Tutorial 01 — MNIST binary train + bench (≈15 min)

Master guide: ../GUIDE_E2E.md · Next: 02_wrap_linear.md

Goal

Train a binary MLP on MNIST and confirm packed kernel correctness/speed.

Steps

pip install -e ".[dev]" -c constraints.txt
python -m bnn.kernels.compile_native
bnn validate-native
bnn export-check
bnn train --epochs 3 --seed 42 --model binary_mlp
bnn bench --reps 5
:: Or verify committed goldens without retraining:
bnn repro

Expect

• Native GEMM err = 0 (Windows MSVC DLL); NumPy path err = 0 everywhere
• Compression 32× exact
• binary_mlp test acc ≥ ~95% when FP is ≥97% (see tests/golden_floors.json)

Notes

Training uses STE (not faster than FP). Inference wins need packed kernels. Full agent/human repro: REPRODUCIBILITY.md.