-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcnn_model.py
189 lines (176 loc) · 9.01 KB
/
cnn_model.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
#!/usr/bin/env python
# -*- coding: utf-8 -*-
import chainer
import chainer.functions as F
import chainer.links as L
from chainer.functions.loss import softmax_cross_entropy
from chainer.functions.evaluation import accuracy
from chainer import reporter
# CONV → Batch → ReLU を1つのノードとして返すクラス
class ConvBlock(chainer.Chain):
def __init__(self, ksize, initializer, n_out=32):
super(ConvBlock, self).__init__()
links = []
if ksize == 3:
links = [('conv1', L.Convolution2D(None, n_out, ksize, pad=1, initialW=initializer))]
links += [('bn1', L.BatchNormalization(n_out))]
elif ksize == 5:
links = [('conv1', L.Convolution2D(None, n_out, ksize, pad=2, initialW=initializer))]
links += [('bn1', L.BatchNormalization(n_out))]
else:
links = [('conv1', L.Convolution2D(None, n_out, 7, pad=3, initialW=initializer))]
links += [('bn1', L.BatchNormalization(n_out))]
for link in links:
self.add_link(*link)
self.forward = links
def __call__(self, x, train):
for name, f in self.forward:
if 'conv1' in name:
x = getattr(self, name)(x)
elif 'bn1' in name:
x = getattr(self, name)(x, not train)
return F.relu(x)
# Batch → ReLU を1つのノードとして返すクラス（単独でReLUノードを使わない場合，不必要）
class ActivationBlock(chainer.Chain):
def __init__(self):
super(ActivationBlock, self).__init__()
links = []
links = [('bn1', L.BatchNormalization(None))]
links += [('_act1', F.ReLU())]
for link in links:
if not link[0].startswith('_'):
self.add_link(*link)
self.forward = links
def __call__(self, x, train):
for name, f in self.forward:
if 'bn1' in name:
x = getattr(self, name)(x, not train)
elif name.startswith('_'):
x = f(x)
return x
# CGP(list)からCNNのモデル構築
# n_in: 入力画像のchannel
# n_out: 畳込み層の出力channel数
# Classifierで定義想定（ model = Classifier(CGP2CNN(cgp)) )
class CGP2CNN(chainer.Chain):
def __init__(self, cgp, n_class, n_in=3, n_out=32, lossfun=softmax_cross_entropy.softmax_cross_entropy, accfun=accuracy.accuracy):
super(CGP2CNN, self).__init__()
self.cgp = cgp
self.pool_size = 2
self.n_out= n_out
initializer = chainer.initializers.HeNormal()
links = []
i = 1
for name, in1, in2 in self.cgp:
if name == 'conv3':
links += [(name+'_'+str(i), L.Convolution2D(None, n_out, 3, pad=1, initialW=initializer))]
elif name == 'conv5':
links += [(name+'_'+str(i), L.Convolution2D(None, n_out, 5, pad=2, initialW=initializer))]
elif name == 'conv7':
links += [(name+'_'+str(i), L.Convolution2D(None, n_out, 7, pad=3, initialW=initializer))]
elif name == 'pool_max':
links += [('_'+name+'_'+str(i), F.MaxPooling2D(self.pool_size, self.pool_size, 0, False))]
elif name == 'pool_max_const':
links += [('_'+name+'_'+str(i), F.MaxPooling2D(3, 1, 1, True))]
elif name == 'pool_ave':
links += [('_'+name+'_'+str(i), F.AveragePooling2D(self.pool_size, self.pool_size, 0, False))]
elif name == 'pool_ave_const':
links += [('_'+name+'_'+str(i), F.AveragePooling2D(3, 1, 1, True))]
elif name == 'ReLU':
links += [('ActivationBlock'+'_'+str(i), ActivationBlock())] # ◆BatchNormalization(None)にしちゃうとBatchNormalizationのパラメータを学習してくれない?（精度が悪くなる）
# links += [('batch_'+str(i), L.BatchNormalization(None))]
# links += [('_'+name+'_'+str(i), F.ReLU())]
elif name == 'tanh':
links += [('batch_'+str(i), L.BatchNormalization(None))]
links += [('_'+name+'_'+str(i), F.Tanh())]
elif name == 'concat':
links += [('_'+name+'_'+str(i), F.Concat())]
elif name == 'sum':
