Skip to main content

Global Average Pooling Layer

Summary​

  • Flatten vs GlobalAveragePooling2D
  • What is GlobalAveragePooling2D layer

Content​

Flatten vs GlobalAveragePooling2D​

If we use Flatten instead of GlobalAveragePooling2D, accuracy of the end model would be quite similar. However, Flatten and GlobalAveragePooling2D works differently.

Flatten​

Flatten flatten the nd vector until the 1st axis

a = tf.constant([[[[1, 2, 3], [2, 2, 3], [16, 2, 3]]]], dtype=tf.float32)

print(tf.keras.layers.Flatten()(a))
print(tf.keras.layers.GlobalAveragePooling2D()(a))
print(tf.reduce_mean(a, axis=[1, 2]))

"""
tf.Tensor([[ 1.  2.  3.  2.  2.  3. 16.  2.  3.]], shape=(1, 9), dtype=float32)
tf.Tensor([[6.3333335 2.        3.       ]], shape=(1, 3), dtype=float32)
tf.Tensor([[6.3333335 2.        3.       ]], shape=(1, 3), dtype=float32)
"""

What is Average Pooling Layer​

In previous example, we were using the EfficientNetB0 to use derive the existing features learnt by the base_model. When training the new model with the base model, we keep the base model unchanged but get the feature vectors from the base model using GlobalAveragePooling2D layer. Here GlobalAveragePooling2D is used to extract the most important features from the base model output.