def compute_loss_of_similarity(self, inputs, mask=None):
_, _, y_pred, _ = inputs
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 * 30 # 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_true = self.get_labels_of_similarity(
y_pred) # 构建标签 (btz,btz) 左右两个btz互为true
y_pred = K.l2_normalize(y_pred, axis=1) # 句向量归一化 (?, 768)
similarities = K.dot(y_pred, K.transpose(y_pred)) # 相似度矩阵 (btz,btz)
similarities = similarities - K.eye(K.shape(
y_pred)[0]) * 1e12 # 排除对角线，因为对角线的是自己跟自己比 (btz,btz) 对角线上的值会变得无穷小
similarities = similarities * 30 # scale (btz,btz)
loss = K.categorical_crossentropy(
y_true, similarities, from_logits=True) # (?,) 由此可以计算一个btz内的句子相似度
return loss
