Tutorial 03 — CIFAR-10 Bi-Real proxy

Master guide: ../GUIDE_E2E.md · Prev: 02 · Next: 04

Goal

Compare FP32 vs Bi-Real-style binary CNN on CIFAR-10 (ImageNet substitute).

pip install datasets   :: first time — dumps to data/cifar10_hf
bnn train-cifar --epochs 5 --subset 20000

Full 50k: --subset 0 (longer). Recipe notes in docs/13_TRAINING_QAT_DISTILL.md.

For the richer image lane (ApproxSign, golden image_cifar.json), prefer 04_image_cifar.md.

Expect

Short schedules leave an accuracy gap (see results/cifar10_proxy.json). Full ImageNet train is an ADR non-goal; use published Bi-Real/ReActNet for scale-up.