回答編集履歴
3
d
test
CHANGED
|
@@ -88,17 +88,17 @@
|
|
|
88
88
|
|
|
89
89
|
|
|
90
90
|
|
|
91
|
-
num_samples = 30
|
|
91
|
+
num_samples = 30 # サンプル数
|
|
92
92
|
|
|
93
|
-
num_classes = 3
|
|
93
|
+
num_classes = 3 # クラス数
|
|
94
94
|
|
|
95
|
-
num_features = 10
|
|
95
|
+
num_features = 10 # 特徴量の次元
|
|
96
96
|
|
|
97
97
|
|
|
98
98
|
|
|
99
99
|
data = np.random.randn(num_samples, num_features)
|
|
100
100
|
|
|
101
|
-
labels = np.random.randint(0,
|
|
101
|
+
labels = np.random.randint(0, num_classes, num_samples)
|
|
102
102
|
|
|
103
103
|
|
|
104
104
|
|
2
d
test
CHANGED
|
@@ -73,3 +73,87 @@
|
|
|
73
73
|
print(np.allclose(tf_ret, np_ret)) # True
|
|
74
74
|
|
|
75
75
|
```
|
|
76
|
+
|
|
77
|
+
|
|
78
|
+
|
|
79
|
+
## 追記
|
|
80
|
+
|
|
81
|
+
|
|
82
|
+
|
|
83
|
+
```python
|
|
84
|
+
|
|
85
|
+
import numpy as np
|
|
86
|
+
|
|
87
|
+
import tensorflow as tf
|
|
88
|
+
|
|
89
|
+
|
|
90
|
+
|
|
91
|
+
num_samples = 30
|
|
92
|
+
|
|
93
|
+
num_classes = 3
|
|
94
|
+
|
|
95
|
+
num_features = 10
|
|
96
|
+
|
|
97
|
+
|
|
98
|
+
|
|
99
|
+
data = np.random.randn(num_samples, num_features)
|
|
100
|
+
|
|
101
|
+
labels = np.random.randint(0, 10, num_samples)
|
|
102
|
+
|
|
103
|
+
|
|
104
|
+
|
|
105
|
+
# TensorFlow
|
|
106
|
+
|
|
107
|
+
#######################################
|
|
108
|
+
|
|
109
|
+
x = tf.placeholder(tf.float32, shape=(None, num_features))
|
|
110
|
+
|
|
111
|
+
y = tf.placeholder(tf.int64, shape=(None,))
|
|
112
|
+
|
|
113
|
+
|
|
114
|
+
|
|
115
|
+
means = []
|
|
116
|
+
|
|
117
|
+
for class_id in range(num_classes):
|
|
118
|
+
|
|
119
|
+
mean = tf.reduce_mean(tf.boolean_mask(x, tf.equal(y, class_id)), axis=0)
|
|
120
|
+
|
|
121
|
+
means.append(mean)
|
|
122
|
+
|
|
123
|
+
concat = tf.stack(means)
|
|
124
|
+
|
|
125
|
+
|
|
126
|
+
|
|
127
|
+
with tf.Session() as sess:
|
|
128
|
+
|
|
129
|
+
tf_ret = sess.run(concat, feed_dict={x: data, y: labels})
|
|
130
|
+
|
|
131
|
+
print(tf_ret.shape, tf_ret) # (3, 10)
|
|
132
|
+
|
|
133
|
+
|
|
134
|
+
|
|
135
|
+
# numpy
|
|
136
|
+
|
|
137
|
+
#######################################
|
|
138
|
+
|
|
139
|
+
means = []
|
|
140
|
+
|
|
141
|
+
for class_id in range(num_classes):
|
|
142
|
+
|
|
143
|
+
mean = np.mean(data[labels==class_id], axis=0)
|
|
144
|
+
|
|
145
|
+
means.append(mean)
|
|
146
|
+
|
|
147
|
+
np_ret = np.stack(means)
|
|
148
|
+
|
|
149
|
+
|
|
150
|
+
|
|
151
|
+
# TensorFlow と numpy の計算が一致するか
|
|
152
|
+
|
|
153
|
+
print(np.allclose(tf_ret, np_ret)) # True
|
|
154
|
+
|
|
155
|
+
```
|
|
156
|
+
|
|
157
|
+
|
|
158
|
+
|
|
159
|
+
各クラスごとに平均を計算するならこのような感じでしょうか？
|
1
d
test
CHANGED
|
@@ -46,7 +46,7 @@
|
|
|
46
46
|
|
|
47
47
|
tf_ret = sess.run(concat, feed_dict={input1: a, input2: b, input3: c})
|
|
48
48
|
|
|
49
|
-
print(tf_ret.shape, tf_ret)
|
|
49
|
+
print(tf_ret.shape, tf_ret) # (3, 10)
|
|
50
50
|
|
|
51
51
|
|
|
52
52
|
|
|
@@ -64,7 +64,7 @@
|
|
|
64
64
|
|
|
65
65
|
|
|
66
66
|
|
|
67
|
-
print(np_ret, np_ret.shape)
|
|
67
|
+
print(np_ret, np_ret.shape) # (3, 10)
|
|
68
68
|
|
|
69
69
|
|
|
70
70
|
|