Keras中的注意力:如何在keras密集层中添加不同的注意力机制？

from keras.layers import Dense, Activation, Multiply, Reshape, Lambda
from keras import backend as K

def attention(inputs):
    attention_weights = Dense(1, activation='tanh')(inputs)
    attention_weights = Activation('softmax')(attention_weights)
    attention_weights = Multiply()([inputs, attention_weights])
    attention_weights = Reshape((-1,))(attention_weights)
    return attention_weights

inputs = Dense(units)(previous_layer)
attention_weights = attention(inputs)
output = Lambda(lambda x: K.sum(x, axis=1))(attention_weights)