質問編集履歴
1
code delete
test
CHANGED
File without changes
|
test
CHANGED
@@ -10,295 +10,7 @@
|
|
10
10
|
|
11
11
|
|
12
12
|
|
13
|
-
```python
|
14
13
|
|
15
|
-
import numpy as np
|
16
|
-
|
17
|
-
import tensorflow as tf
|
18
|
-
|
19
|
-
import csv
|
20
|
-
|
21
|
-
import csv_decode as csvd
|
22
|
-
|
23
|
-
from matplotlib import pyplot as plt
|
24
|
-
|
25
|
-
import cv2
|
26
|
-
|
27
|
-
import os
|
28
|
-
|
29
|
-
import sys
|
30
|
-
|
31
|
-
|
32
|
-
|
33
|
-
# define input data
|
34
|
-
|
35
|
-
# [batch_size, height, width, channel]
|
36
|
-
|
37
|
-
img_input = tf.placeholder(tf.float32, [None, 224, 224, 3])
|
38
|
-
|
39
|
-
|
40
|
-
|
41
|
-
# conv layer 1
|
42
|
-
|
43
|
-
# [height, width, channel, number of filter]
|
44
|
-
|
45
|
-
block1_f1 = tf.Variable(tf.truncated_normal([5, 5, 3, 64], stddev=0.1), name='block1_f1')
|
46
|
-
|
47
|
-
# [batch direction, height direction, width direction, channel direction]
|
48
|
-
|
49
|
-
block1_conv1 = tf.nn.conv2d(img_input, block1_f1, strides=[1, 1, 1, 1], padding='SAME', name='block1_conv1')
|
50
|
-
|
51
|
-
block1_b1 = tf.Variable(tf.constant(0.1, shape=[64]), name='block1_b1')
|
52
|
-
|
53
|
-
block1_h_conv1 = tf.nn.relu(block1_conv1 + block1_b1)
|
54
|
-
|
55
|
-
|
56
|
-
|
57
|
-
# conv layer 2
|
58
|
-
|
59
|
-
block1_f2 = tf.Variable(tf.truncated_normal([5, 5, 64, 64], stddev=0.1), name='block1_f2')
|
60
|
-
|
61
|
-
block1_conv2 = tf.nn.conv2d(block1_h_conv1, block1_f2, strides=[1, 1, 1, 1], padding='SAME', name='block1_conv2')
|
62
|
-
|
63
|
-
block1_b2 = tf.Variable(tf.constant(0.1, shape=[64]), name='block1_b2')
|
64
|
-
|
65
|
-
block1_h_conv2 = tf.nn.relu(block1_conv2 + block1_b2)
|
66
|
-
|
67
|
-
|
68
|
-
|
69
|
-
# pooling layer 1
|
70
|
-
|
71
|
-
# [batch direction, height direction, width direction, channel direction]
|
72
|
-
|
73
|
-
block1_pool = tf.nn.max_pool(block1_h_conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
|
74
|
-
|
75
|
-
|
76
|
-
|
77
|
-
# to the flat tensor
|
78
|
-
|
79
|
-
flatten = tf.reshape(block1_pool, [-1, 112*112*64])
|
80
|
-
|
81
|
-
|
82
|
-
|
83
|
-
# fully connected layer 1
|
84
|
-
|
85
|
-
w_fc1 = tf.Variable(tf.truncated_normal([112*112*64, 4096], stddev=0.1), name='w_fc1')
|
86
|
-
|
87
|
-
b_fc1 = tf.Variable(tf.constant(0.1, shape=[4096]), name='b_fc1')
|
88
|
-
|
89
|
-
h_fc1 = tf.nn.relu(tf.matmul(flatten, w_fc1) + b_fc1)
|
90
|
-
|
91
|
-
|
92
|
-
|
93
|
-
# fully connected layer 2
|
94
|
-
|
95
|
-
w_fc2 = tf.Variable(tf.truncated_normal([4096, 4096], stddev=0.1), name='w_fc2')
