質問編集履歴
3
最終的なコード
test
CHANGED
File without changes
|
test
CHANGED
@@ -6,166 +6,658 @@
|
|
6
6
|
|
7
7
|
|
8
8
|
|
9
|
+
|
10
|
+
|
9
11
|
### 発生している問題・エラーメッセージ
|
10
12
|
|
11
13
|
|
12
14
|
|
13
15
|
```
|
14
16
|
|
15
|
-
|
16
|
-
|
17
|
-
|
18
|
-
|
19
|
-
I tensorflow/core/common_runtime/bfc_allocator.cc:702] Stats:
|
20
|
-
|
21
|
-
Limit: 7582672487
|
22
|
-
|
23
|
-
InUse: 6971735552
|
24
|
-
|
25
|
-
MaxInUse: 7541746432
|
26
|
-
|
27
|
-
NumAllocs: 443
|
28
|
-
|
29
|
-
MaxAllocSize: 3043936256
|
30
|
-
|
31
|
-
|
32
|
-
|
33
|
-
W tensorflow/core/common_runtime/bfc_allocator.cc:274] **************************_**___****************************************************************xxxx
|
34
|
-
|
35
|
-
W tensorflow/core/common_runtime/bfc_allocator.cc:275] Ran out of memory trying to allocate 392.00MiB. See logs for memory state.
|
36
|
-
|
37
|
-
W tensorflow/core/framework/op_kernel.cc:993] Resource exhausted: OOM when allocating tensor with shape[100352,1024]
|
38
|
-
|
39
|
-
Traceback (most recent call last):
|
40
|
-
|
41
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1022, in _do_call
|
42
|
-
|
43
|
-
return fn(*args)
|
44
|
-
|
45
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1004, in _run_fn
|
46
|
-
|
47
|
-
status, run_metadata)
|
48
|
-
|
49
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/contextlib.py", line 88, in __exit__
|
50
|
-
|
51
|
-
next(self.gen)
|
52
|
-
|
53
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/framework/errors_impl.py", line 469, in raise_exception_on_not_ok_status
|
54
|
-
|
55
|
-
pywrap_tensorflow.TF_GetCode(status))
|
56
|
-
|
57
|
-
|
17
|
+
Error when checking model input: expected convolution2d_input_1 to have shape (None, 3, 32, 32) but got array with shape (500, 32, 32, 3)
|
58
|
-
|
59
|
-
[[Node: sub_28 = Sub[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/gpu:0"](add_14, mul_60)]]
|
60
|
-
|
61
|
-
|
62
|
-
|
63
|
-
During handling of the above exception, another exception occurred:
|
64
|
-
|
65
|
-
|
66
|
-
|
67
|
-
Traceback (most recent call last):
|
68
|
-
|
69
|
-
File "n.py", line 299, in <module>
|
70
|
-
|
71
|
-
run_train('9_Layer_CNN')
|
72
|
-
|
73
|
-
File "n.py", line 208, in run_train
|
74
|
-
|
75
|
-
callbacks=[cp])
|
76
|
-
|
77
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/models.py", line 870, in fit
|
78
|
-
|
79
|
-
initial_epoch=initial_epoch)
|
80
|
-
|
81
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/engine/training.py", line 1507, in fit
|
82
|
-
|
83
|
-
initial_epoch=initial_epoch)
|
84
|
-
|
85
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/engine/training.py", line 1156, in _fit_loop
|
86
|
-
|
87
|
-
outs = f(ins_batch)
|
88
|
-
|
89
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 2269, in __call__
|
90
|
-
|
91
|
-
**self.session_kwargs)
|
92
|
-
|
93
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 767, in run
|
94
|
-
|
95
|
-
run_metadata_ptr)
|
96
|
-
|
97
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 965, in _run
|
98
|
-
|
99
|
-
feed_dict_string, options, run_metadata)
|
100
|
-
|
101
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1015, in _do_run
|
102
|
-
|
103
|
-
target_list, options, run_metadata)
|
104
|
-
|
105
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1035, in _do_call
|
106
|
-
|
107
|
-
raise type(e)(node_def, op, message)
|
108
|
-
|
109
|
-
tensorflow.python.framework.errors_impl.ResourceExhaustedError: OOM when allocating tensor with shape[100352,1024]
|
110
|
-
|
111
|
-
[[Node: sub_28 = Sub[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/gpu:0"](add_14, mul_60)]]
|
112
|
-
|
113
|
-
|
114
|
-
|
115
|
-
Caused by op 'sub_28', defined at:
|
116
|
-
|
117
|
-
File "n.py", line 299, in <module>
|
118
|
-
|
119
|
-
run_train('9_Layer_CNN')
|
120
|
-
|
121
|
-
File "n.py", line 208, in run_train
|
122
|
-
|
123
|
-
callbacks=[cp])
|
124
|
-
|
125
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/models.py", line 870, in fit
|
126
|
-
|
127
|
-
initial_epoch=initial_epoch)
|
128
|
-
|
129
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/engine/training.py", line 1490, in fit
|
130
|
-
|
131
|
-
self._make_train_function()
|
132
|
-
|
133
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/engine/training.py", line 1014, in _make_train_function
|
134
|
-
|
135
|
-
self.total_loss)
|
136
|
-
|
137
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/optimizers.py", line 169, in get_updates
|
138
|
-
|
139
|
-
new_p = p + self.momentum * v - lr * g
|
140
|
-
|
141
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/ops/math_ops.py", line 884, in binary_op_wrapper
|
142
|
-
|
143
|
-
return func(x, y, name=name)
|
144
|
-
|
145
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/ops/gen_math_ops.py", line 2775, in _sub
|
146
|
-
|
147
|
-
result = _op_def_lib.apply_op("Sub", x=x, y=y, name=name)
|
148
|
-
|
149
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py", line 763, in apply_op
