質問編集履歴
2
コメント追記。
test
CHANGED
|
File without changes
|
test
CHANGED
|
@@ -13,6 +13,10 @@
|
|
|
13
13
|
（少数点4桁くらい表示していますが、まったく変化せず）。
|
|
14
14
|
|
|
15
15
|
計算グラフがうまくできておらず、誤差逆伝播ができていないのかと推察しています。
|
|
16
|
+
|
|
17
|
+
|
|
18
|
+
|
|
19
|
+
以下、インデントが上手くされておらず、読みにくいコードになっておりすみません。
|
|
16
20
|
|
|
17
21
|
|
|
18
22
|
|
1
#つきのコメントを削除
test
CHANGED
|
File without changes
|
test
CHANGED
|
@@ -16,17 +16,11 @@
|
|
|
16
16
|
|
|
17
17
|
|
|
18
18
|
|
|
19
|
-
# ResNet
|
|
20
|
-
|
|
21
|
-
def _shortcut(inputs, residual): #inputs: ResBlockへの入力そのもの（ショートカットしたデータ）, residual: ResBlockの中で処理されたデータ
|
|
22
|
-
|
|
23
|
-
|
|
19
|
+
def _shortcut(inputs, residual)
|
|
24
20
|
|
|
25
21
|
n_filters = residual.shape[3]
|
|
26
22
|
|
|
27
23
|
print('n_filters=',n_filters)
|
|
28
|
-
|
|
29
|
-
# inputsのチャンネル数をresidualのチャンネル数に合わせる（1x1畳み込み）
|
|
30
24
|
|
|
31
25
|
shortcut = tf.layers.Conv2D(n_filters, kernel_size=[1, 1], strides=[1, 1], padding='VALID')(inputs)
|
|
32
26
|
|
|
@@ -58,8 +52,6 @@
|
|
|
58
52
|
|
|
59
53
|
|
|
60
54
|
|
|
61
|
-
### ネットワーク ###
|
|
62
|
-
|
|
63
55
|
tf.reset_default_graph()
|
|
64
56
|
|
|
65
57
|
is_training = tf.placeholder(tf.bool, shape=())
|
|
@@ -72,49 +64,51 @@
|
|
|
72
64
|
|
|
73
65
|
|
|
74
66
|
|
|
75
|
-
### ネットワーク本体 ###
|
|
76
|
-
|
|
77
67
|
h = tf.layers.Conv2D(filters=32, kernel_size=[7, 7], strides=[1, 1], kernel_initializer=initializer, padding='SAME')(x)
|
|
78
68
|
|
|
79
69
|
h = tf.layers.BatchNormalization()(h, training=is_training)
|
|
80
70
|
|
|
81
71
|
h = tf.nn.relu(h)
|
|
82
72
|
|
|
83
|
-
h = tf.layers.MaxPooling2D(pool_size=[2, 2], strides=2)(h)
|
|
73
|
+
h = tf.layers.MaxPooling2D(pool_size=[2, 2], strides=2)(h)
|
|
84
74
|
|
|
85
75
|
|
|
86
|
-
|
|
87
|
-
h = _resblock(n_filters=64)(h) # 16x16x32 -> 16x16x64
|
|
88
76
|
|
|
89
77
|
h = _resblock(n_filters=64)(h)
|
|
90
78
|
|
|
91
79
|
h = _resblock(n_filters=64)(h)
|
|
92
80
|
|
|
93
|
-
h =
|
|
81
|
+
h = _resblock(n_filters=64)(h)
|
|
94
82
|
|
|
95
|
-
h =
|
|
83
|
+
h = tf.layers.MaxPooling2D(pool_size=[2, 2], strides=2)(h)
|
|
84
|
+
|
|
85
|
+
|
|
96
86
|
|
|
97
87
|
h = _resblock(n_filters=128)(h)
|
|
98
88
|
|
|
99
89
|
h = _resblock(n_filters=128)(h)
|
|
100
90
|
|
|
101
|
-
h =
|
|
91
|
+
h = _resblock(n_filters=128)(h)
|
|
102
92
|
|
|
93
|
+
h = tf.layers.MaxPooling2D(pool_size=[2, 2], strides=2)(h)
|
|
94
|
+
|
|
95
|
+
|
|
96
|
+
|
|
103
|
-
h = _resblock(n_filters=256)(h)
|
|
97
|
+
h = _resblock(n_filters=256)(h)
|
|
104
98
|
|
|
105
99
|
h = _resblock(n_filters=256)(h)
|
|
106
100
|
|
|
107
101
|
h = _resblock(n_filters=256)(h)
|
|
108
102
|
|
|
109
|
-
# Global Average Pooling
|
|
110
103
|
|
|
111
|
-
h = tf.keras.layers.GlobalAveragePooling2D()(h) # 4x4x256 -> 256
|
|
112
104
|
|
|
113
|
-
|
|
105
|
+
h = tf.keras.layers.GlobalAveragePooling2D()(h)
|
|
106
|
+
|
|
107
|
+
|
|
114
108
|
|
|
115
109
|
h = tf.layers.Dense(units=128, activation=tf.nn.relu)(h)
|
|
116
110
|
|
|
117
|
-
y = tf.layers.Dense(units=10, activation=tf.nn.softmax)(h)
|
|
111
|
+
y = tf.layers.Dense(units=10, activation=tf.nn.softmax)(h)
|
|
118
112
|
|
|
119
113
|
|
|
120
114
|
|
|
@@ -126,9 +120,7 @@
|
|
|
126
120
|
|
|
127
121
|
with tf.control_dependencies(update_ops):
|
|
128
122
|
|
|
129
|
-
|
|
123
|
+
optimizer = tf.train.AdamOptimizer(0.01).minimize(cost)
|
|
130
|
-
|
|
131
|
-
optimizer = tf.train.GradientDecentOptimizer(0.01).minimize(cost)
|
|
132
124
|
|
|
133
125
|
|
|
134
126
|
|