Using AveragePooling2D instead of GlobalAveragePooling 2D to replace the Dense layers after an FCN for classification

GlobalAveragePooling2D will downsample an input by taking the average value along the spatial dimensions and return a 1D output by default, unless you set keepdims= True. AveragePooling2D also downsamples an input but takes the average value over an input window defined by the pool_size parameter. So, it will return a 3D output:

import tensorflow as tf
x = tf.constant([[1., 2., 3.],
                 [4., 5., 6.],
                 [7., 8., 9.]])

x = tf.reshape(x, [1, 3, 3, 1]) # Add batch dimension

avg_pool_2d = tf.keras.layers.AveragePooling2D(pool_size=(2, 2),
   strides=(1, 1), padding='valid')
avg_global_2d = tf.keras.layers.GlobalAveragePooling2D()
print(avg_pool_2d(x).shape)
print(avg_global_2d(x).shape)

(1, 2, 2, 1)
(1, 1)

You will probably have to flatten your output from the AveragePooling2D layer if you want to feed it to a Dense layer afterwards:

import tensorflow as tf
x = tf.constant([[1., 2., 3.],
                 [4., 5., 6.],
                 [7., 8., 9.]])

x = tf.reshape(x, [1, 3, 3, 1]) # Add batch dimension

avg_pool_2d = tf.keras.layers.AveragePooling2D(pool_size=(2, 2),
   strides=(1, 1), padding='valid')
avg_global_2d = tf.keras.layers.GlobalAveragePooling2D()
print(tf.keras.layers.Flatten()(avg_pool_2d(x)).shape)
print(avg_global_2d(x).shape)

(1, 4)
(1, 1)

If that is not the case you can just leave it as it is.

Update: GlobalAveragePooling2D and AveragePooling2D can sometimes behave similarly if you adjust the strides and pool_size parameters accordingly:

import tensorflow as tf
x = tf.constant([[1., 2., 3.],
                 [4., 5., 6.],
                 [7., 8., 9.]])

x = tf.reshape(x, [1, 3, 3, 1]) # Add batch dimension

avg_pool_2d = tf.keras.layers.AveragePooling2D(pool_size=(2, 2),
   strides=(3, 3), padding='valid')
avg_global_2d = tf.keras.layers.GlobalAveragePooling2D()

print(avg_pool_2d(x))
print(avg_global_2d(x))

tf.Tensor([[[[3.]]]], shape=(1, 1, 1, 1), dtype=float32)
tf.Tensor([[5.]], shape=(1, 1), dtype=float32)

Or

avg_pool_2d = tf.keras.layers.AveragePooling2D(pool_size=(3, 3),
   strides=(2, 2), padding='valid')
avg_global_2d = tf.keras.layers.GlobalAveragePooling2D()

tf.Tensor([[[[5.]]]], shape=(1, 1, 1, 1), dtype=float32)
tf.Tensor([[5.]], shape=(1, 1), dtype=float32)

Adding to the answer above, global average pooling can be used for taking variable size images as inputs. If the input shape before global pooling is (N,H,W,C) then output will be (N,1,1,C) for keras when keepdims=True. This makes the output of images with different (H,W) produce similar shape outputs.

References:

• https://stats.stackexchange.com/questions/388859/is-it-possible-to-give-variable-sized-images-as-input-to-a-convolutional-neural