|
150
|
-
|
151
|
-
op_def=op_def)
|
152
|
-
|
153
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 2395, in create_op
|
154
|
-
|
155
|
-
original_op=self._default_original_op, op_def=op_def)
|
156
|
-
|
157
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 1264, in __init__
|
158
|
-
|
159
|
-
self._traceback = _extract_stack()
|
160
|
-
|
161
|
-
|
162
|
-
|
163
|
-
ResourceExhaustedError (see above for traceback): OOM when allocating tensor with shape[100352,1024]
|
164
|
-
|
165
|
-
[[Node: sub_28 = Sub[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/gpu:0"](add_14, mul_60)]]
|
166
|
-
|
167
|
-
|
168
|
-
|
169
|
-
|
170
18
|
|
171
19
|
```
|
20
|
+
|
21
|
+
|
22
|
+
|
23
|
+
### 該当のソースコード
|
24
|
+
|
25
|
+
|
26
|
+
|
27
|
+
```
|
28
|
+
|
29
|
+
import numpy as np
|
30
|
+
|
31
|
+
from numpy.random import permutation
|
32
|
+
|
33
|
+
|
34
|
+
|
35
|
+
import os, glob, cv2, math, sys
|
36
|
+
|
37
|
+
import pandas as pd
|
38
|
+
|
39
|
+
|
40
|
+
|
41
|
+
from keras.models import Sequential, model_from_json
|
42
|
+
|
43
|
+
from keras.layers.core import Dense, Dropout, Flatten
|
44
|
+
|
45
|
+
from keras.layers.convolutional import Convolution2D, MaxPooling2D
|
46
|
+
|
47
|
+
from keras.layers.advanced_activations import LeakyReLU
|
48
|
+
|
49
|
+
from keras.callbacks import ModelCheckpoint
|
50
|
+
|
51
|
+
from keras.optimizers import SGD
|
52
|
+
|
53
|
+
from keras.utils import np_utils
|
54
|
+
|
55
|
+
|
56
|
+
|
57
|
+
# seed値
|
58
|
+
|
59
|
+
np.random.seed(1)
|
60
|
+
|
61
|
+
|
62
|
+
|
63
|
+
# 使用する画像サイズ
|
64
|
+
|
65
|
+
img_rows, img_cols = 224, 224
|
66
|
+
|
67
|
+
|
68
|
+
|
69
|
+
# 画像データ 1枚の読み込みとリサイズを行う
|
70
|
+
|
71
|
+
def get_im(path):
|
72
|
+
|
73
|
+
|
74
|
+
|
75
|
+
img = cv2.imread(path)
|
76
|
+
|
77
|
+
resized = cv2.resize(img, (img_cols, img_rows))
|
78
|
+
|
79
|
+
|
80
|
+
|
81
|
+
return resized
|
82
|
+
|
83
|
+
|
84
|
+
|
85
|
+
|
86
|
+
|
87
|
+
# データの読み込み、正規化、シャッフルを行う
|
88
|
+
|
89
|
+
def read_train_data(ho=0, kind='train'):
|
90
|
+
|
91
|
+
|
92
|
+
|
93
|
+
train_data = []
|
94
|
+
|
95
|
+
train_target = []
|
96
|
+
|
97
|
+
|
98
|
+
|
99
|
+
# 学習用データ読み込み
|
100
|
+
|
101
|
+
for j in range(0, 6): # 0~5まで
|
102
|
+
|
103
|
+
|
104
|
+
|
105
|
+
path = '../../data/Caltech-101/'
|
106
|
+
|
107
|
+
path += '%s/%i/*/%i/*.jpg'%(kind, ho, j)
|
108
|
+
|
109
|
+
|
110
|
+
|
111
|
+
files = sorted(glob.glob(path))
|
112
|
+
|
113
|
+
|
114
|
+
|
115
|
+
for fl in files:
|
116
|
+
|
117
|
+
|
118
|
+
|
119
|
+
flbase = os.path.basename(fl)
|
120
|
+
|
121
|
+
|
122
|
+
|
123
|
+
# 画像 1枚 読み込み
|
124
|
+
|
125
|
+
img = get_im(fl)
|
126
|
+
|
127
|
+
img = np.array(img, dtype=np.float32)
|
128
|
+
|
129
|
+
|
130
|
+
|
131
|
+
# 正規化(GCN)実行
|
132
|
+
|
133
|
+
img -= np.mean(img)
|
134
|
+
|
135
|
+
img /= np.std(img)
|
136
|
+
|
137
|
+
|
138
|
+
|
139
|
+
train_data.append(img)
|
140
|
+
|
141
|
+
train_target.append(j)
|
142
|
+
|
143
|
+
|
144
|
+
|
145
|
+
# 読み込んだデータを numpy の array に変換
|
146
|
+
|
147
|
+
train_data = np.array(train_data, dtype=np.float32)
|
148
|
+
|
149
|
+
train_target = np.array(train_target, dtype=np.uint8)
|
150
|
+
|
151
|
+
|
152
|
+
|
153
|
+
# (レコード数,縦,横,channel数) を (レコード数,channel数,縦,横) に変換
|
154
|
+
|
155
|
+
#train_data = train_data.transpose((0, 3, 1, 2))
|
156
|
+
|
157
|
+
|
158
|
+
|
159
|
+
# target を 6次元のデータに変換。
|
160
|
+
|
161
|
+
# ex) 1 -> 0,1,0,0,0,0 2 -> 0,0,1,0,0,0
|
162
|
+
|
163
|
+
train_target = np_utils.to_categorical(train_target, 6)
|
164
|
+
|
165
|
+
|
166
|
+
|
167
|
+
# データをシャッフル
|
168
|
+
|
169
|
+
perm = permutation(len(train_target))
|
170
|
+
|
171
|
+
train_data = train_data[perm]
|
172
|
+
|
173
|
+
train_target = train_target[perm]
|
174
|
+
|
175
|
+
|
176
|
+
|
177
|
+
return train_data, train_target
|
178
|
+
|
179
|
+
|
180
|
+
|
181
|
+
|
182
|
+
|
183
|
+
# テストデータ読み込み
|
184
|
+
|
185
|
+
def load_test(test_class, aug_i):
|
186
|
+
|
187
|
+
|
188
|
+
|
189
|
+
path = '../../data/Caltech-101/test/%i/%i/*.jpg'%(aug_i, test_class)
|
190
|
+
|
191
|
+
|
192
|
+
|
193
|
+
files = sorted(glob.glob(path))
|
194
|
+
|
195
|
+
X_test = []
|
196
|
+
|
197
|
+
X_test_id = []
|
198
|
+
|
199
|
+
|
200
|
+
|
201
|
+
for fl in files:
|
202
|
+
|
203
|
+
flbase = os.path.basename(fl)
|
204
|
+
|
205
|
+
|
206
|
+
|
207
|
+
img = get_im(fl)
|
208
|
+
|
209
|
+
img = np.array(img, dtype=np.float32)
|
210
|
+
|
211
|
+
|
212
|
+
|
213
|
+
# 正規化(GCN)実行
|
214
|
+
|
215
|
+
img -= np.mean(img)
|
216
|
+
|
217
|
+
img /= np.std(img)
|
218
|
+
|
219
|
+
|
220
|
+
|
221
|
+
X_test.append(img)
|
222
|
+
|
223
|
+
X_test_id.append(flbase)
|
224
|
+
|
225
|
+
|
226
|
+
|
227
|
+
# 読み込んだデータを numpy の array に変換
|
228
|
+
|
229
|
+
test_data = np.array(X_test, dtype=np.float32)
|
230
|
+
|
231
|
+
|
232
|
+
|
233
|
+
# (レコード数,縦,横,channel数) を (レコード数,channel数,縦,横) に変換
|
234
|
+
|
235
|
+
#test_data = test_data.transpose((0, 3, 1, 2))
|
236
|
+
|
237
|
+
|
238
|
+
|
239
|
+
return test_data, X_test_id
