質問編集履歴
1
修正
test
CHANGED
|
File without changes
|
test
CHANGED
|
@@ -1,283 +1,5 @@
|
|
|
1
|
-
|
|
1
|
+
長方形の画像について画像認識を行いたいと考えています。
|
|
2
2
|
|
|
3
3
|
様々なアスペクト比の画像が混在しているため、初めは256×256の画像にリサイズした後に、学習することを行っていました。しかし、精度があまりよくなかったため、元の画像サイズに近い、200×400に画像サイズを統一した後学習を行おうと考えています。しかし、エラーが出てしまいその原因を究明することができません。
|
|
4
4
|
|
|
5
5
|
ネットワーク構造は基本256×256の画像の学習で使用していたものと同じです。
|
|
6
|
-
|
|
7
|
-
以下のプログラムを参考にしています。
|
|
8
|
-
|
|
9
|
-
https://github.com/rkuga/PSPNet/blob/master/nets/pspnet.py
|
|
10
|
-
|
|
11
|
-
```pspnet.py
|
|
12
|
-
|
|
13
|
-
import numpy as np
|
|
14
|
-
|
|
15
|
-
import cv2
|
|
16
|
-
|
|
17
|
-
import chainer
|
|
18
|
-
|
|
19
|
-
import chainer.functions as F
|
|
20
|
-
|
|
21
|
-
import chainer.links as L
|
|
22
|
-
|
|
23
|
-
|
|
24
|
-
|
|
25
|
-
class PSPNet(chainer.Chain):
|
|
26
|
-
|
|
27
|
-
# in_channel=1, out_channel=2, input_height=400, input_width =200
|
|
28
|
-
|
|
29
|
-
def __init__(self, in_channel, out_channel, input_height, input_width):
|
|
30
|
-
|
|
31
|
-
self.input_height = input_height
|
|
32
|
-
|
|
33
|
-
self.input_width = input_width
|
|
34
|
-
|
|
35
|
-
super(PSPNet, self).__init__(
|
|
36
|
-
|
|
37
|
-
conv1_1=L.Convolution2D(in_channel, 64, 3, stride=2, pad=1, nobias=True),
|
|
38
|
-
|
|
39
|
-
bn1_1=L.BatchNormalization(64),
|
|
40
|
-
|
|
41
|
-
conv1_2=L.Convolution2D(64, 64, 3, stride=1, pad=1, nobias=True),
|
|
42
|
-
|
|
43
|
-
bn1_2=L.BatchNormalization(64),
|
|
44
|
-
|
|
45
|
-
conv1_3=L.Convolution2D(64, 128, 3, stride=1, pad=1, nobias=True),
|
|
46
|
-
|
|
47
|
-
bn1_3=L.BatchNormalization(128),
|
|
48
|
-
|
|
49
|
-
|
|
50
|
-
|
|
51
|
-
conv2_1_proj = L.Convolution2D(128, 256, 1, stride=1, pad=0, nobias=True),
|
|
52
|
-
|
|
53
|
-
bn2_1_proj=L.BatchNormalization(256),
|
|
54
|
-
|
|
55
|
-
conv2_1_res = BottleNeck_proj(128, 64, 256),
|
|
56
|
-
|
|
57
|
-
conv2_res = Block(2, 256, 64, 256),
|
|
58
|
-
|
|
59
|
-
|
|
60
|
-
|
|
61
|
-
conv3_1_proj = L.Convolution2D(256, 512, 1, stride=1, pad=0, nobias=True),
|
|
62
|
-
|
|
63
|
-
bn3_1_proj=L.BatchNormalization(512),
|
|
64
|
-
|
|
65
|
-
conv3_1_res = BottleNeck_proj(256, 128, 512),
|
|
66
|
-
|
|
67
|
-
conv3_res = Block(3, 512, 128, 512),
|
|
68
|
-
|
|
69
|
-
|
|
70
|
-
|
|
71
|
-
conv4_1_proj = L.Convolution2D(512, 1024, 1, stride=1, pad=0, nobias=True),
|
|
72
|
-
|
|
73
|
-
bn4_1_proj=L.BatchNormalization(1024),
|
|
74
|
-
|
|
75
|
-
conv4_1_res = BottleNeck_proj(512, 256, 1024),
|
|
76
|
-
|
|
77
|
-
