Keras中的注意力:如何在keras密集层中添加不同的注意力机制？

Question

我是Keras的新手，我正在尝试用keras构建一个简单的自动编码器，并带有关注层：

下面是我尝试过的：

data = Input(shape=(w,), dtype=np.float32, name='input_da')
noisy_data = Dropout(rate=0.2, name='drop1')(data)

encoded = Dense(256, activation='relu',
            name='encoded1', **kwargs)(noisy_data)
encoded = Lambda(mvn, name='mvn1')(encoded)

encoded = Dense(128, activation='relu',
            name='encoded2', **kwargs)(encoded)

encoded = Lambda(mvn, name='mvn2')(encoded)
encoded = Dropout(rate=0.5, name='drop2')(encoded)


encoder = Model([data], encoded)
encoded1 = encoder.get_layer('encoded1')
encoded2 = encoder.get_layer('encoded2')


decoded = DenseTied(256, tie_to=encoded2, transpose=True,
            activation='relu', name='decoded2')(encoded)
decoded = Lambda(mvn, name='new_mv')(decoded)


decoded = DenseTied(w, tie_to=encoded1, transpose=True,
            activation='linear', name='decoded1')(decoded)

它看起来像这样：

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
data (InputLayer)            (None, 2693)              0         
_________________________________________________________________
drop1 (Dropout)              (None, 2693)              0         
_________________________________________________________________
encoded1 (Dense)             (None, 256)               689664    
_________________________________________________________________
mvn1 (Lambda)                (None, 256)               0         
_________________________________________________________________
encoded2 (Dense)             (None, 128)               32896     
_________________________________________________________________
mvn2 (Lambda)                (None, 128)               0         
_________________________________________________________________
drop2 (Dropout)              (None, 128)               0         
_________________________________________________________________
decoded2 (DenseTied)         (None, 256)               256       
_________________________________________________________________
mvn3 (Lambda)                (None, 256)               0         
_________________________________________________________________
decoded1 (DenseTied)         (None, 2693)              2693      
=================================================================

在这个模型中我可以在哪里添加关注层？我应该在第一个encoded_output之后和第二个编码输入之前添加吗？

encoded = Lambda(mvn, name='mvn1')(encoded)

    Here?

encoded = Dense(128, activation='relu',
            name='encoded2', **kwargs)(encoded)

我还浏览了这个漂亮的lib：

https://github.com/CyberZHG/keras-self-attention

他们已经实现了各种类型的注意机制，但它是针对顺序模型的。我如何才能在我的模型中添加这些注意力？

我试着用非常简单的注意力：

encoded = Dense(256, activation='relu',
        name='encoded1', **kwargs)(noisy_data)


encoded = Lambda(mvn, name='mvn1')(encoded)

attention_probs = Dense(256, activation='softmax', name='attention_vec')(encoded)
attention_mul = multiply([encoded, attention_probs], name='attention_mul')
attention_mul = Dense(256)(attention_mul)

print(attention_mul.shape)

encoded = Dense(128, activation='relu',
        name='encoded2', **kwargs)(attention_mul)

它在正确的位置吗?我可以在这个模型中添加任何其他注意力机制吗？

Answer 1

我猜你所做的是一种增加注意力的正确方式，因为注意力本身什么都不是，而是可以被可视化为密集层的权重。此外，我认为在编码器之后立即应用注意力是正确的做法，因为它会将注意力应用于任务所需的数据分布中最“信息量”的部分。