我正在尝试在编码器LSTM(很多)和解码器LSTM(很多)之间添加一个Attention层。
但是我的代码似乎仅使一个LSTM解码器输入成为关注层。
如何将Attention层应用于解码器LSTM的所有输入? (Attention层的输出=(None,1440,984))
这是我模型关注层的摘要。
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) (None, 1440, 5) 0
__________________________________________________________________________________________________
bidirectional_1 (Bidirectional) (None, 1440, 984) 1960128 input_1[0][0]
__________________________________________________________________________________________________
dense_1 (Dense) (None, 1440, 1) 985 bidirectional_1[0][0]
__________________________________________________________________________________________________
flatten_1 (Flatten) (None, 1440) 0 dense_1[0][0]
__________________________________________________________________________________________________
activation_1 (Activation) (None, 1440) 0 flatten_1[0][0]
__________________________________________________________________________________________________
repeat_vector_1 (RepeatVector) (None, 984, 1440) 0 activation_1[0][0]
__________________________________________________________________________________________________
permute_1 (Permute) (None, 1440, 984) 0 repeat_vector_1[0][0]
__________________________________________________________________________________________________
multiply_1 (Multiply) (None, 1440, 984) 0 bidirectional_1[0][0]
permute_1[0][0]
__________________________________________________________________________________________________
lambda_1 (Lambda) (None, 984) 0 multiply_1[0][0]
==================================================================================================
Total params: 1,961,113
Trainable params: 1,961,113
Non-trainable params: 0
__________________________________________________________________________________________________
这是我的代码
_input = Input(shape=(self.x_seq_len, self.input_x_shape), dtype='float32')
activations = Bidirectional(LSTM(self.hyper_param['decoder_units'], return_sequences=True), input_shape=(self.x_seq_len, self.input_x_shape,))(_input)
# compute importance for each step
attention = Dense(1, activation='tanh')(activations)
attention = Flatten()(attention)
attention = Activation('softmax')(attention)
attention = RepeatVector(self.hyper_param['decoder_units']*2)(attention)
attention = Permute([2, 1])(attention)
sent_representation = Multiply()([activations, attention])
sent_representation = Lambda(lambda xin: K.sum(xin, axis=-2), output_shape=(self.hyper_param['decoder_units']*2,))(sent_representation)
attn = Model(input=_input, output=sent_representation)
model.add(attn)
#decoder
model.add(LSTM(self.hyper_param['encoder_units'], return_sequences=False, input_shape=(None, self.hyper_param['decoder_units'] * 2 )))
答案 0 :(得分:1)
注意的意思是迭代地获取一个解码器输出值(最后一个隐藏状态),然后使用此“查询”,“参与”所有“值”,这些只是编码器输出的整个列表。
因此,input1 =上一步的解码器隐藏状态:“键”
input2 =所有编码器隐藏状态:“值”
输出=上下文:所有编码器隐藏状态的加权总和
使用上下文,解码器的上一个隐藏状态和上一个转换后的输出生成下一个单词和新的隐藏输出状态,然后再次重复上述过程,直到遇到“ EOS”为止。
您的注意力逻辑本身是完美的(不包括涉及解码器的最后一行)。但是其余的代码丢失了。如果您可以共享完整的代码,则可以为您提供帮助。我认为您定义的注意力逻辑没有错误。
有关更多详细信息,请参阅https://towardsdatascience.com/create-your-own-custom-attention-layer-understand-all-flavours-2201b5e8be9e