Pooling layers (MaxPool, AvgPool) in TensorFlow - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output shape after MaxPooling2D layer

Given the following TensorFlow code, what is the shape of the output tensor after applying the MaxPooling2D layer?

TensorFlow

import tensorflow as tf
input_tensor = tf.random.uniform(shape=(1, 28, 28, 3))
max_pool = tf.keras.layers.MaxPooling2D(pool_size=(2, 2), strides=2)
output_tensor = max_pool(input_tensor)
output_shape = output_tensor.shape

A (1, 7, 7, 3)

B (1, 14, 14, 3)

C (1, 28, 28, 3)

D (1, 13, 13, 3)

💡 Hint

Remember that MaxPooling2D with pool_size (2,2) and stride 2 reduces each spatial dimension by half when padding is 'valid'.

The input has height and width 28. MaxPooling2D with pool size 2 and stride 2 reduces each dimension by half, so output height and width become 14. The batch size and channels remain the same.

🧠 Conceptual

intermediate

Difference between MaxPooling and AveragePooling

Which statement correctly describes the difference between MaxPooling and AveragePooling layers in convolutional neural networks?

A MaxPooling selects the maximum value in each pooling window, while AveragePooling computes the average value in each pooling window.

B MaxPooling computes the average value in each pooling window, while AveragePooling selects the maximum value.

C Both MaxPooling and AveragePooling select the minimum value in each pooling window.

D MaxPooling and AveragePooling both compute the sum of values in each pooling window.

💡 Hint

Think about how each pooling method summarizes the values in a small region.

MaxPooling picks the highest value in the window, emphasizing strong features. AveragePooling computes the mean, smoothing the features.

❓ Hyperparameter

advanced

Effect of stride on output size in pooling layers

If you apply a MaxPooling2D layer with pool_size=(3,3) and stride=1 on an input tensor of shape (1, 10, 10, 1) with padding='valid', what will be the spatial dimensions (height and width) of the output?

A (8, 8)

B (7, 7)

C (10, 10)

D (9, 9)

💡 Hint

Use the formula: output_size = (input_size - pool_size) / stride + 1

With input size 10, pool size 3, stride 1, and valid padding, output size = (10 - 3)/1 + 1 = 7 + 1 = 8 for both height and width.

🔧 Debug

advanced

Identifying error in pooling layer usage

What error will occur when running the following TensorFlow code snippet?

TensorFlow

import tensorflow as tf
input_tensor = tf.random.uniform(shape=(1, 28, 28, 3))
max_pool = tf.keras.layers.MaxPooling2D(pool_size=(2, 2), strides=3)
output_tensor = max_pool(input_tensor)
print(output_tensor.shape)

A ValueError due to stride larger than input dimension

B No error; output shape will be (1, 13, 13, 3)

C TypeError because strides must be a tuple, not an integer

D No error; output shape will be (1, 9, 9, 3)

💡 Hint

Check if strides can be an integer and how output shape is computed.

Strides can be an integer or tuple. With stride 3 and pool size 2, output size = floor((28 - 2)/3) + 1 = 9. No error occurs.

❓ Model Choice

expert

Choosing pooling layer for noise robustness

You want to build a convolutional neural network that is robust to small noisy activations in feature maps. Which pooling layer choice is best to reduce the effect of noise?

A No pooling layer, to preserve all activations including noise

B MaxPooling, because it selects the strongest activation which may amplify noise

C AveragePooling, because it smooths activations and reduces noise impact

D GlobalMaxPooling, because it selects the maximum activation over the entire feature map

💡 Hint

Consider how each pooling method treats small noisy values.

AveragePooling reduces noise by averaging values, smoothing out small noisy activations. MaxPooling can amplify noise by selecting the highest value.