|
240
|
+
|
241
|
+
|
242
|
+
|
243
|
+
|
244
|
+
|
245
|
+
# 9層 CNNモデル 作成
|
246
|
+
|
247
|
+
def layer_9_model():
|
248
|
+
|
249
|
+
|
250
|
+
|
251
|
+
# KerasのSequentialをモデルの元として使用 ---①
|
252
|
+
|
253
|
+
model = Sequential()
|
254
|
+
|
255
|
+
|
256
|
+
|
257
|
+
# 畳み込み層(Convolution)をモデルに追加 ---②
|
258
|
+
|
259
|
+
model.add(Convolution2D(32, 3, 3, border_mode='same', activation='linear',
|
260
|
+
|
261
|
+
input_shape=(img_rows, img_cols, 3)))
|
262
|
+
|
263
|
+
model.add(LeakyReLU(alpha=0.3))
|
264
|
+
|
265
|
+
|
266
|
+
|
267
|
+
model.add(Convolution2D(32, 3, 3, border_mode='same', activation='linear'))
|
268
|
+
|
269
|
+
model.add(LeakyReLU(alpha=0.3))
|
270
|
+
|
271
|
+
|
272
|
+
|
273
|
+
# プーリング層(MaxPooling)をモデルに追加 ---③
|
274
|
+
|
275
|
+
model.add(MaxPooling2D((2, 2), strides=(2, 2)))
|
276
|
+
|
277
|
+
|
278
|
+
|
279
|
+
model.add(Convolution2D(64, 3, 3, border_mode='same', activation='linear'))
|
280
|
+
|
281
|
+
model.add(LeakyReLU(alpha=0.3))
|
282
|
+
|
283
|
+
model.add(Convolution2D(64, 3, 3, border_mode='same', activation='linear'))
|
284
|
+
|
285
|
+
model.add(LeakyReLU(alpha=0.3))
|
286
|
+
|
287
|
+
model.add(MaxPooling2D((2, 2), strides=(2, 2)))
|
288
|
+
|
289
|
+
|
290
|
+
|
291
|
+
model.add(Convolution2D(128, 3, 3, border_mode='same', activation='linear'))
|
292
|
+
|
293
|
+
model.add(LeakyReLU(alpha=0.3))
|
294
|
+
|
295
|
+
model.add(Convolution2D(128, 3, 3, border_mode='same', activation='linear'))
|
296
|
+
|
297
|
+
model.add(LeakyReLU(alpha=0.3))
|
298
|
+
|
299
|
+
model.add(MaxPooling2D((2, 2), strides=(2, 2)))
|
300
|
+
|
301
|
+
|
302
|
+
|
303
|
+
# Flatten層をモデルに追加 -- ④
|
304
|
+
|
305
|
+
model.add(Flatten())
|
306
|
+
|
307
|
+
# 全接続層(Dense)をモデルに追加 --- ⑤
|
308
|
+
|
309
|
+
model.add(Dense(1024, activation='linear'))
|
310
|
+
|
311
|
+
model.add(LeakyReLU(alpha=0.3))
|
312
|
+
|
313
|
+
# Dropout層をモデルに追加 --- ⑥
|
314
|
+
|
315
|
+
model.add(Dropout(0.5))
|
316
|
+
|
317
|
+
model.add(Dense(1024, activation='linear'))
|
318
|
+
|
319
|
+
model.add(LeakyReLU(alpha=0.3))
|
320
|
+
|
321
|
+
model.add(Dropout(0.5))
|
322
|
+
|
323
|
+
# 最終的なアウトプットを作成。 --- ⑦
|
324
|
+
|
325
|
+
model.add(Dense(6, activation='softmax'))
|
326
|
+
|
327
|
+
|
328
|
+
|
329
|
+
# ロス計算や勾配計算に使用する式を定義する。 --- ⑧
|
330
|
+
|
331
|
+
sgd = SGD(lr=1e-3, decay=1e-6, momentum=0.9, nesterov=True)
|
332
|
+
|
333
|
+
model.compile(optimizer=sgd,
|
334
|
+
|
335
|
+
loss='categorical_crossentropy', metrics=["accuracy"])
|
336
|
+
|
337
|
+
return model
|
338
|
+
|
339
|
+
|
340
|
+
|
341
|
+
|
342
|
+
|
343
|
+
# モデルの構成と重みを読み込む
|
344
|
+
|
345
|
+
def read_model(ho, modelStr='', epoch='00'):
|
346
|
+
|
347
|
+
# モデル構成のファイル名
|
348
|
+
|
349
|
+
json_name = 'architecture_%s_%i.json'%(modelStr, ho)
|
350
|
+
|
351
|
+
# モデル重みのファイル名
|
352
|
+
|
353
|
+
weight_name = 'model_weights_%s_%i_%s.h5'%(modelStr, ho, epoch)
|
354
|
+
|
355
|
+
|
356
|
+
|
357
|
+
# モデルの構成を読込み、jsonからモデルオブジェクトへ変換
|
358
|
+
|
359
|
+
model = model_from_json(open(os.path.join('cache', json_name)).read())
|
360
|
+
|
361
|
+
# モデルオブジェクトへ重みを読み込む
|
362
|
+
|
363
|
+
model.load_weights(os.path.join('cache', weight_name))
|
364
|
+
|
365
|
+
|
366
|
+
|
367
|
+
return model
|
368
|
+
|
369
|
+
|
370
|
+
|
371
|
+
|
372
|
+
|
373
|
+
# モデルの構成を保存
|
374
|
+
|
375
|
+
def save_model(model, ho, modelStr=''):
|
376
|
+
|
377
|
+
# モデルオブジェクトをjson形式に変換
|
378
|
+
|
379
|
+
json_string = model.to_json()
|
380
|
+
|
381
|
+
# カレントディレクトリにcacheディレクトリがなければ作成
|
382
|
+
|
383
|
+
if not os.path.isdir('cache'):
|
384
|
+
|
385
|
+
os.mkdir('cache')
|
386
|
+
|
387
|
+
# モデルの構成を保存するためのファイル名
|
388
|
+
|
389
|
+
json_name = 'architecture_%s_%i.json'%(modelStr, ho)
|
390
|
+
|
391
|
+
# モデル構成を保存
|
392
|
+
|
393
|
+
open(os.path.join('cache', json_name), 'w').write(json_string)
|
394
|
+
|
395
|
+
|
396
|
+
|
397
|
+
|
398
|
+
|
399
|
+
def run_train(modelStr=''):
|
400
|
+
|
401
|
+
|
402
|
+
|
403
|
+
# HoldOut 2回行う
|
404
|
+
|
405
|
+
for ho in range(2):
|
406
|
+
|
407
|
+
|
408
|
+
|
409
|
+
# モデルの作成
|
410
|
+
|
411
|
+
model = layer_9_model()
|
412
|
+
|
413
|
+
|
414
|
+
|
415
|
+
# trainデータ読み込み
|
416
|
+
|
417
|
+
t_data, t_target = read_train_data(ho, 'train')
|
418
|
+
|
419
|
+
v_data, v_target = read_train_data(ho, 'valid')
|
420
|
+
|
421
|
+
|
422
|
+
|
423
|
+
# CheckPointを設定。エポック毎にweightsを保存する。
|
424
|
+
|
425
|
+
cp = ModelCheckpoint('./cache/model_weights_%s_%i_{epoch:02d}.h5'%(modelStr, ho),
|
426
|
+
|
427
|
+
monitor='val_loss', save_best_only=False)
|
428
|
+
|
429
|
+
|
430
|
+
|
431
|
+
# train実行
|
432
|
+
|
433
|
+
print(t_data.shape)
|
434
|
+
|
435
|
+
print(t_data.dtype)
|
436
|
+
|
437
|
+
print(v_data.shape)
|
438
|
+
|
439
|
+
print(v_data.dtype)
|
440
|
+
|
441
|
+
print(t_target.shape)
|
442
|
+
|
443
|
+
print(t_target.dtype)
|
444
|
+
|
445
|
+
print(v_target.shape)
|
446
|
+
|
447
|
+
print(v_target.dtype)
|
448
|
+
|
449
|
+
model.fit(t_data, t_target, batch_size=64,
|
450
|
+
|
451
|
+
nb_epoch=40,
