tf.keras.layers.MaxPooling2D : MaxPooling in TensorFlow

Max pooling is a downsampling technique that slides a window (e.g., 2x2) over the input feature map and extracts the maximum value from each window.

This process achieves two key goals:

1. Dimensionality Reduction: Reduces computational complexity by shrinking the feature map size.
2. Translation Invariance: Makes the model robust to small spatial shifts in input features.

In TensorFlow, tf.keras.layers.MaxPooling2D implements max pooling operation. Key parameters include:

• pool_size: Size of the pooling window (e.g., (2, 2)).
• strides: Step size of the window (defaults to pool_size if not specified).
• padding: 'valid' (no padding) or 'same' (pad to retain input size).

Implementation with TensorFlow

Here’s how to implement max pooling using tf.keras.layers.MaxPooling2D:

import tensorflow as tf
import numpy as np

#4x4 grayscale image
input_image = np.array([[
    [[1], [2], [3], [4]],
    [[5], [6], [7], [8]],
    [[9], [10], [11], [12]],
    [[13], [14], [15], [16]]
]], dtype=np.float32)

# MaxPooling2D layer
model = tf.keras.Sequential([
    tf.keras.layers.MaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='valid')
])
output = model.predict(input_image)

print("Input Shape:", input_image.shape)  
print("Output Shape:", output.shape)  
print("Pooled Output:\n", output[0, :, :, 0])

Output:

Input Shape: (1, 4, 4, 1)
Output Shape: (1, 2, 2, 1)
Pooled Output:
[[ 6. 8.]
[14. 16.]]

Max pooling layer is used to reduce overfitting, improve computational efficiency and enhances translation invariance.