使用步幅为1的最大池时的维度问题

Question

正如这篇文章所述，在max-pooling层中使用stride 1不应该减少维度。然而，在进行一些实验时，我观察到了一些不同的东西。

my_model = tf.keras.Sequential([ 
tf.keras.layers.Flatten(input_shape=(10,))])  
my_model.add(tf.keras.layers.Dense(16,activation=tf.identity))
my_model.add(tf.keras.layers.Reshape([16,1]))
my_model.add(tf.keras.layers.MaxPooling1D(pool_size=(4), strides=(1)))
my_model.add(tf.keras.layers.Flatten())
my_model.add(tf.keras.layers.Dense(16,activation=tf.identity))

my_model.summary()   

Model: "sequential_24"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
flatten_35 (Flatten)         (None, 10)                0         
_________________________________________________________________
dense_41 (Dense)             (None, 16)                176       
_________________________________________________________________
reshape_18 (Reshape)         (None, 16, 1)             0         
_________________________________________________________________
max_pooling1d_18 (MaxPooling (None, 13, 1)             0         
_________________________________________________________________
flatten_36 (Flatten)         (None, 13)                0         
_________________________________________________________________
dense_42 (Dense)             (None, 16)                224       
=================================================================
Total params: 400
Trainable params: 400
Non-trainable params: 0

可以看到，使用max-pool确实降低了数据的维数。我误解了什么吗?

Answer 1

您忘记添加padding='same'参数了。默认情况下，填充设置为valid.，这很棘手，因为只有在进行卷积/其他操作之前才会更改padding参数。

import tensorflow as tf
my_model = tf.keras.Sequential([ 
tf.keras.layers.Flatten(input_shape=(10,))])  
my_model.add(tf.keras.layers.Dense(16,activation=tf.identity))
my_model.add(tf.keras.layers.Reshape([16,1]))
my_model.add(tf.keras.layers.MaxPooling1D(pool_size=(4), strides=(1),padding='same'))
my_model.add(tf.keras.layers.Flatten())
my_model.add(tf.keras.layers.Dense(16,activation=tf.identity))

my_model.summary()   

Model: "sequential_4"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
flatten_3 (Flatten)          (None, 10)                0         
_________________________________________________________________
dense_8 (Dense)              (None, 16)                176       
_________________________________________________________________
reshape_2 (Reshape)          (None, 16, 1)             0         
_________________________________________________________________
max_pooling1d_2 (MaxPooling1 (None, 16, 1)             0         
_________________________________________________________________
flatten_4 (Flatten)          (None, 16)                0         
_________________________________________________________________
dense_9 (Dense)              (None, 16)                272       
=================================================================
Total params: 448
Trainable params: 448
Non-trainable params: 0