolution(data=act1, num_filter=int(growth_rate), kernel=(3,3), stride=(1,1), pad=(1,1),
no_bias=True, workspace=workspace, name=name + '_conv1')
if drop_out > 0:
conv1 = mx.symbol.Dropout(data=conv1, p=drop_out, name=name + '_dp1')
#return mx.symbol.Concat(data, conv1, name=name + '_concat0')
return conv1
def DenseBlock(units_num, data, growth_rate, name, bottle_neck=True, drop_out=0.0, bn_mom=0.9, workspace=512):
"""Return DenseBlock Unit symbol for building DenseNet
Parameters
----------
units_num : int
the number of BasicBlock in each DenseBlock
data : str
Input data
growth_rate : int
Number of output channels
drop_out : float
Probability of an element to be zeroed. Default = 0.2
workspace : int
Workspace used in convolution operator
"""
# import pdb
# pdb.set_trace()
for i in range(units_num):
Block = BasicBlock(data, growth_rate=growth_rate, stride=(1,1), name=name + '_unit%d' % (i+1),
bottle_neck=bottle_neck, drop_out=drop_out,
bn_mom=bn_mom, workspace=workspace)
data = mx.symbol.Concat(data, Block, name=name + '_concat%d' % (i+1))
return data
def TransitionBlock(num_stage, data, num_filter, stride, name, drop_out=0.0, bn_mom=0.9, workspace=512):
"""Return TransitionBlock Unit symbol for building DenseNet
Parameters
----------
num_stage : int
Number of stage
data : str
Input data
num : int
Number of output channels
stride : tupe
Stride used in convolution
name : str
Base name of the operators
drop_out : float
Probability of an element to be zeroed. Default = 0.2
workspace : int
Workspace used in convolution operator
"""
bn1 = mx.sym.BatchNorm(data=data, fix_gamma=False, eps=2e-5, momentum=bn_mom, name=name + '_bn1')
act1 = mx.sym.Activation(data=bn1, act_type='relu', name=name + '_relu1')
conv1 = mx.sym.Convolution(data=act1, num_filter=num_filter,
kernel=(1,1), stride=stride, pad=(0,0), no_bias=True,
workspace=workspace, name=name + '_conv1')
if drop_out > 0:
conv1 = mx.symbol.Dropout(data=conv1, p=drop_out, name=name + '_dp1')
return mx.symbol.Pooling(conv1, global_pool=False, kernel=(2,2), stride=(2,2), pool_type='avg', name=name + '_pool%d' % (num_stage+1))
def DenseNet(units, num_stage, growth_rate, num_class, data_type, reduction=0.5, drop_out=0., bottle_neck=True, bn_mom=0.9, workspace=512):
"""Return DenseNet symbol of imagenet
Parameters
----------
units : list
Number of units in each stage
num_stage : int
Number of stage
growth_rate : int
Number of output channels
num_class : int
Ouput size of symbol
data_type : str
the type of dataset
reduction : float
Compression ratio. Default = 0.5
drop_out : float
Probability of an element to be zeroed. Default = 0.2
workspace : int
Workspace used in convolution operator
"""
num_unit = len(units)
assert(num_unit == num_stage)
init_channels = 2 * growth_rate
n_channels = init_channels
data = mx.sym.Variable(name='data')
data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data')
if data_type == 'imagenet':
body = mx.sym.Convolution(data=data, num_filter=growth_rate*2, kernel=(7, 7), stride=(2,2), pad=(3, 3),
no_bias=True, name="conv0", workspace=workspace)
body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
body = mx.sym.Activation(data=body, act_type='relu', name='relu0')
body = mx.symbol.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max')
elif data_type == 'vggface':
body = mx.sym.Convolution(data=data, num_filter=growth_rate*2, kernel=(7, 7), stride=(2,2), pad=(3, 3),
no_bias=True, name="conv0", workspace=workspace)
body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
body = mx.sym.Activation(data=body, act_type='relu', name='relu0')
body = mx.symbol.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max')
elif data_type == 'msface':
body = mx.sym.Convolution(data=data, num_filter=growth_rate*2, kernel=(7, 7), stride=(2,2), pad=(3, 3),
no_bias=True, name="conv0", workspace=workspace)
body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0')
body = mx.sym.Activation(data=body, act_type='relu', name='relu0')
body = mx.symbol.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max')
else:
raise ValueError("do not support {} yet".format(data_type))
for i in range(num_stage-1):
body = DenseBlock(units[i], body, growth_rate=growth_rate, name='DBstage%d' % (i + 1), bottle_neck=bottle_neck, drop_out=drop_out, bn_mom=bn_mom, workspace=workspace)
n_channels += units[i]*growth_rate
n_channels = int(math.floor(n_channels*reduction))
body = TransitionBlock(i, body, n_channels, stride=(1,1), name='TBstage%d' % (i + 1), drop_out=drop_out, bn_mom=bn_mom, workspace=workspace)
body = DenseBlock(units[num_stage-1], body, growth_rate=growth_rate, name='DBstage%d' % (num_stage), bottle_neck=bottle_neck, drop_out=drop_out, bn_mom=bn_mom, workspace=workspace)
bn1 = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1')
relu1 = mx.sym.Activation(data=bn1, act_type='relu', name='relu1')
pool1 = mx.symbol.Pooling(data=relu1, global_pool=True, kernel=(7, 7), pool_type='avg', name='pool1')
flat = mx.symbol.Flatten(data=pool1)
fc1 = mx.symbol.FullyConnected(data=flat, num_hidden=num_class, name='fc1')
return mx.symbol.SoftmaxOutput(data=fc1, name='softmax')