def train_function(inputs): # 重新定义训练函数
grads = embedding_gradients(inputs)[0] # Embedding梯度
delta = epsilon * grads / (np.sqrt((grads**2).sum()) + 1e-8) # 计算扰动
K.set_value(embeddings, K.eval(embeddings) + delta) # 注入扰动
outputs = old_train_function(inputs) # 梯度下降
K.set_value(embeddings, K.eval(embeddings) - delta) # 删除扰动
return outputs
def build(self, input_shape):
output_dim = input_shape[-1]
if not isinstance(output_dim, int):
output_dim = output_dim.value
self.trans = self.add_weight(name='trans',
shape=(output_dim, output_dim),
initializer='glorot_uniform',
trainable=True)
if self.lr_multiplier != 1:
K.set_value(self.trans, K.eval(self.trans) / self.lr_multiplier)
self.trans = self.lr_multiplier * self.trans