Albumentations integration in PyTorch - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output of Albumentations transform on image tensor

Given the following Albumentations transform applied to an image tensor, what is the shape of the output image after the transform?

PyTorch

import albumentations as A
import numpy as np

transform = A.Compose([
    A.HorizontalFlip(p=1.0),
    A.RandomBrightnessContrast(p=1.0)
])

image = np.random.randint(0, 256, (128, 128, 3), dtype=np.uint8)
augmented = transform(image=image)
output_image = augmented['image']
print(output_image.shape)

A (128, 128, 3)

B (3, 128, 128)

C (128, 128)

D (256, 256, 3)

💡 Hint

Albumentations works with HWC format images by default.

Albumentations expects images in height-width-channel format and returns the same shape after applying transforms that do not change size.

❓ Model Choice

intermediate

Choosing the correct way to integrate Albumentations with PyTorch DataLoader

Which option correctly integrates Albumentations transforms into a PyTorch Dataset's __getitem__ method for image augmentation?

PyTorch

import albumentations as A
from torch.utils.data import Dataset
import cv2

class CustomDataset(Dataset):
    def __init__(self, image_paths, transform=None):
        self.image_paths = image_paths
        self.transform = transform

    def __getitem__(self, idx):
        image = cv2.imread(self.image_paths[idx])
        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
        if self.transform:
            augmented = self.transform(image=image)
            image = augmented['image']
        return image

    def __len__(self):
        return len(self.image_paths)

A Apply Albumentations transform inside __getitem__ by passing image as a numpy array and extracting 'image' key from result.

B Apply Albumentations transform outside Dataset class after loading all images as tensors.

C Convert image to tensor first, then apply Albumentations transform directly on tensor inside __getitem__.

D Apply Albumentations transform only on file paths before loading images.

💡 Hint

Albumentations expects numpy arrays, not tensors.

Albumentations works with numpy arrays, so transform must be applied inside __getitem__ before converting to tensor.

❓ Hyperparameter

advanced

Effect of Albumentations 'p' parameter on augmentation frequency

In Albumentations, what does setting the parameter p=0.5 for a transform like HorizontalFlip mean?

A The transform is applied only if the image size is greater than 50 pixels.

B The transform is applied to exactly half of the pixels in the image.

C The transform is applied to every image with 50% probability.

D The transform is applied twice to each image.

💡 Hint

Think about probability of applying the transform per image.

The 'p' parameter controls the chance the transform is applied to each image independently.

❓ Metrics

advanced

Impact of Albumentations on model training metrics

If you add Albumentations augmentations like RandomBrightnessContrast and HorizontalFlip during training, what is the most likely effect on validation accuracy and training loss?

A Training loss and validation accuracy both worsen significantly.

B Training loss decreases faster and validation accuracy drops immediately.

C Both training loss and validation accuracy remain unchanged.

D Training loss may increase initially, but validation accuracy improves due to better generalization.

💡 Hint

Augmentation adds variety to training data.

Augmentation makes training harder initially, increasing loss, but helps model generalize better, improving validation accuracy.

🔧 Debug

expert

Debugging Albumentations transform causing runtime error

You have this Albumentations Compose transform: transform = A.Compose([ A.RandomCrop(width=100, height=100), A.HorizontalFlip(p=0.5), A.Normalize() ]) When applying it to images of size (80, 80, 3), you get a runtime error. What is the cause?

A Compose requires all transforms to have p=1.0.

B RandomCrop size is larger than input image size causing an error.

C Normalize requires images to be float32 but input is uint8.

D HorizontalFlip cannot be applied to images smaller than 100x100.

💡 Hint

Check the sizes used in RandomCrop compared to input image.

RandomCrop cannot crop a region larger than the input image, causing a runtime error.