def __init__(self, optimizer, **kwargs):
super(OptimizerWrapper, self).__init__(**kwargs)
self.optimizer = optimizer
self._optimizer_attributes = []
for k, v in get_all_attributes(self.optimizer).items():
if k not in dir(self):
setattr(self, k, v)
self._optimizer_attributes.append(k)
#! -*- coding: utf-8 -*-
# 自定义层
import tensorflow as tf
from bert4keras.backend import keras, K, get_all_attributes
# 等价于 from keras.layers import *
locals().update(get_all_attributes(keras.layers))
initializers = keras.initializers
activations = keras.activations
def sequence_masking(x, mask, mode=0, axis=None, heads=1):
"""为序列条件mask的函数
mask: 形如(batch_size, seq_len)的0-1矩阵；
mode: 如果是0，则直接乘以mask；
如果是1，则在padding部分减去一个大正数。
axis: 序列所在轴，默认为1；
heads: 相当于batch这一维要被重复的次数。
"""
if mask is None or mode not in [0, 1]:
return x
else:
if heads is not 1:
mask = K.expand_dims(mask, 1)
mask = K.tile(mask, (1, heads, 1))
mask = K.reshape(mask, (-1, K.shape(mask)[2]))
if axis is None:
axis = 1
if axis == -1: