Padding and stride in TensorFlow - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output shape with SAME padding and stride 2

Consider a 2D convolution layer in TensorFlow with input shape (1, 28, 28, 3), kernel size 3x3, stride 2, and padding='SAME'. What is the output shape after applying this convolution?

TensorFlow

import tensorflow as tf
input_tensor = tf.random.uniform((1, 28, 28, 3))
conv_layer = tf.keras.layers.Conv2D(filters=5, kernel_size=3, strides=2, padding='SAME')
output = conv_layer(input_tensor)
print(output.shape)

A (1, 14, 14, 5)

B (1, 12, 12, 5)

C (1, 15, 15, 5)

D (1, 13, 13, 5)

💡 Hint

Padding='SAME' keeps output size roughly input size divided by stride.

With padding='SAME' and stride=2, output height and width are ceil(input_dim / stride). For 28 input, output is ceil(28/2)=14.

🧠 Conceptual

intermediate

Effect of stride on output size

In a convolutional neural network, if you increase the stride from 1 to 3 while keeping padding='VALID' and kernel size fixed, what happens to the output feature map size?

A Output size stays the same because kernel size is fixed.

B Output size increases because stride adds padding.

C Output size becomes zero because stride is too large.

D Output size decreases because stride skips more input positions.

💡 Hint

Think about how stride controls the step size when sliding the kernel.

Increasing stride makes the kernel jump more pixels, reducing the number of positions it fits, so output size decreases.

❓ Metrics

advanced

Calculate output size with padding='VALID' and stride 2

Given an input image of size 32x32, a convolutional layer with kernel size 5x5, stride 2, and padding='VALID', what is the output width and height?

A 16 x 16

B 15 x 15

C 14 x 14

D 13 x 13

💡 Hint

Use formula: output = floor((input - kernel + 1) / stride)

Output size = floor((32 - 5 + 1) / 2) = floor(28 / 2) = 14.

🔧 Debug

advanced

Identify the error in convolution output shape calculation

A user writes this code to compute output shape after a convolution: input_shape = (1, 64, 64, 3) kernel_size = 7 stride = 2 padding = 'VALID' output_height = (input_shape[1] - kernel_size) // stride output_width = (input_shape[2] - kernel_size) // stride print(f'Output shape: (1, {output_height}, {output_width}, 10)') What is wrong with this calculation?

A Padding='VALID' requires adding padding size to input

B The kernel size should be subtracted with +1: (input - kernel + 1)

C Output channels should be input channels minus kernel size

D The stride should be added, not divided

💡 Hint

Recall the formula for output size with VALID padding.

The formula for output size with VALID padding is floor((input - kernel + 1) / stride), missing the +1 causes wrong output size.

❓ Model Choice

expert

Choosing padding and stride for preserving spatial dimensions

You want to design a convolutional layer that keeps the input image size exactly the same after convolution. Which combination of padding and stride should you choose?

A padding='SAME' and stride=1

B padding='VALID' and stride=1

C padding='SAME' and stride=2

D padding='VALID' and stride=2

💡 Hint

Think about how padding and stride affect output size.

Padding='SAME' with stride=1 keeps output size equal to input size by adding zero padding as needed.