質問編集履歴
2
タグの追加
title
CHANGED
|
File without changes
|
body
CHANGED
|
File without changes
|
1
ソースコードの簡略化、実行環境追記
title
CHANGED
|
@@ -1,1 +1,1 @@
|
|
|
1
|
-
CNN-Autoencoderでlossがnanばかり出て、Accuracyも出
|
|
1
|
+
CNN-Autoencoderでlossがnanばかり出て、Accuracyも出ずに困っています。
|
body
CHANGED
|
@@ -1,27 +1,35 @@
|
|
|
1
1
|
失礼します。
|
|
2
|
-
CNN-Autoencoderを用いた画像分類器を作ろうとしていて、folder1内の28*28画像(460枚)を読み込んでCNN-Autoencoderで学習させて、更にfolder1内の画像を用いてテストした後にテスト画像の復元(decoded)を出力しようとしています。以下の様にlossやAccuracyの値が出ません。更には復元画像もほとんど真っ
|
|
2
|
+
CNN-Autoencoderを用いた画像分類器を作ろうとしていて、folder1内の28*28画像(460枚)を読み込んでCNN-Autoencoderで学習させて、更にfolder1内の画像を用いてテストした後にテスト画像の復元(decoded)を出力しようとしています。学習を進めていくと以下の様にlossやAccuracyの値が出ません。更には復元画像もほとんど真っ白な画像しか出ず、明らかに学習できていません。
|
|
3
3
|
```
|
|
4
|
-
step, loss, accuracy = 1: -204.758 0.000
|
|
5
|
-
step, loss, accuracy = 2: -734.519 0.000
|
|
6
|
-
step, loss, accuracy =
|
|
4
|
+
step, loss, accuracy = 1: nan 0.935
|
|
5
|
+
step, loss, accuracy = 2: nan 0.935
|
|
6
|
+
step, loss, accuracy = 3: nan 0.891
|
|
7
|
-
step, loss, accuracy = 4: nan 0.
|
|
7
|
+
step, loss, accuracy = 4: nan 0.891
|
|
8
8
|
step, loss, accuracy = 5: nan 0.913
|
|
9
|
+
step, loss, accuracy = 6: nan 0.978
|
|
9
|
-
step, loss, accuracy =
|
|
10
|
+
step, loss, accuracy = 7: nan 0.913
|
|
11
|
+
step, loss, accuracy = 8: nan 0.935
|
|
10
|
-
step, loss, accuracy =
|
|
12
|
+
step, loss, accuracy = 9: nan 0.848
|
|
11
|
-
step, loss, accuracy = 8: nan 0.848
|
|
12
|
-
step, loss, accuracy =
|
|
13
|
+
step, loss, accuracy = 10: nan 0.913
|
|
14
|
+
step, loss, accuracy = 11: nan 0.935
|
|
13
|
-
step, loss, accuracy =
|
|
15
|
+
step, loss, accuracy = 12: nan 0.870
|
|
16
|
+
step, loss, accuracy = 13: nan 0.848
|
|
17
|
+
step, loss, accuracy = 14: nan 0.913
|
|
18
|
+
step, loss, accuracy = 15: nan 0.935
|
|
19
|
+
step, loss, accuracy = 16: nan 0.935
|
|
20
|
+
step, loss, accuracy = 17: nan 0.913
|
|
21
|
+
step, loss, accuracy = 18: nan 0.913
|
|
22
|
+
step, loss, accuracy = 19: nan 0.913
|
|
23
|
+
step, loss, accuracy = 20: nan 0.826
|
|
24
|
+
loss (test) = nan
|
|
25
|
+
accuracy(test) = 0.9347826
|
|
14
26
|
```
|
|
15
27
|
|
|
16
|
-
以下のコードを使用しています。
|
|
28
|
+
以下のコードを使用しています。
|
|
17
|
-
|
|
29
|
+
コードはこの方 https://qiita.com/TomokIshii/items/26b7414bdb3cd3052786 のコードを使わせて頂きました。ネットワークの定義はより高レベルに書き換えました(tf.nn.conv2dでなくtf.layers.conv2dを使用)。
|
|
18
30
|
|
|
19
31
|
`cnn_ae.py`
|
|
20
32
|
```
|
|
21
|
-
from __future__ import absolute_import
|
|
22
|
-
from __future__ import division
|
|
23
|
-
from __future__ import print_function
|
|
24
|
-
|
|
25
33
|
import numpy as np
|
|
26
34
|
import matplotlib as mpl
|
|
27
35
|
import random
|
|
@@ -40,91 +48,32 @@
|
|
|
40
48
|
two_layer = 7
|
|
41
49
|
|
|
42
50
|
|
|
43
|
-
|
|
44
|
-
# Up-sampling 2-D Layer (deconvolutoinal Layer)
|
|
45
|
-
class Conv2Dtranspose(object):
|
|
46
|
-
'''
|
|
47
|
-
