如何在 CNN 中将维度为 1x1xC 的特征图的形状更改为 NxNxC 维度的特征图？

Question

image_size = 256 #image size which is input shape of the densenet121
base_model=get_base_model('densenet121',image_size) #calling densenet121 using transfer learning
base_in=base_model.input #base_in.shape = (None, 256, 256, 3)
base_out=base_model.output #base_out.shape = (None, 8, 8, 1024)

#function which performs attention which is to be added on the densenet121
def BAM_Channel_Attention(inputs):
  shape=K.int_shape(inputs)
  x=tf.keras.layers.AveragePooling2D(pool_size=(shape[1],shape[2]))(inputs)
  dense = layers.Dense(1280, activation="relu")(x)
  reduction = Conv2D(shape[3]/16,1, padding='same')(dense)
  dense = layers.Dense(1280, activation="relu")(reduction)
  out = tf.keras.layers.BatchNormalization()(dense)
  return out

上面的函数在这里被调用：

#calling the attention function with the output of the densenet121
x=BAM_Channel_Attention(base_out) #

print(x.shape)

我得到了输出(None, 1, 1, 1280)。如何将特征图从 1x1x1280 更改为 8x8x1280？

Answer 1

最简单但可能不是最实用的选项是在BAM_Channel_Attention 输出上使用tf.repeat：

import tensorflow as tf

batch_size = 2
data = tf.random.normal((batch_size, 1, 1, 1280))
output = tf.repeat(tf.repeat(data, repeats=8, axis=1), repeats=8, axis=2)
print(output.shape)

(2, 8, 8, 1280)

可能更好的方法是在BAM_Channel_Attention 函数中更改AveragePooling2D 层中的pool_size，因为它会将您的空间维度从(8, 8) 下采样到(1,1)。您可以将 pool_size 从 (8, 8) 更改为 (1, 1)：

import tensorflow as tf

batch_size = 2
inputs = tf.keras.layers.Input((8, 8, 1024))
x = tf.keras.layers.AveragePooling2D(pool_size=(1,1))(inputs)
dense = tf.keras.layers.Dense(1280, activation="relu")(inputs)
reduction = tf.keras.layers.Conv2D(1024/16, 1, padding='same')(dense)
dense = tf.keras.layers.Dense(1280, activation="relu")(reduction)
outputs = tf.keras.layers.BatchNormalization()(dense)

model = tf.keras.Model(inputs, outputs)
print(model(tf.random.normal((batch_size, 8, 8, 1024))).shape)

(2, 8, 8, 1280)

或者完全省略AveragePooling2D层，你也会得到你想要的结果。