"""Train a binary perceptron using a discrete weight-flip update.""" from pathlib import Path from model import OneBitNOT DATA = [(0, 1), (1, 0)] def errors(model: OneBitNOT) -> int: return sum(model.predict(x) != target for x, target in DATA) def main() -> None: model = OneBitNOT(weight_bit=0) # Start with w=+1 print(f"Before: bit={model.weight_bit}, weight={model.weight:+d}, errors={errors(model)}") # Flip the only trainable bit if it improves the objective. candidate = OneBitNOT(weight_bit=model.weight_bit ^ 1) if errors(candidate) < errors(model): model = candidate print(f"After: bit={model.weight_bit}, weight={model.weight:+d}, errors={errors(model)}") assert errors(model) == 0 model.save(Path(__file__).resolve().parent.parent / "model.bin") for x, target in DATA: print(f"input={x}, prediction={model.predict(x)}, target={target}") if __name__ == "__main__": main()