constructor's args:
|
|
48
|
-
input : input image (2D matrix)
|
|
49
|
-
output_siz : output image size
|
|
50
|
-
in_ch : number of incoming image channel
|
|
51
|
-
out_ch : number of outgoing image channel
|
|
52
|
-
patch_siz : filter(patch) size
|
|
53
|
-
'''
|
|
54
|
-
|
|
55
|
-
def __init__(self, input, output_siz, in_ch, out_ch, patch_siz, activation='relu'):
|
|
56
|
-
self.input = input
|
|
57
|
-
self.rows = output_siz[0]
|
|
58
|
-
self.cols = output_siz[1]
|
|
59
|
-
self.out_ch = out_ch
|
|
60
|
-
self.activation = activation
|
|
61
|
-
|
|
62
|
-
wshape = [patch_siz[0], patch_siz[1], out_ch, in_ch] # note the arguments order
|
|
63
|
-
|
|
64
|
-
w_cvt = tf.Variable(tf.truncated_normal(wshape, stddev=0.1),
|
|
65
|
-
trainable=True)
|
|
66
|
-
b_cvt = tf.Variable(tf.constant(0.1, shape=[out_ch]),
|
|
67
|
-
trainable=True)
|
|
68
|
-
self.batsiz = tf.shape(input)[0]
|
|
69
|
-
self.w = w_cvt
|
|
70
|
-
self.b = b_cvt
|
|
71
|
-
self.params = [self.w, self.b]
|
|
72
|
-
|
|
73
|
-
def output(self):
|
|
74
|
-
shape4D = [self.batsiz, self.rows, self.cols, self.out_ch]
|
|
75
|
-
linout = tf.nn.conv2d_transpose(self.input, self.w, output_shape=shape4D,
|
|
76
|
-
strides=[1, 2, 2, 1], padding='SAME') + self.b
|
|
77
|
-
if self.activation == 'relu':
|
|
78
|
-
self.output = tf.nn.relu(linout)
|
|
79
|
-
elif self.activation == 'sigmoid':
|
|
80
|
-
self.output = tf.sigmoid(linout)
|
|
81
|
-
else:
|
|
82
|
-
self.output = linout
|
|
83
|
-
|
|
84
|
-
return self.output
|
|
85
|
-
|
|
86
|
-
|
|
87
51
|
# Create the model
|
|
88
52
|
def mk_nn_model(x):
|
|
89
53
|
# Encoding phase
|
|
90
54
|
x_image = tf.reshape(x, [-1, image_size, image_size, 1])
|
|
91
|
-
conv1 = Convolution2D(x_image, (image_size, image_size), 1, 16,
|
|
92
|
-
(3, 3), activation='relu')
|
|
93
|
-
conv1_out = conv1.output()
|
|
94
55
|
|
|
56
|
+
conv1 = tf.layers.conv2d(inputs=x_image, filters=16, kernel_size=(3, 3), padding='same', activation=tf.nn.relu)
|
|
95
|
-
pool1 =
|
|
57
|
+
pool1 = tf.layers.max_pooling2d(conv1, pool_size=(2, 2), strides=(2, 2), padding='same')
|
|
96
|
-
pool1_out = pool1.output()
|
|
97
58
|
|
|
59
|
+
conv2 = tf.layers.conv2d(inputs=pool1, filters=8, kernel_size=(3, 3), padding='same', activation=tf.nn.relu)
|
|
98
|
-
|
|
60
|
+
pool2 = tf.layers.max_pooling2d(conv2, pool_size=(2, 2), strides=(2, 2), padding='same')
|
|
99
|
-
(3, 3), activation='relu')
|
|
100
|
-
conv2_out = conv2.output()
|
|
101
61
|
|
|
102
|
-
|
|
62
|
+
conv3 = tf.layers.conv2d(pool2, filters=8, kernel_size=(3, 3), padding='same', activation=tf.nn.relu)
|
|
103
|
-
pool2_out = pool2.output()
|
|
104
|
-
|
|
105
|
-
|
|
63
|
+
encoded = tf.layers.max_pooling2d(conv3, pool_size=(2, 2), strides=(2, 2), padding='same')
|
|
106
|
-
conv3_out = conv3.output()
|
|
107
|
-
|
|
108
|
-
pool3 = MaxPooling2D(conv3_out)
|
|
109
|
-
pool3_out = pool3.output()