|
96
|
-
|
97
|
-
b_fc2 = tf.Variable(tf.constant(0.1, shape=[4096]), name='b_fc2')
|
98
|
-
|
99
|
-
h_fc2 = tf.nn.relu(tf.matmul(h_fc1, w_fc2) + b_fc2)
|
100
|
-
|
101
|
-
|
102
|
-
|
103
|
-
# output layer
|
104
|
-
|
105
|
-
# class = ['barcode', 'tag'] thus [4096, 2]
|
106
|
-
|
107
|
-
w_output = tf.Variable(tf.truncated_normal([4096, 2], stddev=0.1), name='w_output')
|
108
|
-
|
109
|
-
b_output = tf.Variable(tf.constant(0.1, shape=[2]), name='b_output')
|
110
|
-
|
111
|
-
out = tf.nn.softmax(tf.matmul(h_fc2, w_output) + b_output)
|
112
|
-
|
113
|
-
|
114
|
-
|
115
|
-
# define training data
|
116
|
-
|
117
|
-
# class = ['barcode', 'tag'] thus [None, 2]
|
118
|
-
|
119
|
-
y = tf.placeholder(tf.float32, [None, 2])
|
120
|
-
|
121
|
-
# loss function: cross entropy
|
122
|
-
|
123
|
-
loss = tf.reduce_mean(-tf.reduce_sum(y*tf.log(out + 1e-5), axis=[1]))
|
124
|
-
|
125
|
-
|
126
|
-
|
127
|
-
# training
|
128
|
-
|
129
|
-
train_step = tf.train.GradientDescentOptimizer(0.01).minimize(loss)
|
130
|
-
|
131
|
-
|
132
|
-
|
133
|
-
# evaluation
|
134
|
-
|
135
|
-
correct = tf.equal(tf.argmax(out, 1), tf.argmax(y, 1))
|
136
|
-
|
137
|
-
# tf.reduce_mean: average calculate (ex)[0,1,1] --> 2/3
|
138
|
-
|
139
|
-
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
|
140
|
-
|
141
|
-
|
142
|
-
|
143
|
-
# initialization
|
144
|
-
|
145
|
-
init = tf.global_variables_initializer()
|
146
|
-
|
147
|
-
|
148
|
-
|
149
|
-
|
150
|
-
|
151
|
-
with tf.Session() as sess:
|
152
|
-
|
153
|
-
sess.run(init)
|
154
|
-
|
155
|
-
print('initilize now')
|
156
|
-
|
157
|
-
|
158
|
-
|
159
|
-
# load images: (height, width, channel)
|
160
|
-
|
161
|
-
""" img = [[[R,G,B],[R,G,B],...],
|
162
|
-
|
163
|
-
[[R,G,B],[R,G,B],...],
|
164
|
-
|
165
|
-
[...................],
|
166
|
-
|
167
|
-
[...................],
|
168
|
-
|
169
|
-
[...................]]"""
|
170
|
-
|
171
|
-
files_train = os.listdir('images/train')
|
172
|
-
|
173
|
-
img_train = []
|
174
|
-
|
175
|
-
normalize = np.ones((224,224,3)) * 255.0
|
176
|
-
|
177
|
-
for file in files_train:
|
178
|
-
|
179
|
-
img = plt.imread('images/train/' + file)
|
180
|
-
|
181
|
-
img = cv2.resize(img, (224,224))
|
182
|
-
|
183
|
-
img = img / normalize
|
184
|
-
|
185
|
-
img_train.append(img)
|
186
|
-
|
187
|
-
print('loading train images now')
|
188
|
-
|
189
|
-
|
190
|
-
|
191
|
-
print('loading train csv file now')
|
192
|
-
|
193
|
-
labels_train_csv = csvd.read_csv('data/train_labels.csv')
