Data augmentation for images in TensorFlow - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output shape after applying random flip augmentation

Consider the following TensorFlow code that applies a random horizontal flip to a batch of images. What is the shape of the output tensor?

TensorFlow

import tensorflow as tf

images = tf.random.uniform([16, 64, 64, 3])  # batch of 16 images, 64x64 RGB
augmented_images = tf.image.random_flip_left_right(images)
output_shape = augmented_images.shape
print(output_shape)

A (16, 64, 64, 3)

B (64, 64, 3)

C (16, 128, 128, 3)

D (16, 64, 64)

💡 Hint

The random flip changes pixel order but not the shape.

The random_flip_left_right function flips images horizontally but does not change their shape or batch size. So the output shape remains (16, 64, 64, 3).

❓ Model Choice

intermediate

Choosing augmentation for small dataset

You have a small image dataset and want to reduce overfitting by augmenting data during training. Which augmentation technique is best suited to increase data variety without changing image content drastically?

A Random rotation by 90 degrees increments only

B Random cropping with resizing back to original size

C Random color inversion (negative colors)

D Random Gaussian noise addition

💡 Hint

Think about preserving main image features while increasing variety.

Random cropping with resizing keeps the main content but changes the framing, increasing variety without drastic content changes. Rotations by 90 degrees may be too rigid, color inversion changes semantics, and noise addition may harm learning.

❓ Metrics

advanced

Effect of augmentation on validation accuracy

A model is trained twice: once with data augmentation and once without. The validation accuracy with augmentation is 85%, without augmentation is 78%. What does this difference most likely indicate?

A Augmentation reduced model capacity

B Augmentation caused overfitting on training data

C Augmentation improved model generalization to unseen data

D Augmentation caused data leakage

💡 Hint

Higher validation accuracy usually means better generalization.

Data augmentation increases training data variety, helping the model generalize better and improving validation accuracy. Overfitting would lower validation accuracy, and data leakage would cause suspiciously high accuracy.

🔧 Debug

advanced

Debugging augmentation pipeline error

You run this TensorFlow code but get a runtime error. What is the cause? ```python import tensorflow as tf images = tf.random.uniform([8, 32, 32, 3]) augmented = tf.image.random_brightness(images, max_delta=0.5) augmented = tf.image.random_flip_left_right(augmented) augmented = tf.image.random_crop(augmented, size=[8, 28, 28, 3]) ```

A random_crop cannot be applied after random_flip_left_right

B random_brightness requires images to be uint8, not float32

C random_flip_left_right only works on grayscale images

D random_crop size must be smaller than input size but batch dimension cannot be included

💡 Hint

Check the shape argument for random_crop carefully.

tf.image.random_crop expects the size argument to exclude the batch dimension. Including batch size causes a runtime error. The batch dimension must be handled separately.

🧠 Conceptual

expert

Why use augmentation only on training data?

Why is data augmentation typically applied only to training images and not to validation or test images?

A To simulate more diverse training data and fairly evaluate model on original data

B Because augmentation algorithms are too slow for validation and test sets

C Because validation and test images are already augmented by default

D To prevent the model from memorizing augmented images in validation and test sets

💡 Hint

Think about the purpose of validation and test sets.

Augmentation increases training data diversity to improve learning. Validation and test sets must remain unchanged to fairly measure model performance on real, unaltered data.