Custom Dataset Class

1. Background Knowledge

PyTorch's Dataset class is the foundation for custom data handling in deep learning pipelines. It enables on-the-fly data generation, crucial for large-scale datasets that exceed memory limits.

Key Concepts

• torch.utils.data.Dataset: Abstract base class requiring init, len, and getitem implementations.
• len(self): Returns dataset size (total samples).
• getitem(self, idx): Returns data for index idx (feature, label pair).
• Memory Efficiency: Generates samples dynamically via getitem, avoiding full in-memory storage—ideal for infinite or massive datasets.
• Integration: Works with DataLoader for batching, shuffling, parallelism.
• Tensors: Use torch.tensor() for PyTorch compatibility (GPU acceleration, autograd).

Prerequisites: Basic PyTorch (torch), indexing, tensor operations.

2. Algorithm Approach

Procedural Generation: For index $i$, compute:

• Feature: $\mathbf{x}_i = [i, i+1, i+2]$ as 1D tensor.
• Label: $y_i = 3i + 3 =$\sum$$\mathbf{x}$_i$.

No complex algorithms needed—direct mathematical mapping from index to sample. This is a synthetic dataset generator, common for prototyping, testing, or simulating sequences.

3. Step-by-Step Strategy

1. Inherit Dataset:

from torch.utils.data import Dataset
import torch

class SequenceDataset(Dataset):
def __init__(self, size: int):
self.size = size # Total samples

2. Implement len:

def __len__(self) -> int:
return self.size

3. Implement getitem:

• Generate feature tensor: torch.tensor([idx, idx+1, idx+2], dtype=torch.float32)
• Compute label: torch.tensor(3*idx + 3, dtype=torch.float32) (or int)
• Return tuple (feature, label)

4. Handle Output Format: Test converts tensors to list/int via .tolist()/.item()—return raw tensors.

5. Test:

dataset = SequenceDataset(5)
feature, label = dataset # torch.tensor([0.,1.,2.]), torch.tensor(3.)

Complete Solution:

class SequenceDataset(Dataset):
 def __init__(self, size: int):
 self.size = size

 def __len__(self):
 return self.size

 def __getitem__(self, idx):
 feature = torch.tensor([idx, idx+1, idx+2], dtype=torch.float32)
 label = torch.tensor(3 * idx + 3, dtype=torch.float32)
 return feature, label

4. Common Pitfalls

• Returning Lists: getitem must return tensors, not lists (DataLoader expects tensors).
• Index Bounds: No explicit check needed—DataLoader handles 0 <= idx < len.
• Dtype Mismatch: Use torch.float32 for features, torch.float32/torch.long for labels. Test uses .item() (expects scalar tensor).
• Integer Overflow: Negligible for typical size (PyTorch tensors handle large ints).
• 2D Features: Constraints allow 1D/2D, but spec is 1D—keep simple.
• No Randomness: Deterministic generation; avoid random unless seeded.

5. Time & Space Complexity

Scalability: Handles arbitrary size (e.g., $10^9$) without memory issues—perfect for streaming/large-scale training.