-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathcenter_loss.py
118 lines (99 loc) · 5.01 KB
/
center_loss.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Sun Sep 16 17:14:49 2018
@author: wsw
"""
# center loss
import tensorflow as tf
def get_center_loss(features,labels,alpha=0.5,num_classes=10):
'''
获取center loss及center的更新op
Arguments:
features: Tensor,表征样本特征,一般使用某个fc层的输出,shape应该为[batch_size, feature_length].
labels: Tensor,表征样本label,非one-hot编码,shape应为[batch_size].
alpha: 0-1之间的数字,控制样本类别中心的学习率,细节参考原文.
num_classes: 整数,表明总共有多少个类别,网络分类输出有多少个神经元这里就取多少.
Return：
loss: Tensor,可与softmax loss相加作为总的loss进行优化.
centers: Tensor,存储样本中心值的Tensor，仅查看样本中心存储的具体数值时有用.
centers_update_op: op,用于更新样本中心的op，在训练时需要同时运行该op，否则样本中心不会更新
'''
# get feature dimension
len_features = features.get_shape()[1]
# initailizer class center
# 设置trainable=False是因为样本中心不是由梯度进行更新的
centers = tf.get_variable('centers',
[num_classes,len_features],
dtype=tf.float32,
initializer=tf.constant_initializer(0),
trainable=False)
# get mini-batch center
# batchxfeature_len
# 根据样本label,获取mini-batch中每一个样本对应的中心值
labels = tf.reshape(labels, [-1])
centers_batch = tf.gather(centers,labels)
# compute center loss
loss = tf.reduce_mean(tf.reduce_sum(tf.squared_difference(features,centers_batch),axis=-1))
# loss = tf.nn.l2_loss(features-centers_batch)
# update center vector
# 当前mini-batch的特征值与它们对应的中心值之间的差
diff = centers_batch - features
# count example nums for each class
# 获取mini-batch中同一类别样本出现的次数,了解原理请参考原文公式(4)
unique_label,idx,count = tf.unique_with_counts(labels)
# get every instance appear times
appear_times = tf.gather(count,idx)
appear_times = tf.reshape(appear_times,shape=[-1,1])
# delta c
diff = diff/tf.cast((1+appear_times),dtype=tf.float32)
diff = alpha*diff
# update center
# centers[labels[i]] - delta_diff[i]
center_update_op = tf.scatter_sub(centers,labels,diff)
return loss,centers,center_update_op
'''
def get_center_loss(features, labels, alpha=0.5, num_classes=10):
"""获取center loss及center的更新op
Arguments:
features: Tensor,表征样本特征,一般使用某个fc层的输出,shape应该为[batch_size, feature_length].
labels: Tensor,表征样本label,非one-hot编码,shape应为[batch_size].
alpha: 0-1之间的数字,控制样本类别中心的学习率,细节参考原文.
num_classes: 整数,表明总共有多少个类别,网络分类输出有多少个神经元这里就取多少.
Return：
loss: Tensor,可与softmax loss相加作为总的loss进行优化.
centers: Tensor,存储样本中心值的Tensor，仅查看样本中心存储的具体数值时有用.
centers_update_op: op,用于更新样本中心的op，在训练时需要同时运行该op，否则样本中心不会更新
"""
# 获取特征的维数，例如256维
# print features.get_shape()
len_features = features.get_shape()[1]
# 建立一个Variable,shape为[num_classes, len_features]，用于存储整个网络的样本中心，
# 设置trainable=False是因为样本中心不是由梯度进行更新的
centers = tf.get_variable('centers',
[num_classes, len_features],
dtype=tf.float32,
initializer=tf.constant_initializer(0),
trainable=False)
# 将label展开为一维的，输入如果已经是一维的，则该动作其实无必要
labels = tf.reshape(labels, [-1])
# 根据样本label,获取mini-batch中每一个样本对应的中心值
centers_batch = tf.gather(centers, labels)
# 计算loss
# loss = tf.nn.l2_loss(features - centers_batch)
loss = tf.reduce_mean(tf.reduce_sum(tf.squared_difference(features,centers_batch),axis=-1))
# 当前mini-batch的特征值与它们对应的中心值之间的差
diff = centers_batch - features
# 获取mini-batch中同一类别样本出现的次数,了解原理请参考原文公式(4)
unique_label, unique_idx, unique_count = tf.unique_with_counts(labels)
appear_times = tf.gather(unique_count, unique_idx)
appear_times = tf.reshape(appear_times, [-1, 1])
diff = diff / tf.cast((1 + appear_times), tf.float32)
diff = alpha * diff
centers_update_op = tf.scatter_sub(centers, labels, diff)
return loss, centers, centers_update_op
'''
if __name__ == '__main__':
xs = tf.random_normal(shape=[128,2])
ys = tf.ones(shape=128,dtype=tf.int64)
loss,center,op = get_center_loss(xs,ys)