|
452
|
+
|
453
|
+
verbose=1,
|
454
|
+
|
455
|
+
validation_data=(v_data, v_target),
|
456
|
+
|
457
|
+
shuffle=True,
|
458
|
+
|
459
|
+
callbacks=[cp])
|
460
|
+
|
461
|
+
|
462
|
+
|
463
|
+
|
464
|
+
|
465
|
+
# モデルの構成を保存
|
466
|
+
|
467
|
+
save_model(model, ho, modelStr)
|
468
|
+
|
469
|
+
|
470
|
+
|
471
|
+
|
472
|
+
|
473
|
+
|
474
|
+
|
475
|
+
# テストデータのクラスを推測
|
476
|
+
|
477
|
+
def run_test(modelStr, epoch1, epoch2):
|
478
|
+
|
479
|
+
|
480
|
+
|
481
|
+
# クラス名取得
|
482
|
+
|
483
|
+
columns = []
|
484
|
+
|
485
|
+
for line in open("../../data/Caltech-101/label.csv", 'r'):
|
486
|
+
|
487
|
+
sp = line.split(',')
|
488
|
+
|
489
|
+
for column in sp:
|
490
|
+
|
491
|
+
columns.append(column.split(":")[1])
|
492
|
+
|
493
|
+
|
494
|
+
|
495
|
+
# テストデータが各クラスに分かれているので、
|
496
|
+
|
497
|
+
# 1クラスずつ読み込んで推測を行う。
|
498
|
+
|
499
|
+
for test_class in range(0, 6):
|
500
|
+
|
501
|
+
|
502
|
+
|
503
|
+
yfull_test = []
|
504
|
+
|
505
|
+
|
506
|
+
|
507
|
+
# データ拡張した画像を読み込むために5回繰り返す
|
508
|
+
|
509
|
+
for aug_i in range(0,5):
|
510
|
+
|
511
|
+
|
512
|
+
|
513
|
+
# テストデータを読み込む
|
514
|
+
|
515
|
+
test_data, test_id = load_test(test_class, aug_i)
|
516
|
+
|
517
|
+
|
518
|
+
|
519
|
+
# HoldOut 2回繰り返す
|
520
|
+
|
521
|
+
for ho in range(2):
|
522
|
+
|
523
|
+
|
524
|
+
|
525
|
+
if ho == 0:
|
526
|
+
|
527
|
+
epoch_n = epoch1
|
528
|
+
|
529
|
+
else:
|
530
|
+
|
531
|
+
epoch_n = epoch2
|
532
|
+
|
533
|
+
|
534
|
+
|
535
|
+
# 学習済みモデルの読み込み
|
536
|
+
|
537
|
+
model = read_model(ho, modelStr, epoch_n)
|
538
|
+
|
539
|
+
|
540
|
+
|
541
|
+
# 推測の実行
|
542
|
+
|
543
|
+
test_p = model.predict(test_data, batch_size=128, verbose=1)
|
544
|
+
|
545
|
+
|
546
|
+
|
547
|
+
yfull_test.append(test_p)
|
548
|
+
|
549
|
+
|
550
|
+
|
551
|
+
# 推測結果の平均化
|
552
|
+
|
553
|
+
test_res = np.array(yfull_test[0])
|
554
|
+
|
555
|
+
for i in range(1,10):
|
556
|
+
|
557
|
+
test_res += np.array(yfull_test[i])
|
558
|
+
|
559
|
+
test_res /= 10
|
560
|
+
|
561
|
+
|
562
|
+
|
563
|
+
# 推測結果とクラス名、画像名を合わせる
|
564
|
+
|
565
|
+
result1 = pd.DataFrame(test_res, columns=columns)
|
566
|
+
|
567
|
+
result1.loc[:, 'img'] = pd.Series(test_id, index=result1.index)
|
568
|
+
|
569
|
+
|
570
|
+
|
571
|
+
# 順番入れ替え
|
572
|
+
|
573
|
+
result1 = result1.ix[:,[6, 0, 1, 2, 3, 4, 5]]
|
574
|
+
|
575
|
+
|
576
|
+
|
577
|
+
if not os.path.isdir('subm'):
|
578
|
+
|
579
|
+
os.mkdir('subm')
|
580
|
+
|
581
|
+
sub_file = './subm/result_%s_%i.csv'%(modelStr, test_class)
|
582
|
+
|
583
|
+
|
584
|
+
|
585
|
+
# 最終推測結果を出力する
|
586
|
+
|
587
|
+
result1.to_csv(sub_file, index=False)
|
588
|
+
|
589
|
+
|
590
|
+
|
591
|
+
# 推測の精度を測定する。
|
592
|
+
|
593
|
+
# 一番大きい値が入っているカラムがtest_classであるレコードを探す
|
594
|
+
|
595
|
+
one_column = np.where(np.argmax(test_res, axis=1)==test_class)
|
596
|
+
|
597
|
+
print ("正解数 " + str(len(one_column[0])))
|
598
|
+
|
599
|
+
print ("不正解数 " + str(test_res.shape[0] - len(one_column[0])))
|
600
|
+
|
601
|
+
|
602
|
+
|
603
|
+
|
604
|
+
|
605
|
+
|
606
|
+
|
607
|
+
|
608
|
+
|
609
|
+
# 実行した際に呼ばれる
|
610
|
+
|
611
|
+
if __name__ == '__main__':
|
612
|
+
|
613
|
+
|
614
|
+
|
615
|
+
# 引数を取得
|
616
|
+
|
617
|
+
# [1] = train or test
|
618
|
+
|
619
|
+
# [2] = test時のみ、使用Epoch数 1
|
620
|
+
|
621
|
+
# [3] = test時のみ、使用Epoch数 2
|
622
|
+
|
623
|
+
param = sys.argv
|
624
|
+
|
625
|
+
|
626
|
+
|
627
|
+
if len(param) < 2:
|
628
|
+
|
629
|
+
sys.exit ("Usage: python 9_Layer_CNN.py [train, test] [1] [2]")
|
630
|
+
|
631
|
+
|
632
|
+
|
633
|
+
# train or test
|
634
|
+
|
635
|
+
run_type = param[1]
|
636
|
+
|
637
|
+
|
638
|
+
|
639
|
+
if run_type == 'train':
|
640
|
+
|
641
|
+
run_train('9_Layer_CNN')
|
642
|
+
|
643
|
+
elif run_type == 'test':
|
644
|
+
|
645
|
+
# testの場合、使用するエポック数を引数から取得する
|
646
|
+
|
647
|
+
if len(param) == 4:
|
648
|
+
|
649
|
+
epoch1 = "%02d"%(int(param[2])-1)
|
650
|
+
|
651
|
+
epoch2 = "%02d"%(int(param[3])-1)
|
652
|
+
|
653
|
+
run_test('9_Layer_CNN', epoch1, epoch2)
|
654
|
+
|
655
|
+
else:
|
656
|
+
|
657
|
+
sys.exit ("Usage: python 9_Layer_CNN.py [train, test] [1] [2]")
|
658
|
+
|
659
|
+
else:
|
660
|
+
|
661
|
+
sys.exit ("Usage: python 9_Layer_CNN.py [train, test] [1] [2]")
|
662
|
+
|
663
|
+
```
|
2
質問事項の追加をしました。
test
CHANGED
File without changes
|
test
CHANGED
@@ -12,650 +12,160 @@
|
|
12
12
|
|
13
13
|
```
|
14
14
|
|
15
|
+
|
16
|
+
|
17
|
+
|
18
|
+
|
19
|
+
I tensorflow/core/common_runtime/bfc_allocator.cc:702] Stats:
|
20
|
+
|
21
|
+
Limit: 7582672487
|
22
|
+
|
23
|
+
InUse: 6971735552
|
24
|
+
|
25
|
+
MaxInUse: 7541746432
|
26
|
+
|
27
|
+
NumAllocs: 443
|
28
|
+
|
29
|
+
MaxAllocSize: 3043936256
|
30
|
+
|
31
|
+
|
32
|
+
|
33
|
+
W tensorflow/core/common_runtime/bfc_allocator.cc:274] **************************_**___****************************************************************xxxx
|
34
|
+
|
35
|
+
W tensorflow/core/common_runtime/bfc_allocator.cc:275] Ran out of memory trying to allocate 392.00MiB. See logs for memory state.
