损失函数Center Loss 代码解析

center loss来自ECCV2016的一篇论文：A Discriminative Feature Learning Approach for Deep Face Recognition。 
论文链接：http://ydwen.github.io/papers/WenECCV16.pdf 
代码链接：https://github.com/davidsandberg/facenet

理论解析请参看 https://blog.csdn.net/u014380165/article/details/76946339

下面给出centerloss的计算公式以及更新公式

下面的代码是facenet作者利用tensorflow实现的centerloss代码

def center_loss(features, label, alfa, nrof_classes):
    """Center loss based on the paper "A Discriminative Feature Learning Approach for Deep Face Recognition"
       (http://ydwen.github.io/papers/WenECCV16.pdf)
       https://blog.csdn.net/u014380165/article/details/76946339
    """
    nrof_features = features.get_shape()[1]　　#训练过程中，需要保存当前所有类中心的全连接预测特征centers， 每个batch的计算都要先读取已经保存的centers
    centers = tf.get_variable(‘centers‘, [nrof_classes, nrof_features], dtype=tf.float32,
        initializer=tf.constant_initializer(0), trainable=False)
    label = tf.reshape(label, [-1])
    centers_batch = tf.gather(centers, label)#获取当前batch对应的类中心特征
    diff = (1 - alfa) * (centers_batch - features)#计算当前的类中心与特征的差异，用于Cj的的梯度更新
    centers = tf.scatter_sub(centers, label, diff)#更新梯度Cj，对于上图中步骤6，tensorflow会将该变量centers保留下来，用于计算下一个batch的centerloss
    loss = tf.reduce_mean(tf.square(features - centers_batch))#计算当前的centerloss 对应于Lc
    return loss, centers

原文地址：https://www.cnblogs.com/adong7639/p/9090421.html