Data augmentation in pipeline in TensorFlow - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output of TensorFlow data augmentation pipeline

What is the shape of the output tensor after applying this augmentation pipeline to a batch of 32 images of size 64x64x3?

TensorFlow

import tensorflow as tf

batch_size = 32
image_size = 64

# Create dummy batch of images
images = tf.random.uniform([batch_size, image_size, image_size, 3])

# Define augmentation pipeline
augmentation = tf.keras.Sequential([
    tf.keras.layers.RandomFlip('horizontal'),
    tf.keras.layers.RandomRotation(0.1),
    tf.keras.layers.RandomZoom(0.2)
])

augmented_images = augmentation(images)
output_shape = augmented_images.shape

A (32, 62, 62, 3)

B (32, 64, 64, 3)

C (32, 64, 64, 1)

D (64, 64, 3)

💡 Hint

Augmentation layers keep the batch size and image dimensions the same.

The augmentation layers like RandomFlip, RandomRotation, and RandomZoom in TensorFlow keep the input shape unchanged. They only modify pixel values or positions but do not change the tensor shape.

🧠 Conceptual

intermediate

Purpose of data augmentation in training pipelines

Why do we use data augmentation in machine learning training pipelines?

A To artificially increase the diversity of training data and reduce overfitting

B To remove noise from the training data

C To convert images to grayscale for simpler models

D To reduce the size of the training dataset for faster training

💡 Hint

Think about how augmentation affects model generalization.

Data augmentation creates new variations of existing data by applying transformations. This helps the model learn to generalize better and reduces overfitting.

❓ Hyperparameter

advanced

Choosing augmentation parameters for RandomRotation

Which of these RandomRotation parameters will rotate images up to 45 degrees in either direction?

A tf.keras.layers.RandomRotation(factor=0.125)

B tf.keras.layers.RandomRotation(factor=0.25)

C tf.keras.layers.RandomRotation(factor=0.5)

D tf.keras.layers.RandomRotation(factor=1.0)

💡 Hint

Rotation factor is a fraction of 2π radians (360 degrees).

RandomRotation's factor determines the maximum rotation angle of factor * 360 degrees in either direction. 0.125 × 360° = 45°, so rotation up to ±45°.

❓ Metrics

advanced

Effect of data augmentation on training and validation accuracy

If a model trained without augmentation has training accuracy 98% and validation accuracy 75%, what is the most likely effect of adding data augmentation?

A Both training and validation accuracy increase

B Training accuracy increases, validation accuracy decreases

C Training accuracy decreases, validation accuracy increases

D Both training and validation accuracy decrease

💡 Hint

Augmentation makes training harder but improves generalization.

Augmentation adds variability making training harder, so training accuracy may drop. But it helps the model generalize better, improving validation accuracy.

🔧 Debug

expert

Debugging augmentation pipeline causing shape mismatch error

Given this pipeline, which option explains why a shape mismatch error occurs during training? ```python import tensorflow as tf augmentation = tf.keras.Sequential([ tf.keras.layers.RandomFlip('horizontal'), tf.keras.layers.RandomRotation(0.2), tf.keras.layers.Resizing(128, 128) ]) # Input images are 64x64x3 images = tf.random.uniform([16, 64, 64, 3]) augmented = augmentation(images) model = tf.keras.Sequential([ tf.keras.layers.InputLayer(input_shape=(64, 64, 3)), tf.keras.layers.Conv2D(32, 3, activation='relu'), tf.keras.layers.Flatten(), tf.keras.layers.Dense(10) ]) model(augmented) ```

A Resizing layer is missing required arguments for interpolation

B RandomFlip changes the number of channels from 3 to 1, causing mismatch

C RandomRotation outputs a batch size different from input batch size

D The model expects input shape (64,64,3) but augmentation outputs (128,128,3), causing mismatch

💡 Hint

Check the input shape expected by the model vs output shape of augmentation.

The augmentation pipeline resizes images to 128x128, but the model expects 64x64 input, causing a shape mismatch error.