"""Next Shift inference (NumPy only): adapt to a few labeled target images and predict. Usage: python inference.py --selftest from inference import load, predict, adapt """ import json import os import sys HERE = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, HERE) import numpy as np # noqa: E402 import shift_core as sc # noqa: E402 from nlrt import load_model # noqa: E402 def load(path=os.path.join(HERE, "model.json")): spec = load_model(path) t = spec["tensors"] spec["top"] = {"w1": t["w1"], "b1": t["b1"], "w2": t["w2"], "b2": t["b2"]} return spec def adapt(spec, images, labels): """images: (N, 16, 16) in [0, 1] from the target environment; returns adapted top-layer weights.""" hy = spec["meta"]["hyper"] F = sc.conv_features(spec["nets"]["conv_target"], images) return sc.finetune_top(spec["top"], F, np.asarray(labels), hy["ft_steps"], hy["ft_lr"]) def predict(spec, images, top=None): """Class probabilities (N, 10) for target-environment images (16x16, values in [0, 1]).""" F = sc.conv_features(spec["nets"]["conv_target"], images) return sc.softmax(sc.mlp_logits(top or spec["top"], F)[0]) def _selftest(): spec = load() with open(os.path.join(HERE, "selftest.json"), encoding="utf-8") as f: t = json.load(f) imgs = np.asarray(t["images"], dtype=float) / 255.0 top = adapt(spec, imgs[:len(t["labels"])], t["labels"]) P = predict(spec, imgs, top) diff = float(np.abs(P - np.asarray(t["expected_proba"])).max()) ok = diff < 1e-6 print(json.dumps({"passed": ok, "max_abs_diff": diff, "n_images": len(imgs), "n_labels": len(t["labels"])})) return ok if __name__ == "__main__": if "--selftest" in sys.argv: sys.exit(0 if _selftest() else 1) print(__doc__)