|
36
|
+
|
37
|
+
W tensorflow/core/framework/op_kernel.cc:993] Resource exhausted: OOM when allocating tensor with shape[100352,1024]
|
38
|
+
|
15
39
|
Traceback (most recent call last):
|
16
40
|
|
17
|
-
|
41
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1022, in _do_call
|
18
42
|
|
19
|
-
|
43
|
+
return fn(*args)
|
20
44
|
|
21
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/
|
45
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1004, in _run_fn
|
22
46
|
|
23
|
-
|
47
|
+
status, run_metadata)
|
24
48
|
|
25
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/
|
49
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/contextlib.py", line 88, in __exit__
|
26
50
|
|
27
|
-
|
51
|
+
next(self.gen)
|
28
52
|
|
53
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/framework/errors_impl.py", line 469, in raise_exception_on_not_ok_status
|
54
|
+
|
55
|
+
pywrap_tensorflow.TF_GetCode(status))
|
56
|
+
|
57
|
+
tensorflow.python.framework.errors_impl.ResourceExhaustedError: OOM when allocating tensor with shape[100352,1024]
|
58
|
+
|
59
|
+
[[Node: sub_28 = Sub[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/gpu:0"](add_14, mul_60)]]
|
60
|
+
|
61
|
+
|
62
|
+
|
63
|
+
During handling of the above exception, another exception occurred:
|
64
|
+
|
65
|
+
|
66
|
+
|
67
|
+
Traceback (most recent call last):
|
68
|
+
|
69
|
+
File "n.py", line 299, in <module>
|
70
|
+
|
71
|
+
run_train('9_Layer_CNN')
|
72
|
+
|
73
|
+
File "n.py", line 208, in run_train
|
74
|
+
|
75
|
+
callbacks=[cp])
|
76
|
+
|
29
|
-
|
77
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/models.py", line 870, in fit
|
78
|
+
|
79
|
+
initial_epoch=initial_epoch)
|
80
|
+
|
81
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/engine/training.py", line 1507, in fit
|
82
|
+
|
83
|
+
initial_epoch=initial_epoch)
|
84
|
+
|
85
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/engine/training.py", line 1156, in _fit_loop
|
86
|
+
|
87
|
+
outs = f(ins_batch)
|
88
|
+
|
89
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 2269, in __call__
|
90
|
+
|
91
|
+
**self.session_kwargs)
|
92
|
+
|
93
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 767, in run
|
94
|
+
|
95
|
+
run_metadata_ptr)
|
96
|
+
|
97
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 965, in _run
|
98
|
+
|
99
|
+
feed_dict_string, options, run_metadata)
|
100
|
+
|
101
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1015, in _do_run
|
102
|
+
|
103
|
+
target_list, options, run_metadata)
|
104
|
+
|
105
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1035, in _do_call
|
106
|
+
|
107
|
+
raise type(e)(node_def, op, message)
|
108
|
+
|
109
|
+
tensorflow.python.framework.errors_impl.ResourceExhaustedError: OOM when allocating tensor with shape[100352,1024]
|
110
|
+
|
111
|
+
[[Node: sub_28 = Sub[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/gpu:0"](add_14, mul_60)]]
|
112
|
+
|
113
|
+
|
114
|
+
|
115
|
+
Caused by op 'sub_28', defined at:
|
116
|
+
|
117
|
+
File "n.py", line 299, in <module>
|
118
|
+
|
119
|
+
run_train('9_Layer_CNN')
|
120
|
+
|
121
|
+
File "n.py", line 208, in run_train
|
122
|
+
|
123
|
+
callbacks=[cp])
|
124
|
+
|
125
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/models.py", line 870, in fit
|
126
|
+
|
127
|
+
initial_epoch=initial_epoch)
|
128
|
+
|
129
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/engine/training.py", line 1490, in fit
|
130
|
+
|
131
|
+
self._make_train_function()
|
132
|
+
|
133
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/engine/training.py", line 1014, in _make_train_function
|
134
|
+
|
135
|
+
self.total_loss)
|
136
|
+
|
137
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/optimizers.py", line 169, in get_updates
|
138
|
+
|
139
|
+
new_p = p + self.momentum * v - lr * g
|
140
|
+
|
141
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/ops/math_ops.py", line 884, in binary_op_wrapper
|
142
|
+
|
143
|
+
return func(x, y, name=name)
|
144
|
+
|
145
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/ops/gen_math_ops.py", line 2775, in _sub
|
146
|
+
|
147
|
+
result = _op_def_lib.apply_op("Sub", x=x, y=y, name=name)
|
148
|
+
|
149
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py", line 763, in apply_op
|
150
|
+
|
151
|
+
op_def=op_def)
|
152
|
+
|
153
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 2395, in create_op
|
154
|
+
|
155
|
+
original_op=self._default_original_op, op_def=op_def)
|
156
|
+
|
157
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 1264, in __init__
|
158
|
+
|
159
|
+
self._traceback = _extract_stack()
|
160
|
+
|
161
|
+
|
162
|
+
|
163
|
+
ResourceExhaustedError (see above for traceback): OOM when allocating tensor with shape[100352,1024]
|
164
|
+
|
165
|
+
[[Node: sub_28 = Sub[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/gpu:0"](add_14, mul_60)]]
|
30
166
|
|
31
167
|
|
32
168
|
|
33
169
|
|
34
170
|
|
35
171
|
```
|
36
|
-
|
37