conv4_res = Block(7, 1024, 256, 1024, Dilate=2),
|
|
78
|
-
|
|
79
|
-
conv4_res2 = Block(15, 1024, 256, 1024, Dilate=2),
|
|
80
|
-
|
|
81
|
-
|
|
82
|
-
|
|
83
|
-
conv5_1_proj = L.Convolution2D(1024, 2048, 1, stride=1, pad=0, nobias=True),
|
|
84
|
-
|
|
85
|
-
bn5_1_proj=L.BatchNormalization(2048),
|
|
86
|
-
|
|
87
|
-
conv5_1_res = BottleNeck_proj(1024, 512, 2048),
|
|
88
|
-
|
|
89
|
-
conv5_res = Block(2, 2048, 512, 2048, Dilate=4),
|
|
90
|
-
|
|
91
|
-
|
|
92
|
-
|
|
93
|
-
conv5_3_pool1_conv = L.Convolution2D(2048, 512, 1, stride=1, pad=0, nobias=True),
|
|
94
|
-
|
|
95
|
-
bn5_3_pool1_conv=L.BatchNormalization(512),
|
|
96
|
-
|
|
97
|
-
conv5_3_pool2_conv = L.Convolution2D(2048, 512, 1, stride=1, pad=0, nobias=True),
|
|
98
|
-
|
|
99
|
-
bn5_3_pool2_conv=L.BatchNormalization(512),
|
|
100
|
-
|
|
101
|
-
conv5_3_pool3_conv = L.Convolution2D(2048, 512, 1, stride=1, pad=0, nobias=True),
|
|
102
|
-
|
|
103
|
-
bn5_3_pool3_conv=L.BatchNormalization(512),
|
|
104
|
-
|
|
105
|
-
conv5_3_pool6_conv = L.Convolution2D(2048, 512, 1, stride=1, pad=0, nobias=True),
|
|
106
|
-
|
|
107
|
-
bn5_3_pool6_conv=L.BatchNormalization(512),
|
|
108
|
-
|
|
109
|
-
|
|
110
|
-
|
|
111
|
-
conv5_4 = L.Convolution2D(4096, 512, 3, stride=1, pad=1, nobias=True),
|
|
112
|
-
|
|
113
|
-
bn5_4=L.BatchNormalization(512),
|
|
114
|
-
|
|
115
|
-
|
|
116
|
-
|
|
117
|
-
conv6 = L.Convolution2D(512, out_channel, 1, stride=1, pad=0, nobias=True),
|
|
118
|
-
|
|
119
|
-
|
|
120
|
-
|
|
121
|
-
|
|
122
|
-
|
|
123
|
-
)
|
|
124
|
-
|
|
125
|
-
|
|
126
|
-
|
|
127
|
-
|
|
128
|
-
|
|
129
|
-
def __call__(self, x):
|
|
130
|
-
|
|
131
|
-
conv1_1 = F.relu(self.bn1_1(self.conv1_1(x)))
|
|
132
|
-
|
|
133
|
-
conv1_2 = F.relu(self.bn1_2(self.conv1_2(conv1_1)))
|
|
134
|
-
|
|
135
|
-
conv1_3 = F.relu(self.bn1_3(self.conv1_3(conv1_2)))
|
|
136
|
-
|
|
137
|
-
pool1 = F.max_pooling_2d(conv1_3,3,stride=2)
|
|
138
|
-
|
|
139
|
-
|
|
140
|
-
|
|
141
|
-
conv2_1_proj = self.bn2_1_proj(self.conv2_1_proj(pool1))
|
|
142
|
-
|
|
143
|
-
conv2_1 = self.conv2_1_res(pool1, conv2_1_proj)
|
|
144
|
-
|
|
145
|
-
conv2_3 = self.conv2_res(conv2_1)
|
|
146
|
-
|
|
147
|
-
|
|
148
|
-
|
|
149
|
-
conv3_1_proj = self.bn3_1_proj(self.conv3_1_proj(conv2_3))
|
|
150
|
-
|
|
151
|
-
conv3_1 = self.conv3_1_res(conv2_3, conv3_1_proj)
|
|
152
|
-
|
|
153
|
-
conv3_4 = self.conv3_res(conv3_1)
|
|
154
|
-
|
|
155
|
-
|
|
156
|
-
|
|
157
|
-
conv4_1_proj = self.bn4_1_proj(self.conv4_1_proj(conv3_4))
