Keras中具有屏蔽支持的平均池或最大池 [英] Mean or max pooling with masking support in Keras

问题描述

...
print('Build model...')
model = Sequential()
model.add(Embedding(max_features, 128))
model.add(LSTM(size, return_sequences=True, dropout_W=0.2 dropout_U=0.2)) 
model.add(GlobalAveragePooling1D())
model.add(Dense(1))
model.add(Activation('sigmoid'))
....

我需要能够将LSTM层之后的样本中所有时间步长的向量的平均值或最大值取下，然后再将该平均值或最大值提供给Keras中的密集层.

I need to be able to take the mean or max of the vectors for all time steps in a sample after LSTM layer before giving this mean or max vector to the dense layer in Keras.

我认为timedistributedmerge能够做到这一点，但已弃用.使用return_sequences=True，我可以获取LSTM层之后的样本中所有时间步长的向量.但是，GlobalAveragePooling1D()与掩码不兼容，它考虑了所有时间步长，而我只需要非掩码时间步长.

I think timedistributedmerge was able to do this but it was deprecated. Using return_sequences=True I can obtain the vectors for all time steps in a sample after the LSTM layer. However, GlobalAveragePooling1D() is not compatible with masking and it considers all time steps whereas I need only the non-masked time steps.

我看到了推荐Lambda层的帖子，但这些帖子也没有考虑屏蔽.任何帮助将不胜感激.

I saw posts recommending the Lambda layer but these also do not take masking into account. Any help would be appreciated.

推荐答案

Jacoxu的答案是正确的.但是，如果您使用的是用于keras的tensorflow后端，则Tensor类型不支持dimshuffle函数，请尝试使用此方法.

Jacoxu's answer is right. But if you are using a tensorflow backend for keras, the Tensor type doesn't support dimshuffle function, try this instead.

def call(self, x, mask=None):
    if mask is not None:
        # mask (batch, time)
        mask = K.cast(mask, K.floatx())
        # mask (batch, x_dim, time)
        mask = K.repeat(mask, x.shape[-1])
        # mask (batch, time, x_dim)
        mask = tf.transpose(mask, [0,2,1])
        x = x * mask
    return K.sum(x, axis=1) / K.sum(mask, axis=1)

这篇关于Keras中具有屏蔽支持的平均池或最大池的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！