mirror of
https://github.com/MLSysBook/TinyTorch.git
synced 2026-05-28 06:16:13 -05:00
3981032e35
Key Changes: - Implemented CrossEntropyBackward for gradient computation - Integrated CrossEntropyLoss into enable_autograd() patching - Created comprehensive loss gradient test suite - Milestone 03: MLP digits classifier (77.5% accuracy) - Shipped tiny 8x8 digits dataset (67KB) for instant demos - Updated DataLoader module with ASCII visualizations Tests: - All 3 losses (MSE, BCE, CrossEntropy) now have gradient flow - MLP successfully learns digit classification (6.9% → 77.5%) - Integration tests pass Technical: - CrossEntropyBackward: softmax - one_hot gradient - Numerically stable via log-softmax - Works with raw class labels (no one-hot needed)