Vijay Janapa Reddi 98a02d0efa Simplify Module 08: Focus on DataLoader mechanics, not dataset downloads ...

Removed synthetic download functions (download_mnist, download_cifar10):
- These were placeholder stubs generating random noise
- Conflicted with 'Real Data, Real Systems' philosophy
- Added scope creep (dataset management vs data loading)

Module 08 now focuses purely on:
✅ Dataset abstraction (interface design)
✅ TensorDataset implementation (in-memory wrapper)
✅ DataLoader mechanics (batching, shuffling, iteration)

Real datasets handled in examples/milestones:
- datasets/tiny/digits_8x8.npz ships with repo (instant)
- Milestone 03: MNIST download + training
- Milestone 04: CIFAR-10 download + CNN training

Separation of concerns:
- Module 08: Learn DataLoader abstraction (synthetic test data)
- Examples: Apply DataLoader to real data (actual datasets)

This follows PyTorch's pattern:
- torch.utils.data.DataLoader (abstraction)
- torchvision.datasets (actual data)

Tests still pass 100% with simplified synthetic data.

2025-09-30 15:10:08 -04:00

..

source

Simplify Module 08: Focus on DataLoader mechanics, not dataset downloads

2025-09-30 15:10:08 -04:00