https://github.com/busyyang/keras_centerloss

https://github.com/busyyang/keras_centerloss

Last synced: about 1 month ago
JSON representation

• Host: GitHub
• URL: https://github.com/busyyang/keras_centerloss
• Owner: busyyang
• Created: 2020-12-28T06:33:47.000Z (about 4 years ago)
• Default Branch: main
• Last Pushed: 2020-12-28T06:35:40.000Z (about 4 years ago)
• Last Synced: 2023-10-19T21:08:16.885Z (about 1 year ago)
• Language: Python
• Size: 8.79 KB
• Stars: 1
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        

## 中心损失函数CenterLoss的实现

在Keras中可通过继承`Layer`的类自定义一个中心损失函数层，在这个新建的层中只需要重写`build`,`call`,`compute_output_shape`函数，其中在`build`中添加一个centers的权重，用于存放中心点，其大小为`num_class x num_features`,即模型输出数x特征数。由于centers不是由梯度来学习的，所以需要设置`trainable=False`。

~~~py

class CenterLossLayer(Layer):

    def __init__(self, num_class, num_features, alpha=0.5, **kwargs):

        # num_class为模型总的分类数量

        # num_features为模型倒数第二层的节点数，即为特征点数

        # alpha为centers的学习率，避免个别标注错误的点对中心点的较大扰动

        super().__init__(**kwargs)

        self.alpha = alpha

        self.num_class = num_class

        self.num_features = num_features

    def build(self, input_shape):

        self.centers = self.add_weight(name='centers',

                                       shape=(self.num_class, self.num_features),

                                       initializer='uniform', trainable=False)

        super().build(input_shape)

    def call(self, x, mask=None):

        """

        计算中心损失

        :param x: 为centerLossLayer层的输入，为2个tensor,

                    x[0]为模型的特征向量(batch_size x num_features)，

                    x[1]为label标签，非onehot(batch_size x 1)

        :param mask: None

        :return:

        """

        # 将x[1]转化为batch_size长度的向量，并转化为int32数据类型

        labels = tf.reshape(x[1], [-1])

        labels = tf.cast(labels, 'int32')

        # 获取当前batch下涉及到可能需要更新的centers

        center_batch = tf.gather(self.centers, labels)

        # 计算当前batch下中心更新量

        diff = center_batch - x[0]

        # 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)

        # 更新centers: 用减法的形式更新，需要更新的是self.centers, 更新的index是labels表示，更新量为self.alpha * diff

        self.centers = tf.scatter_sub(self.centers, labels, self.alpha * diff)

        # 计算当前batch下的centerloss

        self.result = x[0] - K.dot(tf.one_hot(tf.cast(x[1], 'int32'), self.num_class), self.centers)

        self.result = 0.5 * K.sum(self.result ** 2, axis=1, keepdims=True)

        return self.result

    def compute_output_shape(self, input_shape):

        return K.int_shape(self.result)

~~~

由于在CenterLossLayer.call中已经返回了centerloss，所以在这部分的损失函数中直接返回中心损失层的输出即可。即需要定义一个损失函数为：

~~~py

def zero_loss(y_true, y_pred):

    return y_pred

~~~

#### issue

如果需要获取centers,按理说只需要`model.get_layer('centerlosslayer').getweights()[0]`即可，但是会报一个需要feed a value for placeholder 'input_2'的错误，找了很久没有找到为什么。 

#### 参考

1. [A Discriminative Feature Learning Approach for Deep Face Recognition](https://ydwen.github.io/papers/WenECCV16.pdf)

2. [handongfeng/MNIST-center-loss](https://github.com/handongfeng/MNIST-center-loss/blob/master/centerLoss_MNIST.py)

3. [在Keras使用center-losss损失函数\Keras自定义损失函数](https://blog.csdn.net/DumpDoctorWang/article/details/84204476)