Compose transforms in PyTorch - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output of composed image transforms

What is the output shape of the tensor after applying the composed transforms to a 28x28 grayscale image?

PyTorch

from torchvision import transforms
from PIL import Image
import torch

# Create a dummy grayscale image of size 28x28
image = Image.new('L', (28, 28))

# Compose transforms: Resize to 32x32, convert to tensor, normalize
transform = transforms.Compose([
    transforms.Resize((32, 32)),
    transforms.ToTensor(),
    transforms.Normalize((0.5,), (0.5,))
])

transformed_image = transform(image)
output_shape = transformed_image.shape

A torch.Size([1, 32, 32])

B torch.Size([3, 28, 28])

C torch.Size([3, 32, 32])

D torch.Size([1, 28, 28])

💡 Hint

Think about the number of channels in a grayscale image and what Resize and ToTensor do.

The original image is grayscale, so it has 1 channel. Resize changes the spatial size to 32x32. ToTensor converts the image to a tensor with shape (channels, height, width). Normalize does not change the shape. So the output shape is (1, 32, 32).

❓ Model Choice

intermediate

Choosing the correct transform composition for data augmentation

Which composed transform will correctly apply random horizontal flip and convert the image to a tensor for training a model?

A transforms.ToTensor(transforms.RandomHorizontalFlip())

B transforms.Compose([transforms.ToTensor(), transforms.RandomHorizontalFlip()])

C transforms.RandomHorizontalFlip(transforms.ToTensor())

D transforms.Compose([transforms.RandomHorizontalFlip(), transforms.ToTensor()])

💡 Hint

Remember the order of transforms matters: image operations before tensor conversion.

RandomHorizontalFlip expects a PIL image input, so it must come before ToTensor. Option D correctly composes the transforms in the right order. Options B, C, and D are incorrect because they either apply tensor transforms before image transforms or misuse the API.

❓ Hyperparameter

advanced

Effect of probability parameter in RandomHorizontalFlip

What is the effect of setting the probability parameter p=0.0 in transforms.RandomHorizontalFlip() inside a Compose?

A Images will never be flipped horizontally during transform.

B Images will always be flipped horizontally during transform.

C Images will be flipped horizontally with 50% chance.

D Transforms will raise an error due to invalid probability.

💡 Hint

Probability p controls how often the flip happens.

The parameter p in RandomHorizontalFlip controls the probability of flipping. p=0.0 means no flips occur, so images remain unchanged. p=1.0 would flip all images. p=0.5 is the default for random flipping.

🔧 Debug

advanced

Identify the error in this Compose transform code

What error will this code raise when applying the transform to a PIL image?

PyTorch

from torchvision import transforms
transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.RandomRotation(30)
])

from PIL import Image
image = Image.new('RGB', (64, 64))
transformed = transform(image)

A No error, code runs successfully.

B TypeError: img should be PIL Image or ndarray. Got Tensor instead.

C AttributeError: 'Tensor' object has no attribute 'rotate'.

D ValueError: angle must be a number.

💡 Hint

Check the expected input types for RandomRotation.

RandomRotation expects a PIL Image or numpy ndarray as input. Since ToTensor converts the image to a tensor first, RandomRotation receives a tensor and raises a TypeError.

🧠 Conceptual

expert

Understanding Compose transform order impact on model input

Why is the order of transforms in torchvision.transforms.Compose critical when preparing data for a neural network?

A Because transforms are applied in parallel, so order is irrelevant.

B Because the order only affects training speed, not data correctness.

C Because some transforms expect PIL images and others expect tensors, so wrong order causes errors or wrong data.

D Because Compose automatically sorts transforms by type, so order does not matter.

💡 Hint

Think about input types each transform expects and produces.

Transforms like RandomRotation or RandomHorizontalFlip expect PIL images or numpy arrays, while ToTensor converts images to tensors. Applying tensor-only transforms before image transforms causes errors or incorrect data. Hence, order matters to ensure correct data flow.