|
|
158
|
-
|
|
159
|
-
conv4_1 = self.conv4_1_res(conv3_4, conv4_1_proj)
|
|
160
|
-
|
|
161
|
-
conv4_23_1 = self.conv4_res(conv4_1)
|
|
162
|
-
|
|
163
|
-
conv4_23 = self.conv4_res2(conv4_23_1)
|
|
164
|
-
|
|
165
|
-
|
|
166
|
-
|
|
167
|
-
conv5_1_proj = self.bn5_1_proj(self.conv5_1_proj(conv4_23))
|
|
168
|
-
|
|
169
|
-
conv5_1 = self.conv5_1_res(conv4_23, conv5_1_proj)
|
|
170
|
-
|
|
171
|
-
conv5_3 = self.conv5_res(conv5_1)
|
|
172
|
-
|
|
173
|
-
|
|
174
|
-
|
|
175
|
-
_,c,h,w = conv5_3.data.shape
|
|
176
|
-
|
|
177
|
-
conv5_3_pool1 = F.average_pooling_2d(conv5_3, h, stride=h)
|
|
178
|
-
|
|
179
|
-
print(conv5_3_pool1.data.shape)
|
|
180
|
-
|
|
181
|
-
conv5_3_pool1_conv = self.bn5_3_pool1_conv(self.conv5_3_pool1_conv(conv5_3_pool1))
|
|
182
|
-
|
|
183
|
-
conv5_3_pool1_interp = F.resize_images(conv5_3_pool1_conv, (h,w))
|
|
184
|
-
|
|
185
|
-
|
|
186
|
-
|
|
187
|
-
conv5_3_pool2 = F.average_pooling_2d(conv5_3, int(h/2), stride=int(h/2))
|
|
188
|
-
|
|
189
|
-
conv5_3_pool2_conv = self.bn5_3_pool2_conv(self.conv5_3_pool2_conv(conv5_3_pool2))
|
|
190
|
-
|
|
191
|
-
conv5_3_pool2_interp = F.resize_images(conv5_3_pool2_conv, (h,w))
|
|
192
|
-
|
|
193
|
-
|
|
194
|
-
|
|
195
|
-
conv5_3_pool3 = F.average_pooling_2d(conv5_3, int(h/3), stride=int(h/3))
|
|
196
|
-
|
|
197
|
-
conv5_3_pool3_conv = self.bn5_3_pool3_conv(self.conv5_3_pool3_conv(conv5_3_pool3))
|
|
198
|
-
|
|
199
|
-
conv5_3_pool3_interp = F.resize_images(conv5_3_pool3_conv, (h,w))
|
|
200
|
-
|
|
201
|
-
|
|
202
|
-
|
|
203
|
-
conv5_3_pool6 = F.average_pooling_2d(conv5_3, int(h/6), stride=int(h/6))
|
|
204
|
-
|
|
205
|
-
conv5_3_pool6_conv = self.bn5_3_pool6_conv(self.conv5_3_pool6_conv(conv5_3_pool6))
|
|
206
|
-
|
|
207
|
-
conv5_3_pool6_interp = F.resize_images(conv5_3_pool6_conv, (h,w))
|
|
208
|
-
|
|
209
|
-
|
|
210
|
-
|
|
211
|
-
conv5_3_concat = F.concat((conv5_3, conv5_3_pool6_interp))
|
|
212
|
-
|
|
213
|
-
conv5_3_concat = F.concat((conv5_3_concat, conv5_3_pool3_interp))
|
|
214
|
-
|
|
215
|
-
conv5_3_concat = F.concat((conv5_3_concat, conv5_3_pool2_interp))
|
|
216
|
-
|
|
217
|
-
conv5_3_concat = F.concat((conv5_3_concat, conv5_3_pool1_interp))
|
|
218
|
-
|
|
219
|
-
|
|
220
|
-
|
|
221
|
-
conv5_4 = F.relu(self.bn5_4(self.conv5_4(conv5_3_concat)))
|
|
222
|
-
|
|
223
|
-
conv5_4 = F.dropout(conv5_4, ratio=.1)
|
|
224
|
-
|
|
225
|
-
|
|
226
|
-
|
|
227
|
-
conv6 = self.conv6(conv5_4)
|
|
228
|
-
|
|
229
|
-
|
|
230
|
-
|
|
