"""Watch Reflex play one game in the terminal on a laptop CPU, one forward pass per move. python -m reflex.play --ckpt reflex.pt [--seed 0] [--every 500] """ import argparse, time import torch from .env import ACTIONS, reset, spawn from .evaluate import policy from .model import load def show(b, score, moves, p): tiles = [[(1 << int(x)) if x else 0 for x in row] for row in b.tolist()] print(f"\nmove {moves} score {score} " + " ".join(f"{a} {q:.2f}" for a, q in zip(ACTIONS, p.tolist()))) for row in tiles: print(" ".join(f"{t:>6}" if t else " ." for t in row)) def main(): ap = argparse.ArgumentParser() ap.add_argument("--ckpt", default="reflex.pt") ap.add_argument("--seed", type=int, default=0) ap.add_argument("--every", type=int, default=500, help="print the board every N moves") args = ap.parse_args() torch.set_num_threads(max(1, torch.get_num_threads())) model = load(args.ckpt, "cpu") gen = torch.Generator().manual_seed(args.seed) b, score, moves, t = reset(1, "cpu", gen), 0, 0, 0.0 policy(model, b) # warm-up: builds the move lookup table once while True: s = time.perf_counter() p, legal, after, reward = policy(model, b) t += time.perf_counter() - s if not legal.any(): break a = int(p.argmax(1)) if moves % args.every == 0: show(b[0], score, moves, p[0]) score += int(reward[0, a]); moves += 1 b = spawn(after[:, a], gen) show(b[0], score, moves, p[0]) print(f"\ngame over: score {score}, {moves} moves, max tile {1 << int(b.max())}, " f"{t / max(moves, 1) * 1000:.2f} ms per move on cpu (features + forward)") if __name__ == "__main__": main()