|
-
```該当ソースコード
|
38
|
-
|
39
|
-
import numpy as np
|
40
|
-
|
41
|
-
from numpy.random import permutation
|
42
|
-
|
43
|
-
|
44
|
-
|
45
|
-
import os, glob, cv2, math, sys
|
46
|
-
|
47
|
-
import pandas as pd
|
48
|
-
|
49
|
-
|
50
|
-
|
51
|
-
from keras.models import Sequential, model_from_json
|
52
|
-
|
53
|
-
from keras.layers.core import Dense, Dropout, Flatten
|
54
|
-
|
55
|
-
from keras.layers.convolutional import Convolution2D, MaxPooling2D
|
56
|
-
|
57
|
-
from keras.layers.advanced_activations import LeakyReLU
|
58
|
-
|
59
|
-
from keras.callbacks import ModelCheckpoint
|
60
|
-
|
61
|
-
from keras.optimizers import SGD
|
62
|
-
|
63
|
-
from keras.utils import np_utils
|
64
|
-
|
65
|
-
|
66
|
-
|
67
|
-
# seed値
|
68
|
-
|
69
|
-
np.random.seed(1)
|
70
|
-
|
71
|
-
|
72
|
-
|
73
|
-
# 使用する画像サイズ
|
74
|
-
|
75
|
-
img_rows, img_cols = 224, 224
|
76
|
-
|
77
|
-
|
78
|
-
|
79
|
-
# 画像データ 1枚の読み込みとリサイズを行う
|
80
|
-
|
81
|
-
def get_im(path):
|
82
|
-
|
83
|
-
|
84
|
-
|
85
|
-
img = cv2.imread(path)
|
86
|
-
|
87
|
-
resized = cv2.resize(img, (img_cols, img_rows))
|
88
|
-
|
89
|
-
|
90
|
-
|
91
|
-
return resized
|
92
|
-
|
93
|
-
|
94
|
-
|
95
|
-
|
96
|
-
|
97
|
-
# データの読み込み、正規化、シャッフルを行う
|
98
|
-
|
99
|
-
def read_train_data(ho=0, kind='train'):
|
100
|
-
|
101
|
-
|
102
|
-
|
103
|
-
train_data = []
|
104
|
-
|
105
|
-
train_target = []
|
106
|
-
|
107
|
-
|
108
|
-
|
109
|
-
# 学習用データ読み込み
|
110
|
-
|
111
|
-
for j in range(0, 6): # 0~5まで
|
112
|
-
|
113
|
-
|
114
|
-
|
115
|
-
path = '../../data/Caltech-101/'
|
116
|
-
|
117
|
-
path += '%s/%i/*/%i/*.jpg'%(kind, ho, j)
|
118
|
-
|
119
|
-
|
120
|
-
|
121
|
-
files = sorted(glob.glob(path))
|
122
|
-
|
123
|
-
|
124
|
-
|
125
|
-
for fl in files:
|
126
|
-
|
127
|
-
|
128
|
-
|
129
|
-
flbase = os.path.basename(fl)
|
130
|
-
|
131
|
-
|
132
|
-
|
133
|
-
# 画像 1枚 読み込み
|
134
|
-
|
135
|
-
img = get_im(fl)
|
136
|
-
|
137
|
-
img = np.array(img, dtype=np.float32)
|
138
|
-
|
139
|
-
|
140
|
-
|
141
|
-
# 正規化(GCN)実行
|
142
|
-
|
143
|
-
img -= np.mean(img)
|
144
|
-
|
145
|
-
img /= np.std(img)
|
146
|
-
|
147
|
-
|
148
|
-
|
149
|
-
train_data.append(img)
|
150
|
-
|
151
|
-
train_target.append(j)
|
152
|
-
|
153
|
-
|
154
|
-
|
155
|
-
# 読み込んだデータを numpy の array に変換
|
156
|
-
|
157
|
-
train_data = np.array(train_data, dtype=np.float32)
|
158
|
-
|
159
|
-
train_target = np.array(train_target, dtype=np.uint8)
|
160
|
-
|
161
|
-
|
162
|
-
|
163
|
-
# (レコード数,縦,横,channel数) を (レコード数,channel数,縦,横) に変換
|
164
|
-
|
165
|
-
#train_data = train_data.transpose((0, 3, 1, 2))
|
166
|
-
|
167
|
-
|
168
|
-
|
169
|
-
# target を 6次元のデータに変換。
|
170
|
-
|
171
|
-
# ex) 1 -> 0,1,0,0,0,0 2 -> 0,0,1,0,0,0
|
172
|
-
|
173
|
-
train_target = np_utils.to_categorical(train_target, 6)
|
174
|
-
|
175
|
-
|
176
|
-
|
177
|
-
# データをシャッフル
|
178
|
-
|
179
|
-
perm = permutation(len(train_target))
|
180
|
-
|
181
|
-
train_data = train_data[perm]
|
182
|
-
|
183
|
-
train_target = train_target[perm]
|
184
|
-
|
185
|
-
|
186
|
-
|
187
|
-
return train_data, train_target
|
188
|
-
|
189
|
-
|
190
|
-
|
191
|
-
|
192
|
-
|
193
|
-
# テストデータ読み込み
|
194
|
-
|
195
|
-
def load_test(test_class, aug_i):
|
196
|
-
|
197
|
-
|
198
|
-
|
199
|
-
path = '../../data/Caltech-101/test/%i/%i/*.jpg'%(aug_i, test_class)
|
200
|
-
|
201
|
-
|
202
|
-
|
203
|
-
files = sorted(glob.glob(path))
|
204
|
-
|
205
|
-
X_test = []
|
206
|
-
|
207
|
-
X_test_id = []
|
208
|
-
|
209
|
-
|
210
|
-
|
211
|
-
for fl in files:
|
212
|
-
|
213
|
-
flbase = os.path.basename(fl)
|
214
|
-
|
215
|
-
|
216
|
-
|
217
|
-
img = get_im(fl)
|
218
|
-
|
219
|
-
img = np.array(img, dtype=np.float32)
|
220
|
-
|
221
|
-
|
222
|
-
|
223
|
-
# 正規化(GCN)実行
|
224
|
-
|
225
|
-
img -= np.mean(img)
|
226
|
-
|
227
|
-
img /= np.std(img)
|
228
|
-
|
229
|
-
|
230
|
-
|
231
|
-
X_test.append(img)
|
232
|
-
|
233
|
-
X_test_id.append(flbase)
|
234
|
-
|
235
|
-
|
236
|
-
|
237
|
-
# 読み込んだデータを numpy の array に変換
|
238
|
-
|
239
|
-
test_data = np.array(X_test, dtype=np.float32)
|
240
|
-
|
241
|
-
|
242
|
-
|
243
|
-
# (レコード数,縦,横,channel数) を (レコード数,channel数,縦,横) に変換
|
244
|
-
|
245
|
-
#test_data = test_data.transpose((0, 3, 1, 2))
|
246
|
-
|
247
|
-
|
248
|
-
|
249
|
-
return test_data, X_test_id
|
250
|
-
|
251
|
-
|
252
|
-
|
253
|
-
|
254
|
-
|
255
|
-
# 9層 CNNモデル 作成
|
256
|
-
|
257
|
-
def layer_9_model():
|
258
|
-
|
259
|
-
|
260
|
-
|
261
|
-
# KerasのSequentialをモデルの元として使用 ---①
|
262
|
-
|
263
|
-
model = Sequential()
|
264
|
-
|
265
|
-
|
266
|
-
|
267
|
-
# 畳み込み層(Convolution)をモデルに追加 ---②
|
268
|
-
|
269
|
-
model.add(Convolution2D(32, 3, 3, border_mode='same', activation='linear',
|
270
|
-
|
271
|
-
input_shape=(img_rows, img_cols, 3)))
|
272
|
-
|
273
|
-
model.add(LeakyReLU(alpha=0.3))
|
274
|
-
|
275
|
-
|
276
|
-
|
277
|
-
model.add(Convolution2D(32, 3, 3, border_mode='same', activation='linear'))
|
278
|
-
|
279
|
-
model.add(LeakyReLU(alpha=0.3))
|
280
|
-
|
281
|
-
|
282
|
-
|
283
|