links += [('_'+name+'_'+str(i), F.Concat())]
elif name == 'ConvBlock3':
links += [(name+'_'+str(i), ConvBlock(3, initializer))]
elif name == 'ConvBlock5':
links += [(name+'_'+str(i), ConvBlock(5, initializer))]
elif name == 'ConvBlock7':
links += [(name+'_'+str(i), ConvBlock(7, initializer))]
elif name == 'full':
links += [(name+'_'+str(i), L.Linear(None, n_class, initialW=initializer))]
i += 1
for link in links:
if not link[0].startswith('_'):
self.add_link(*link)
self.forward = links
self.train = True
self.lossfun = lossfun
self.accfun = accfun
self.loss = None
self.accuracy = None
self.outputs = [None for _ in range(len(self.cgp))]
self.param_num = 0
def __call__(self, x, t):
xp = chainer.cuda.get_array_module(x)
outputs = self.outputs
outputs[0] = x # 原画像
nodeID = 1
param_num = 0
for name, f in self.forward:
if 'conv' in name:
outputs[nodeID] = getattr(self, name)(outputs[self.cgp[nodeID][1]])
nodeID += 1
param_num += (f.W.shape[0]*f.W.shape[2]*f.W.shape[3]*f.W.shape[1]+f.W.shape[0])
elif 'ConvBlock' in name:
outputs[nodeID] = getattr(self, name)(outputs[self.cgp[nodeID][1]], self.train)
nodeID += 1
elif 'ActivationBlock' in name:
outputs[nodeID] = getattr(self, name)(outputs[self.cgp[nodeID][1]], self.train)
nodeID += 1
# elif 'batch' in name:
# outputs[nodeID] = getattr(self, name)(outputs[self.cgp[nodeID][1]], not self.train)
# param_num += outputs[nodeID].data.shape[1]*2
# elif 'ReLU' in name or 'tanh' in name:
# outputs[nodeID] = f(outputs[nodeID])
# nodeID += 1
elif name.startswith('_') and 'concat' not in name and 'sum' not in name:
outputs[nodeID] = f(outputs[self.cgp[nodeID][1]])
nodeID += 1
elif 'concat' in name:
in_data = [outputs[self.cgp[nodeID][1]], outputs[self.cgp[nodeID][2]]]
small_in_id, large_in_id = (0, 1) if in_data[0].shape[2] < in_data[1].shape[2] else (1, 0)
pool_num = xp.floor(xp.log2(in_data[large_in_id].shape[2] / in_data[small_in_id].shape[2]))
for _ in xp.arange(pool_num):
in_data[large_in_id] = F.max_pooling_2d(in_data[large_in_id], self.pool_size, self.pool_size, 0, False)
outputs[nodeID] = f(in_data[0], in_data[1])
nodeID += 1
elif 'sum' in name:
in_data = [outputs[self.cgp[nodeID][1]], outputs[self.cgp[nodeID][2]]]
# 画像サイズに関するチェック
small_in_id, large_in_id = (0, 1) if in_data[0].shape[2] < in_data[1].shape[2] else (1, 0)
pool_num = xp.floor(xp.log2(in_data[large_in_id].shape[2] / in_data[small_in_id].shape[2]))
for _ in xp.arange(pool_num):
in_data[large_in_id] = F.max_pooling_2d(in_data[large_in_id], self.pool_size, self.pool_size, 0, False)
# channel sizeに関するチェック
small_ch_id, large_ch_id = (0, 1) if in_data[0].shape[1] < in_data[1].shape[1] else (1, 0)
pad_num = int(in_data[large_ch_id].shape[1] - in_data[small_ch_id].shape[1])
tmp = in_data[large_ch_id][:, :pad_num, :, :]
in_data[small_ch_id] = F.concat((in_data[small_ch_id], tmp * 0), axis=1)
outputs[nodeID] = in_data[0] + in_data[1]
nodeID += 1
else:
outputs[nodeID] = getattr(self, name)(outputs[self.cgp[nodeID][1]])
nodeID += 1
param_num += f.W.data.shape[0] * f.W.data.shape[1] + f.b.data.shape[0]
self.param_num = param_num
if self.train:
# self.loss = F.softmax_cross_entropy(outputs[-1], t)
# self.accuracy = F.accuracy(outputs[-1], t)
# return self.loss
self.loss = None
self.accuracy = None
self.loss = self.lossfun(outputs[-1], t)
reporter.report({'loss': self.loss}, self)
self.accuracy = self.accfun(outputs[-1], t)
reporter.report({'accuracy': self.accuracy}, self)
return self.loss
# return outputs[-1]
else:
return outputs[-1]