def build(self, input_shape):
output_dim = input_shape[-1]
if not isinstance(output_dim, int):
output_dim = output_dim.value
if self.hidden_dim is None:
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
else:
self.l_trans = self.add_weight(name='l_trans',
shape=(output_dim, self.hidden_dim),
initializer='glorot_uniform',
trainable=True)
self.r_trans = self.add_weight(name='r_trans',
shape=(output_dim, self.hidden_dim),
initializer='glorot_uniform',
trainable=True)
if self.lr_multiplier != 1:
K.set_value(self.l_trans,
K.eval(self.l_trans) / self.lr_multiplier)
self.l_trans = self.lr_multiplier * self.l_trans
K.set_value(self.r_trans,
K.eval(self.r_trans) / self.lr_multiplier)
self.r_trans = self.lr_multiplier * self.r_trans
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