-
# プーリング層(MaxPooling)をモデルに追加 ---③
|
284
|
-
|
285
|
-
model.add(MaxPooling2D((2, 2), strides=(2, 2)))
|
286
|
-
|
287
|
-
|
288
|
-
|
289
|
-
model.add(Convolution2D(64, 3, 3, border_mode='same', activation='linear'))
|
290
|
-
|
291
|
-
model.add(LeakyReLU(alpha=0.3))
|
292
|
-
|
293
|
-
model.add(Convolution2D(64, 3, 3, border_mode='same', activation='linear'))
|
294
|
-
|
295
|
-
model.add(LeakyReLU(alpha=0.3))
|
296
|
-
|
297
|
-
model.add(MaxPooling2D((2, 2), strides=(2, 2)))
|
298
|
-
|
299
|
-
|
300
|
-
|
301
|
-
model.add(Convolution2D(128, 3, 3, border_mode='same', activation='linear'))
|
302
|
-
|
303
|
-
model.add(LeakyReLU(alpha=0.3))
|
304
|
-
|
305
|
-
model.add(Convolution2D(128, 3, 3, border_mode='same', activation='linear'))
|
306
|
-
|
307
|
-
model.add(LeakyReLU(alpha=0.3))
|
308
|
-
|
309
|
-
model.add(MaxPooling2D((2, 2), strides=(2, 2)))
|
310
|
-
|
311
|
-
|
312
|
-
|
313
|
-
# Flatten層をモデルに追加 -- ④
|
314
|
-
|
315
|
-
model.add(Flatten())
|
316
|
-
|
317
|
-
# 全接続層(Dense)をモデルに追加 --- ⑤
|
318
|
-
|
319
|
-
model.add(Dense(1024, activation='linear'))
|
320
|
-
|
321
|
-
model.add(LeakyReLU(alpha=0.3))
|
322
|
-
|
323
|
-
# Dropout層をモデルに追加 --- ⑥
|
324
|
-
|
325
|
-
model.add(Dropout(0.5))
|
326
|
-
|
327
|
-
model.add(Dense(1024, activation='linear'))
|
328
|
-
|
329
|
-
model.add(LeakyReLU(alpha=0.3))
|
330
|
-
|
331
|
-
model.add(Dropout(0.5))
|
332
|
-
|
333
|
-
# 最終的なアウトプットを作成。 --- ⑦
|
334
|
-
|
335
|
-
model.add(Dense(6, activation='softmax'))
|
336
|
-
|
337
|
-
|
338
|
-
|
339
|
-
# ロス計算や勾配計算に使用する式を定義する。 --- ⑧
|
340
|
-
|
341
|
-
sgd = SGD(lr=1e-3, decay=1e-6, momentum=0.9, nesterov=True)
|
342
|
-
|
343
|
-
model.compile(optimizer=sgd,
|
344
|
-
|
345
|
-
loss='categorical_crossentropy', metrics=["accuracy"])
|
346
|
-
|
347
|
-
return model
|
348
|
-
|
349
|
-
|
350
|
-
|
351
|
-
|
352
|
-
|
353
|
-
# モデルの構成と重みを読み込む
|
354
|
-
|
355
|
-
def read_model(ho, modelStr='', epoch='00'):
|
356
|
-
|
357
|
-
# モデル構成のファイル名
|
358
|
-
|
359
|
-
json_name = 'architecture_%s_%i.json'%(modelStr, ho)
|
360
|
-
|
361
|
-
# モデル重みのファイル名
|
362
|
-
|
363
|
-
weight_name = 'model_weights_%s_%i_%s.h5'%(modelStr, ho, epoch)
|
364
|
-
|
365
|
-
|
366
|
-
|
367
|
-
# モデルの構成を読込み、jsonからモデルオブジェクトへ変換
|
368
|
-
|
369
|
-
model = model_from_json(open(os.path.join('cache', json_name)).read())
|
370
|
-
|
371
|
-
# モデルオブジェクトへ重みを読み込む
|
372
|
-
|
373
|
-
model.load_weights(os.path.join('cache', weight_name))
|
374
|
-
|
375
|
-
|
376
|
-
|
377
|
-
return model
|
378
|
-
|
379
|
-
|
380
|
-
|
381
|
-
|
382
|
-
|
383
|
-
# モデルの構成を保存
|
384
|
-
|
385
|
-
def save_model(model, ho, modelStr=''):
|
386
|
-
|
387
|
-
# モデルオブジェクトをjson形式に変換
|
388
|
-
|
389
|
-
json_string = model.to_json()
|
390
|
-
|
391
|
-
# カレントディレクトリにcacheディレクトリがなければ作成
|
392
|
-
|
393
|
-
if not os.path.isdir('cache'):
|
394
|
-
|
395
|
-
os.mkdir('cache')
|
396
|
-
|
397
|
-
# モデルの構成を保存するためのファイル名
|
398
|
-
|
399
|
-
json_name = 'architecture_%s_%i.json'%(modelStr, ho)
|
400
|
-
|
401
|
-
# モデル構成を保存
|
402
|
-
|
403
|
-
open(os.path.join('cache', json_name), 'w').write(json_string)
|
404
|
-
|
405
|
-
|
406
|
-
|
407
|
-
|
408
|
-
|
409
|
-
def run_train(modelStr=''):
|
410
|
-
|
411
|
-
|
412
|
-
|
413
|
-
# HoldOut 2回行う
|
414
|
-
|
415
|
-
for ho in range(2):
|
416
|
-
|
417
|
-
|
418
|
-
|
419
|
-
# モデルの作成
|
420
|
-
|
421
|
-
model = layer_9_model()
|
422
|
-
|
423
|
-
|
424
|
-
|
425
|
-
# trainデータ読み込み
|
426
|
-
|
427
|
-
t_data, t_target = read_train_data(ho, 'train')
|
428
|
-
|
429
|
-
v_data, v_target = read_train_data(ho, 'valid')
|
430
|
-
|
431
|
-
|
432
|
-
|
433
|
-
# CheckPointを設定。エポック毎にweightsを保存する。
|
434
|
-
|
435
|
-
cp = ModelCheckpoint('./cache/model_weights_%s_%i_{epoch:02d}.h5'%(modelStr, ho),
|
436
|
-
|
437
|
-
monitor='val_loss', save_best_only=False)
|
438
|
-
|
439
|
-
|
440
|
-
|
441
|
-
# train実行
|
442
|
-
|
443
|
-
model.fit(t_data, t_target, batch_size=64,
|
444
|
-
|
445
|
-
nb_epoch=40,
|
446
|
-
|
447
|
-
verbose=1,
|
448
|
-
|
449
|
-
validation_data=(v_data, v_target),
|
450
|
-
|
451
|
-
shuffle=True,
|
452
|
-
|
453
|
-
callbacks=[cp])
|
454
|
-
|
455
|
-
|
456
|
-
|
457
|
-
|
458
|
-
|
459
|
-
# モデルの構成を保存
|
460
|
-
|
461
|
-
save_model(model, ho, modelStr)
|
462
|
-
|
463
|
-
|
464
|
-
|
465
|
-
|
466
|
-
|
467
|
-
|
468
|
-
|
469
|
-
# テストデータのクラスを推測
|
470
|
-
|
471
|
-
def run_test(modelStr, epoch1, epoch2):
|
472
|
-
|
473
|
-
|
474
|
-
|
475
|
-
# クラス名取得
|
476
|
-
|
477
|
-
columns = []
|
478
|
-
|
479
|
-
for line in open("../../data/Caltech-101/label.csv", 'r'):
|
480
|
-
|
481
|
-
sp = line.split(',')
|
482
|
-
|
483
|
-
for column in sp:
|
484
|
-
|
485
|
-
columns.append(column.split(":")[1])
|
486
|
-
|
487
|
-
|
488
|
-
|
489
|
-
# テストデータが各クラスに分かれているので、
|
490
|
-
|
491
|
-
# 1クラスずつ読み込んで推測を行う。
|
492
|
-
|
493
|
-
for test_class in range(0, 6):
|
494
|
-
|
495
|
-
|
496
|
-
|
497
|
-
yfull_test = []
|
498
|
-
|
499
|
-
|
500
|
-
|
501
|
-
# データ拡張した画像を読み込むために5回繰り返す
|
502
|
-
|
