"""Run the trained model from the command line, or dump reference outputs for the JavaScript parity test. python infer.py "a red fox jumping over a car in the desert" python infer.py --parity ../tests/parity.json """ import json import random import sys from pathlib import Path import torch import gen import tok from model import TinyGPT HERE = Path(__file__).resolve().parent def load(dev="cpu"): cfg = json.loads((HERE.parent / "web" / "model" / "model.json").read_text())["config"] m = TinyGPT(cfg["vocab"], cfg["ctx"], cfg["d"], cfg["layers"], cfg["heads"]) m.load_state_dict(torch.load(HERE / "model.pt", map_location=dev)) return m.to(dev).eval() def run(model, text): return [tok.VOCAB[i] for i in model.generate(tok.encode_prompt(text), tok.EOS)] if __name__ == "__main__": model = load() if sys.argv[1] == "--parity": r = random.Random(424242) prompts = [gen.sample(r)[0] for _ in range(300)] Path(sys.argv[2]).write_text(json.dumps([[p, run(model, p)] for p in prompts], ensure_ascii=False)) print("wrote", len(prompts)) else: for p in sys.argv[1:]: print(p, "\n ->", " ".join(run(model, p)))