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()[]
#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(), trainable=False)
label = tf.reshape(label, [-])
centers_batch = tf.gather(centers, label)#获取当前batch对应的类中心特征
diff = ( - 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,这里facenet的作者做了一个 1-alfa操作,比较奇怪,和原论文不同
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