Data augmentation with transforms in PyTorch - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output of a composed transform on an image tensor

Given the following PyTorch code applying data augmentation transforms, what is the shape of the output tensor after the transforms?

PyTorch

import torch
from torchvision import transforms
from PIL import Image

transform = transforms.Compose([
    transforms.Resize((64, 64)),
    transforms.RandomHorizontalFlip(p=1.0),
    transforms.ToTensor()
])

# Simulate a PIL image as a tensor with shape (3, 100, 100)
# Normally transforms expect PIL images, so convert tensor to PIL Image
image = Image.fromarray((torch.rand(3, 100, 100).permute(1, 2, 0).mul(255).byte().numpy()))

output = transform(image)

print(output.shape)

A torch.Size([1, 3, 64, 64])

B torch.Size([64, 64, 3])

C torch.Size([3, 64, 64])

D torch.Size([3, 100, 100])

💡 Hint

Remember that Resize changes the image size and ToTensor converts to a tensor with channels first.

The Resize transform changes the image size to 64x64 pixels. The RandomHorizontalFlip flips the image horizontally but does not change shape. ToTensor converts the image to a tensor with shape (channels, height, width). So the output shape is (3, 64, 64).

❓ Model Choice

intermediate

Choosing the best augmentation for rotation invariance

You want to train a model to recognize objects regardless of their orientation. Which data augmentation transform should you add to your training pipeline to help the model learn rotation invariance?

A transforms.RandomRotation(degrees=45)

B transforms.RandomHorizontalFlip(p=0.5)

C transforms.ColorJitter(brightness=0.5)

D transforms.RandomCrop(size=32)

💡 Hint

Think about which transform changes the orientation of the image.

RandomRotation rotates the image by a random angle up to 45 degrees, helping the model learn to recognize objects in different orientations. HorizontalFlip flips the image but only horizontally, ColorJitter changes colors, and RandomCrop changes the size and position but not orientation.

❓ Hyperparameter

advanced

Effect of probability parameter in RandomHorizontalFlip

In the transform transforms.RandomHorizontalFlip(p=0.7), what does the parameter p=0.7 control?

A The probability that the image will be converted to grayscale

B The probability that the image will be flipped horizontally during augmentation

C The angle in degrees to flip the image

D The scale factor for resizing the image

💡 Hint

Think about what 'p' usually means in random transforms.

The parameter p in RandomHorizontalFlip sets the chance that the image will be flipped horizontally. Here, there is a 70% chance the image flips and 30% chance it stays the same.

🔧 Debug

advanced

Debugging a transform pipeline error

You have this transform pipeline: transform = transforms.Compose([ transforms.ToTensor(), transforms.RandomCrop(32), transforms.Normalize(mean=[0.5], std=[0.5]) ]) When you run it on a PIL image of size 28x28, you get an error. What is the cause?

A RandomCrop(32) fails because the image is smaller than the crop size

B ToTensor() must be after Normalize()

C Normalize expects 3 channels but the image has 1 channel

D Transforms.Compose requires a list of functions, not a list of transforms

💡 Hint

Check the image size compared to the crop size.

RandomCrop(32) tries to crop a 32x32 patch, but the input image is only 28x28, so it raises an error. To fix, either resize the image first or use a smaller crop size.

🧠 Conceptual

expert

Why use data augmentation in training deep learning models?

Which of the following best explains why data augmentation is important when training deep learning models?

A It removes noisy data points from the training set

B It reduces the training time by simplifying the dataset

C It guarantees the model will achieve 100% accuracy on the training set

D It artificially increases the size and diversity of the training data to reduce overfitting and improve generalization

💡 Hint

Think about how data augmentation affects the training data and model learning.

Data augmentation creates new variations of existing data, helping the model learn more robust features and reducing overfitting. It does not reduce training time or guarantee perfect accuracy.