def log_norm_step(self, inputs, states):
"""递归求解归一化因子"""
inputs, mask = inputs[:, :-1], inputs[:, -1:]
states = K.expand_dims(states[0], 2) # batch_size,output_dim, 1
trans = K.expand_dims(self.trans, 0) # 1, output_dim, output_dim
outputs = K.logsumexp(states + trans, 1)
outputs += inputs
outputs = mask * outputs + (1 - mask) * states[:, :, 0]
return outputs, [outputs]
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
mask = K.all(K.greater(y_pred, -1e6), axis=2, keepdims=True)
mask = K.cast(mask, K.floatx())
# 计算目标分数
y_true, y_pred = y_true * mask, y_pred * mask
target_score = self.path_score(y_pred, y_true)
# 递归计算log Z
init_states = [y_pred[:, 0]]
y_pred = K.concatenate([y_pred, mask], axis=2)
input_length = K.int_shape(y_pred[:, 1:])[1]
log_norm, _, _ = K.rnn(self.log_norm_step,
y_pred[:, 1:],
init_states,
input_length=input_length) # 最后一步的log Z向量
log_norm = K.logsumexp(log_norm, 1) # logsumexp得标量
# 计算损失 -log p
return log_norm - target_score