KERAS和tensroflow实现Attention机制

一、Keras实现Attention

from keras import backend as K
from keras.engine.topology import Layer

class Position_Embedding_Attention(Layer):

    def __init__(self, size=None, mode='sum', **kwargs):
        self.size = size #必须为偶数
        self.mode = mode
        super(Position_Embedding, self).__init__(**kwargs)

    def call(self, x):
        if (self.size == None) or (self.mode == 'sum'):
            self.size = int(x.shape[-1])
        batch_size,seq_len = K.shape(x)[0],K.shape(x)[1]
        position_j = 1. / K.pow(10000., 2 * K.arange(self.size / 2, dtype='float32') / self.size)
        position_j = K.expand_dims(position_j, 0)
        position_i = K.cumsum(K.ones_like(x[:,:,0]), 1)-1 #K.arange不支持变长，只好用这种方法生成
        position_i = K.expand_dims(position_i, 2)
        position_ij = K.dot(position_i, position_j)
        position_ij = K.concatenate([K.cos(position_ij), K.sin(position_ij)], 2)
        if self.mode == 'sum':
            return position_ij + x
        elif self.mode == 'concat':
            return K.concatenate([position_ij, x], 2)

    def compute_output_shape(self, input_shape):
        if self.mode == 'sum':
            return input_shape
        elif self.mode == 'concat':
            return (input_shape[0], input_shape[1], input_shape[2]+self.size)

class Multi_Head_AttentionAttention(Layer):

    def __init__(self, nb_head, size_per_head, **kwargs):
        self.nb_head = nb_head
        self.size_per_head = size_per_head
        self.output_dim = nb_head*size_per_head
        super(Attention, se