def sparse_categorical_crossentropy(y_true, y_pred):
# y_true需要重新明确一下shape和dtype
y_true = K.reshape(y_true, K.shape(y_pred)[:-1])
y_true = K.cast(y_true, 'int32')
y_true = K.one_hot(y_true, K.shape(y_pred)[2])
# 计算交叉熵
return K.mean(K.categorical_crossentropy(y_true, y_pred))
def compute_loss_of_similarity(self, inputs):
y_true = self.get_labels_of_similarity(inputs) # 构建标签
y_pred = K.l2_normalize(inputs, axis=1) # 句向量归一化
similarities = K.dot(y_pred, K.transpose(y_pred)) # 相似度矩阵
similarities = similarities - K.eye(K.shape(y_pred)[0]) * 1e12 # 排除对角线
similarities = similarities * self.scale # scale
loss = K.categorical_crossentropy(
y_true, similarities, from_logits=True
)
return loss
def compute_loss_of_similarity(self, inputs, mask=None):
# _, _, _, y_pred, _ = inputs
_, _, _, _, y_pred = inputs # use last layer's logits
y_true = self.get_labels_of_similarity(y_pred) # 构建标签
y_pred = K.l2_normalize(y_pred, axis=1) # 句向量归一化
similarities = K.dot(y_pred, K.transpose(y_pred)) # 相似度矩阵
similarities = similarities - K.eye(K.shape(y_pred)[0]) * 1e12 # 排除对角线
similarities = similarities * 20 # scale
loss = K.categorical_crossentropy(
y_true, similarities, from_logits=True
)
self.add_metric(loss, 'sim_loss')
return loss
def cal_ce(self, inputs):
source, target = inputs
source_t = K.softmax(source / self.temperature)
target_t = K.softmax(target / self.temperature)
ce = K.categorical_crossentropy(source_t, target_t)
return ce