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 sparse_categorical_crossentropy(y_true, y_pred):
"""自定义稀疏交叉熵
这主要是因为keras自带的sparse_categorical_crossentropy不支持求二阶梯度。
"""
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)[-1])
return K.categorical_crossentropy(y_true, y_pred)
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
def compute_loss(self, inputs, mask=None):
q_start_in, q_end_in, q_label_in, ps_category, ps_heads, ps_tails = inputs
if mask is None:
mask = 1.0
else:
mask = K.cast(mask, K.floatx())
loss0 = K.sparse_categorical_crossentropy(q_label_in,
ps_category,
from_logits=True)
loss0 = K.mean(loss0)
loss0 = K.sum(loss0 * mask) / K.sum(mask)
loss1 = K.categorical_crossentropy(q_start_in,
ps_heads,
from_logits=True)
loss1 = K.mean(loss1)
ps_tails = ps_tails - (1 - K.cumsum(q_start_in, axis=1)) * 1e10
loss2 = K.mean(
K.categorical_crossentropy(q_end_in, ps_tails, from_logits=True))
loss2 = K.mean(loss2)
# 总的loss
return loss0 + loss1 + loss2
