visualize_dataset.py

import argparse
import os
import random

import imageio
import numpy as np


def load_shard(path: str):
    data = np.load(path, allow_pickle=True)
    frames = data["frames"]
    w = data["w"]
    h = data["h"]
    return frames, w, h


def to_gray(frame: np.ndarray, alphabet_size: int) -> np.ndarray:
    scale = int(255 / max(1, alphabet_size - 1))
    return (frame.astype(np.uint8) * scale)


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--data", default="./nca_dataset_set2/data")
    parser.add_argument("--out", default="./sample_rollout.gif")
    parser.add_argument("--alphabet", type=int, default=32)
    parser.add_argument("--max-frames", type=int, default=200)
    args = parser.parse_args()

    shard_files = [
        os.path.join(args.data, f)
        for f in os.listdir(args.data)
        if f.startswith("shard_") and f.endswith(".npz")
    ]
    if not shard_files:
        raise FileNotFoundError("No shard_*.npz files found in the data folder.")

    shard_path = random.choice(shard_files)
    frames_list, w_list, h_list = load_shard(shard_path)

    idx = random.randrange(len(frames_list))
    frames = frames_list[idx]

    images = []
    max_frames = min(args.max_frames, frames.shape[0])
    for t in range(max_frames):
        images.append(to_gray(frames[t], args.alphabet))

    imageio.mimsave(args.out, images, fps=20)
    print(f"Saved: {args.out}")


if __name__ == "__main__":
    main()