Custom detection dataset in PyTorch - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output of Custom Dataset __getitem__ method

Given the following PyTorch custom dataset code snippet, what is the output type of the __getitem__ method when called with an index?

PyTorch

from torch.utils.data import Dataset

class CustomDetectionDataset(Dataset):
    def __init__(self, images, annotations):
        self.images = images
        self.annotations = annotations

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

    def __getitem__(self, idx):
        image = self.images[idx]
        target = self.annotations[idx]
        return image, target

A A list containing image and target tensors

B Only the image tensor at the given index

C Only the target dictionary at the given index

D A tuple containing (image, target) where image is a tensor and target is a dictionary

💡 Hint

Look at what the __getitem__ method returns explicitly.

The __getitem__ method returns two objects as a tuple: the image and the target (annotation). This is a common pattern in detection datasets where targets are dictionaries with bounding boxes and labels.

❓ data_output

intermediate

Number of items in Custom Detection Dataset

If a CustomDetectionDataset is initialized with a list of 150 images and a list of 150 corresponding annotations, what will be the output of len(dataset)?

PyTorch

dataset = CustomDetectionDataset(images=[...150 images...], annotations=[...150 annotations...])
print(len(dataset))

A 150

B 0

C Depends on the length of annotations only

D Raises a TypeError

💡 Hint

Check the __len__ method implementation.

The __len__ method returns the length of the images list, which is 150 in this case.

🔧 Debug

advanced

Fixing a KeyError in Custom Detection Dataset

Consider this code snippet inside a custom dataset __getitem__ method: image = self.images[idx] target = self.annotations[idx] boxes = target['boxes'] Sometimes, running this code raises a KeyError: 'boxes'. Which option fixes this error by ensuring 'boxes' key exists in target?

PyTorch

def __getitem__(self, idx):
    image = self.images[idx]
    target = self.annotations[idx]
    boxes = target['boxes']
    return image, target

A Use boxes = target.get('boxes', torch.empty((0,4))) to provide an empty tensor if 'boxes' is missing

B Replace boxes = target['boxes'] with boxes = target['box']

C Remove the line boxes = target['boxes'] entirely

D Change target = self.annotations[idx] to target = self.annotations.get(idx)

💡 Hint

Use a safe way to access dictionary keys that might be missing.

Using target.get('boxes', torch.empty((0,4))) returns an empty tensor if 'boxes' key is missing, preventing KeyError.

🚀 Application

advanced

Transforming Images in Custom Detection Dataset

You want to apply a random horizontal flip to images and their bounding boxes in your custom detection dataset. Which approach correctly applies the transform inside __getitem__?

PyTorch

def __getitem__(self, idx):
    image = self.images[idx]
    target = self.annotations[idx]
    if random.random() > 0.5:
        # Apply horizontal flip to image and boxes
        image = image.flip(-1)
        boxes = target['boxes']
        width = image.shape[-1]
        boxes[:, [0, 2]] = width - boxes[:, [2, 0]]
        target['boxes'] = boxes
    return image, target

A The boxes coordinates are not updated correctly after flip

B The image.flip(-1) should be image.flip(0) to flip horizontally

C This code correctly flips both image and bounding boxes horizontally

D The target dictionary should not be modified inside __getitem__

💡 Hint

Check how image.flip and box coordinate updates correspond to horizontal flip.

image.flip(-1) flips the image horizontally. Updating boxes by width - boxes[:, [2, 0]] correctly adjusts x coordinates after flip.

🧠 Conceptual

expert

Why use a custom detection dataset in PyTorch?

Which of the following is the best reason to create a custom detection dataset class in PyTorch instead of using a built-in dataset?

A To automatically generate synthetic images

B To handle unique image and annotation formats not supported by built-in datasets

C To avoid using DataLoader for batching

D Because built-in datasets are slower to load images

💡 Hint

Think about dataset flexibility and formats.

Custom datasets allow handling of unique data formats and annotations that built-in datasets do not support.