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 compute_loss_of_scl(self, inputs, mask=None):
y_pred, y_true = inputs
label_mask = self.get_label_mask(y_true)
y_pred = K.l2_normalize(y_pred, axis=1) # 特征向量归一化
similarities = K.dot(y_pred, K.transpose(y_pred)) # 相似矩阵
similarities = similarities - K.eye(K.shape(y_pred)[0]) * 1e12 # 排除对角线,即 i == j
similarities = similarities / self.T # Temperature scale
similarities = K.exp(similarities) # exp
sum_similarities = K.sum(similarities, axis=-1, keepdims=True) # sum i != k
scl = similarities / sum_similarities
scl = K.log(scl + K.epsilon()) # sum log
scl = -K.sum(scl * label_mask, axis=1, keepdims=True) / (K.sum(label_mask, axis=1, keepdims=True) + K.epsilon())
return K.mean(scl)