503
|
-
for aug_i in range(0,5):
|
504
|
-
|
505
|
-
|
506
|
-
|
507
|
-
# テストデータを読み込む
|
508
|
-
|
509
|
-
test_data, test_id = load_test(test_class, aug_i)
|
510
|
-
|
511
|
-
|
512
|
-
|
513
|
-
# HoldOut 2回繰り返す
|
514
|
-
|
515
|
-
for ho in range(2):
|
516
|
-
|
517
|
-
|
518
|
-
|
519
|
-
if ho == 0:
|
520
|
-
|
521
|
-
epoch_n = epoch1
|
522
|
-
|
523
|
-
else:
|
524
|
-
|
525
|
-
epoch_n = epoch2
|
526
|
-
|
527
|
-
|
528
|
-
|
529
|
-
# 学習済みモデルの読み込み
|
530
|
-
|
531
|
-
model = read_model(ho, modelStr, epoch_n)
|
532
|
-
|
533
|
-
|
534
|
-
|
535
|
-
# 推測の実行
|
536
|
-
|
537
|
-
test_p = model.predict(test_data, batch_size=128, verbose=1)
|
538
|
-
|
539
|
-
|
540
|
-
|
541
|
-
yfull_test.append(test_p)
|
542
|
-
|
543
|
-
|
544
|
-
|
545
|
-
# 推測結果の平均化
|
546
|
-
|
547
|
-
test_res = np.array(yfull_test[0])
|
548
|
-
|
549
|
-
for i in range(1,10):
|
550
|
-
|
551
|
-
test_res += np.array(yfull_test[i])
|
552
|
-
|
553
|
-
test_res /= 10
|
554
|
-
|
555
|
-
|
556
|
-
|
557
|
-
# 推測結果とクラス名、画像名を合わせる
|
558
|
-
|
559
|
-
result1 = pd.DataFrame(test_res, columns=columns)
|
560
|
-
|
561
|
-
result1.loc[:, 'img'] = pd.Series(test_id, index=result1.index)
|
562
|
-
|
563
|
-
|
564
|
-
|
565
|
-
# 順番入れ替え
|
566
|
-
|
567
|
-
result1 = result1.ix[:,[6, 0, 1, 2, 3, 4, 5]]
|
568
|
-
|
569
|
-
|
570
|
-
|
571
|
-
if not os.path.isdir('subm'):
|
572
|
-
|
573
|
-
os.mkdir('subm')
|
574
|
-
|
575
|
-
sub_file = './subm/result_%s_%i.csv'%(modelStr, test_class)
|
576
|
-
|
577
|
-
|
578
|
-
|
579
|
-
# 最終推測結果を出力する
|
580
|
-
|
581
|
-
result1.to_csv(sub_file, index=False)
|
582
|
-
|
583
|
-
|
584
|
-
|
585
|
-
# 推測の精度を測定する。
|
586
|
-
|
587
|
-
# 一番大きい値が入っているカラムがtest_classであるレコードを探す
|
588
|
-
|
589
|
-
one_column = np.where(np.argmax(test_res, axis=1)==test_class)
|
590
|
-
|
591
|
-
print ("正解数 " + str(len(one_column[0])))
|
592
|
-
|
593
|
-
print ("不正解数 " + str(test_res.shape[0] - len(one_column[0])))
|
594
|
-
|
595
|
-
|
596
|
-
|
597
|
-
|
598
|
-
|
599
|
-
|
600
|
-
|
601
|
-
|
602
|
-
|
603
|
-
# 実行した際に呼ばれる
|
604
|
-
|
605
|
-
if __name__ == '__main__':
|
606
|
-
|
607
|
-
|
608
|
-
|
609
|
-
# 引数を取得
|
610
|
-
|
611
|
-
# [1] = train or test
|
612
|
-
|
613
|
-
# [2] = test時のみ、使用Epoch数 1
|
614
|
-
|
615
|
-
# [3] = test時のみ、使用Epoch数 2
|
616
|
-
|
617
|
-
param = sys.argv
|
618
|
-
|
619
|
-
|
620
|
-
|
621
|
-
if len(param) < 2:
|
622
|
-
|
623
|
-
sys.exit ("Usage: python 9_Layer_CNN.py [train, test] [1] [2]")
|
624
|
-
|
625
|
-
|
626
|
-
|
627
|
-
# train or test
|
628
|
-
|
629
|
-
run_type = param[1]
|
630
|
-
|
631
|
-
|
632
|
-
|
633
|
-
if run_type == 'train':
|
634
|
-
|
635
|
-
run_train('9_Layer_CNN')
|
636
|
-
|
637
|
-
elif run_type == 'test':
|
638
|
-
|
639
|
-
# testの場合、使用するエポック数を引数から取得する
|
640
|
-
|
641
|
-
if len(param) == 4:
|
642
|
-
|
643
|
-
epoch1 = "%02d"%(int(param[2])-1)
|
644
|
-
|
645
|
-
epoch2 = "%02d"%(int(param[3])-1)
|
646
|
-
|
647
|
-
run_test('9_Layer_CNN', epoch1, epoch2)
|
648
|
-
|
649
|
-
else:
|
650
|
-
|
651
|
-
sys.exit ("Usage: python 9_Layer_CNN.py [train, test] [1] [2]")
|
652
|
-
|
653
|
-
else:
|
654
|
-
|
655
|
-
sys.exit ("Usage: python 9_Layer_CNN.py [train, test] [1] [2]")
|
656
|
-
|
657
|
-
|
658
|
-
|
659
|
-
|
660
|
-
|
661
|
-
```
|
1
コード2
test
CHANGED
File without changes
|
test
CHANGED
@@ -12,39 +12,23 @@
|
|
12
12
|
|
13
13
|
```
|
14
14
|
|
15
|
-
Using TensorFlow backend.
|
16
|
-
|
17
|
-
|
18
|
-
|
19
|
-
|
20
|
-
|
21
15
|
Traceback (most recent call last):
|
22
16
|
|
23
|
-
|
17
|
+
|
24
|
-
|
25
|
-
|
18
|
+
|
26
|
-
|
27
|
-
|
19
|
+
|
28
|
-
|
29
|
-
|
20
|
+
|
30
|
-
|
31
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/models.py", line 672, in fit
|
32
|
-
|
33
|
-
initial_epoch=initial_epoch)
|
34
|
-
|
35
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/keras/engine/training.py", line 1116, in fit
|
36
|
-
|
37
|
-
batch_size=batch_size)
|
38
|
-
|
39
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/
|
21
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/framework/common_shapes.py", line 610, in call_cpp_shape_fn
|
40
|
-
|
22
|
+
|
41
|
-
e
|
23
|
+
debug_python_shape_fn, require_shape_fn)
|
42
|
-
|
24
|
+
|
43
|
-
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/
|
25
|
+
File "/home/a/libraries/anaconda3/envs/main/lib/python3.6/site-packages/tensorflow/python/framework/common_shapes.py", line 675, in _call_cpp_shape_fn_impl
|
44
|
-
|
26
|
+
|
45
|
-
s
|
27
|
+
raise ValueError(err.message)
|
46
|
-
|
28
|
+
|
47
|
-
ValueError:
|
29
|
+
ValueError: Negative dimension size caused by subtracting 2 from 1 for 'max_pooling2d_2/MaxPool' (op: 'MaxPool') with input shapes: [?,112,1,64].
|
30
|
+
|
31
|
+
|
48
32
|
|
49
33
|
|
50
34
|
|
@@ -284,7 +268,7 @@
|
|
284
268
|
|
285
269
|
model.add(Convolution2D(32, 3, 3, border_mode='same', activation='linear',
|
286
270
|
|
287
|
-
|
271
|
+
input_shape=(img_rows, img_cols, 3)))
|
288
272
|
|
289
273
|
model.add(LeakyReLU(alpha=0.3))
|
290
274
|
|