keras写一个Focal Loss损失函数

A：

import tensorflow.keras.backend as K

def FocalLoss(gamma=2., alpha=.25):
    def focal_loss(y_true, y_pred):
        # calculate cross entropy
        ce_loss = K.categorical_crossentropy(y_true, y_pred, from_logits=False)
        # calculate exponential focal loss
        alpha_factor = K.ones_like(y_true) * alpha
        alpha_factor = K.tf.where(K.equal(y_true, 1), alpha_factor, 1 - alpha_factor)
        focal_weight = K.tf.where(K.equal(y_true, 1), 1 - y_pred, y_pred)
        focal_weight = alpha_factor * focal_weight ** gamma
        # calculate final focal loss
        focal_loss = focal_weight * ce_loss
        return K.mean(focal_loss, axis=-1)
    return focal_loss

在使用时，可以按以下方式将损失函数添加到模型中：

model.compile(loss=FocalLoss(gamma=2., alpha=.25), optimizer='adam')