def compute_mask(self, inputs, mask=None):
"""
保证第一个token不被mask
"""
mask = super(Embedding, self).compute_mask(inputs, mask)
if mask is not None:
mask1 = K.ones_like(mask[:, :1], dtype='bool')
mask2 = mask[:, 1:]
return K.concatenate([mask1, mask2], 1)
def dense_loss(self, y_true, y_pred):
"""y_true需要是one hot形式
"""
# 导出mask并转换数据类型
if self.input_mask is None:
mask = None
else:
mask = K.cast(self.input_mask, K.floatx())
# 计算目标分数
target_score = self.target_score(y_true, y_pred, mask)
# 递归计算log Z
init_states = [y_pred[:, 0]]
if mask is None:
mask = K.ones_like(y_pred[:, :, :1])
else:
mask = K.expand_dims(mask, 2)
y_pred = K.concatenate([y_pred, mask])
log_norm, _, _ = K.rnn(self.log_norm_step,
y_pred[:, 1:],
init_states) # 最后一步的log Z向量
log_norm = tf.reduce_logsumexp(log_norm, 1) # logsumexp得标量
# 计算损失 -log p
return log_norm - target_score