|
194
|
-
|
195
|
-
# labels = [['IMG_6188.JPG', 'barcode'],...]
|
196
|
-
|
197
|
-
labels_train = csvd.fname_and_label(labels_train_csv)
|
198
|
-
|
199
|
-
labels_one_hot_train = []
|
200
|
-
|
201
|
-
for i in range(len(labels_train)):
|
202
|
-
|
203
|
-
labels_one_hot_train.append([0, 0])
|
204
|
-
|
205
|
-
print('making train one-hot labels of all zero')
|
206
|
-
|
207
|
-
for i in range(len(labels_train)):
|
208
|
-
|
209
|
-
if labels_train[i][1] == 'barcode':
|
210
|
-
|
211
|
-
labels_one_hot_train[i][0] = 1
|
212
|
-
|
213
|
-
if labels_train[i][1] == 'tag':
|
214
|
-
|
215
|
-
labels_one_hot_train[i][1] = 1
|
216
|
-
|
217
|
-
print('making true train one-hot labels')
|
218
|
-
|
219
|
-
|
220
|
-
|
221
|
-
# load images: (height, width, channel)
|
222
|
-
|
223
|
-
""" img = [[[R,G,B],[R,G,B],...],
|
224
|
-
|
225
|
-
[[R,G,B],[R,G,B],...],
|
226
|
-
|
227
|
-
[...................],
|
228
|
-
|
229
|
-
[...................],
|
230
|
-
|
231
|
-
[...................]]"""
|
232
|
-
|
233
|
-
files_test = os.listdir('images/test')
|
234
|
-
|
235
|
-
img_test = []
|
236
|
-
|
237
|
-
for file in files_test:
|
238
|
-
|
239
|
-
img = plt.imread('images/test/' + file)
|
240
|
-
|
241
|
-
img = cv2.resize(img, (224,224))
|
242
|
-
|
243
|
-
img = img / normalize
|
244
|
-
|
245
|
-
img_test.append(img)
|
246
|
-
|
247
|
-
print('loading test images now')
|
248
|
-
|
249
|
-
|
250
|
-
|
251
|
-
print('loading test csv file now')
|
252
|
-
|
253
|
-
labels_test_csv = csvd.read_csv('data/test_labels.csv')
|
254
|
-
|
255
|
-
# labels = [['IMG_6188.JPG', 'barcode'],...]
|
256
|
-
|
257
|
-
labels_test = csvd.fname_and_label(labels_test_csv)
|
258
|
-
|
259
|
-
labels_one_hot_test = []
|
260
|
-
|
261
|
-
for i in range(len(labels_test)):
|
262
|
-
|
263
|
-
labels_one_hot_test.append([0, 0])
|
264
|
-
|
265
|
-
print('making test one-hot labels of all zero')
|
266
|
-
|
267
|
-
for i in range(len(labels_test)):
|
268
|
-
|
269
|
-
if labels_test[i][1] == 'barcode':
|
270
|
-
|
271
|
-
labels_one_hot_test[i][0] = 1
|
272
|
-
|
273
|
-
if labels_test[i][1] == 'tag':
|
274
|
-
|
275
|
-
labels_one_hot_test[i][1] = 1
|
276
|
-
|
277
|
-
print('making true test one-hot labels')
|
278
|
-
|
279
|
-
|
280
|
-
|
281
|
-
for step in range(1000):
|
282
|
-
|
283
|
-
for i in range(len(img_train)):
|
284
|
-
|
285
|
-
sess.run(train_step, feed_dict={img_input:img_train[i], y:labels_one_hot_train[i]})
|
286
|
-
|
287
|
-
print('training now: ', i, ' steps')
|
288
|
-
|
289
|
-
|
290
|
-
|
291
|
-
acc_val = 0
|
292
|
-
|
293
|
-
for i in range(len(img_test)):
|
294
|
-
|
295
|
-
acc_val += sess.run(accuracy, feed_dict={img_input:img_test[i], y:labels_one_hot_test[i]})
|
296
|
-
|
297
|
-
acc_val = acc_val / len(img_test)
|
298
|
-
|
299
|
-
print('step: ', step+1, ' accuracy: ', acc_val)
|
300
|
-
|
301
|
-
```
|
302
14
|
|
303
15
|
実行時のコマンドプロンプトは以下のように表示されたきり動きません。
|
304
16
|
|