|
|
110
64
|
# at this point the representation is (8, 4, 4) i.e. 128-dimensional
|
|
111
65
|
# Decoding phase
|
|
112
|
-
|
|
66
|
+
upsample1 = tf.image.resize_images(encoded, size=(two_layer, two_layer), method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
|
|
113
|
-
|
|
67
|
+
conv4 = tf.layers.conv2d(inputs=upsample1, filters=8, kernel_size=(3, 3), padding='same', activation=tf.nn.relu)
|
|
114
|
-
conv_t1_out = conv_t1.output()
|
|
115
68
|
|
|
116
|
-
|
|
69
|
+
upsample2 = tf.image.resize_images(conv4, size=(one_layer, one_layer), method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
|
|
117
|
-
|
|
70
|
+
conv5 = tf.layers.conv2d(inputs=upsample2, filters=8, kernel_size=(3, 3), padding='same', activation=tf.nn.relu)
|
|
118
|
-
conv_t2_out = conv_t2.output()
|
|
119
71
|
|
|
120
|
-
|
|
72
|
+
upsample3 = tf.image.resize_images(conv5, size=(image_size, image_size), method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
|
|
121
|
-
|
|
73
|
+
conv6 = tf.layers.conv2d(inputs=upsample3, filters=16, kernel_size=(3, 3), padding='same', activation=tf.nn.relu)
|
|
122
|
-
conv_t3_out = conv_t3.output()
|
|
123
74
|
|
|
124
|
-
|
|
75
|
+
logits = tf.layers.conv2d(inputs=conv6, filters=1, kernel_size=(3, 3), padding='same', activation=None)
|
|
125
|
-
activation='sigmoid')
|
|
126
|
-
decoded =
|
|
76
|
+
decoded = tf.nn.sigmoid(logits)
|
|
127
|
-
|
|
128
77
|
decoded = tf.reshape(decoded, [-1, image_size*image_size])
|
|
129
78
|
|
|
130
79
|
cross_entropy = -1. * x * tf.log(decoded) - (1. - x) * tf.log(1. - decoded)
|
|
@@ -154,7 +103,7 @@
|
|
|
154
103
|
loss, decoded, accuracy = mk_nn_model(x)
|
|
155
104
|
train_step = tf.train.AdagradOptimizer(learning_rate).minimize(loss)
|
|
156
105
|
|
|
157
|
-
epochs =
|
|
106
|
+
epochs = 20
|
|
158
107
|
batch_size = data_number // 10
|
|
159
108
|
|
|
160
109
|
init = tf.initialize_all_variables()
|
|
@@ -172,9 +121,9 @@
|
|
|
172
121
|
# generate decoded image with test data
|
|
173
122
|
X = np.random.permutation(fileset)
|
|
174
123
|
test_image = X[0:batch_size]
|
|
175
|
-
decoded_imgs =
|
|
124
|
+
decoded_imgs, test_loss, test_accuracy = sess.run([decoded, loss, accuracy], feed_dict={x: test_image})
|
|
176
|
-
print('loss (test) = ',
|
|
125
|
+
print('loss (test) = ', test_loss)
|
|
177
|
-
print('accuracy(test) = ',
|
|
126
|
+
print('accuracy(test) = ', test_accuracy)
|
|
178
127
|
|
|
179
128
|
x_test = test_image
|
|
180
129
|
n = 10 # how many digits we will display
|
|
@@ -198,6 +147,11 @@
|
|
|
198
147
|
plt.savefig('images.png')
|
|
199
148
|
|
|
200
149
|
```
|
|
150
|
+
####実行環境
|
|
151
|
+
OS Windows10 64bit
|
|
152
|
+
プロセッサ Intel(R) Core(TM)i7-8700k CPU @ 3.70GHz
|
|
153
|
+
RAM 32.0 GB
|
|
154
|
+
Anaconda Prompt
|
|
201
155
|
|
|
202
156
|
####試したこと
|
|
203
157
|
・decoded直前でsigmoidをsoftmaxにした。その結果、nanばかり表示されてしまい、全く改善せず。
|