231
|
-
conv6_interp = F.resize_images(conv6, (self.input_height,self.input_width))
|
|
232
|
-
|
|
233
|
-
|
|
234
|
-
|
|
235
|
-
return conv6_interp
|
|
236
|
-
|
|
237
|
-
```
|
|
238
|
-
|
|
239
|
-
以下エラー内容です。
|
|
240
|
-
|
|
241
|
-
`assert y_w > 0, 'Width in the output should be positive.'`から、average_pooling_2dにおいて幅が負の値になっているということのようですが、その理由が分かりません。どこを改善すればよいのでしょうか？わかる方がいましたら教えていただけると幸いです。よろしくお願いいたします。
|
|
242
|
-
|
|
243
|
-
|
|
244
|
-
|
|
245
|
-
```
|
|
246
|
-
|
|
247
|
-
C:\Users\hoge\Anaconda3\lib\site-packages\chainer\functions\normalization\batch_normalization.py:67: UserWarning: A batch with no more than one sample has been given to F.batch_normalization. F.batch_normalization will always output a zero tensor for such batches. This could be caused by incorrect configuration in your code (such as running evaluation while chainer.config.train=True), but could also happen in the last batch of training if non-repeating iterator is used.
|
|
248
|
-
|
|
249
|
-
UserWarning)
|
|
250
|
-
|
|
251
|
-
Traceback (most recent call last):
|
|
252
|
-
|
|
253
|
-
File "PSPtrain.py", line 63, in main
|
|
254
|
-
|
|
255
|
-
z = net(x)
|
|
256
|
-
|
|
257
|
-
File "C:\Users\hoge\pspnet.py", line 166, in __call__
|
|
258
|
-
|
|
259
|
-
conv5_3_pool2 = F.average_pooling_2d(conv5_3, int(h/2), stride=int(h/2))
|
|
260
|
-
|
|
261
|
-
File "C:\Users\hoge\Anaconda3\lib\site-packages\chainer\functions\pooling\average_pooling_2d.py", line 214, in average_pooling_2d
|
|
262
|
-
|
|
263
|
-
return AveragePooling2D(ksize, stride, pad, False).apply((x,))[0]
|
|
264
|
-
|
|
265
|
-
File "C:\Users\hoge\Anaconda3\lib\site-packages\chainer\function_node.py", line 258, in apply
|
|
266
|
-
|
|
267
|
-
outputs = self.forward(in_data)
|
|
268
|
-
|
|
269
|
-
File "C:\Users\hoge\Anaconda3\lib\site-packages\chainer\function_node.py", line 367, in forward
|
|
270
|
-
|
|
271
|
-
return self.forward_gpu(inputs)
|
|
272
|
-
|
|
273
|
-
File "C:\Users\hoge\Anaconda3\lib\site-packages\chainer\functions\pooling\average_pooling_2d.py", line 56, in forward_gpu
|
|
274
|
-
|
|
275
|
-
return super(AveragePooling2D, self).forward_gpu(x)
|
|
276
|
-
|
|
277
|
-
File "C:\Users\hoge\Anaconda3\lib\site-packages\chainer\functions\pooling\pooling_2d.py", line 55, in forward_gpu
|
|
278
|
-
|
|
279
|
-
assert y_w > 0, 'Width in the output should be positive.'
|
|
280
|
-
|
|
281
|
-
AssertionError: Width in the output should be positive.
|
|
282
|
-
|
|
283
|
-
```
|