| |
| """Generate a small CLEVRER-lite dataset (GIF videos + JSON annotations). |
| |
| Examples: |
| python -m clevrer_lite.generate_dataset --num-videos 8 --num-frames 64 --frame-size 128 |
| python -m clevrer_lite.generate_dataset --out data/big --num-videos 1000 --num-frames 128 |
| """ |
|
|
| import argparse |
| import json |
| import os |
| import time |
|
|
| import numpy as np |
| from PIL import Image |
|
|
| from . import CLEVRERLiteEnv, config, simulate |
|
|
|
|
| def describe_events(spec, ev): |
| """Human-readable event descriptions for the annotation file.""" |
| objs = {o["idx"]: o for o in spec} |
|
|
| def ref(i): |
| return "the {} {}".format(objs[i]["color"], objs[i]["shape"]) |
|
|
| collisions = [{ |
| "frame": e["frame"], |
| "objects": e["objects"], |
| "description": f"{ref(e['objects'][0])} collided with {ref(e['objects'][1])}", |
| } for e in ev["collisions"]] |
|
|
| motion = [] |
| for o, sts in ev["starts"].items(): |
| for s in sts: |
| cause = None |
| for e in ev["collisions"]: |
| a, b = e["objects"] |
| if o in (a, b) and abs(e["frame"] - s) <= config.CAUSE_WINDOW_FRAMES: |
| cause = e |
| break |
| motion.append({"object": o, "type": "start", "frame": s, |
| "cause_collision": cause}) |
| return {"collision_events": collisions, "motion_events": motion, |
| "initially_moving": ev["initially_moving"]} |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser(description="Generate a CLEVRER-lite dataset") |
| ap.add_argument("--out", default="output") |
| ap.add_argument("--num-videos", type=int, default=8) |
| ap.add_argument("--num-frames", type=int, default=64) |
| ap.add_argument("--frame-size", type=int, default=128) |
| ap.add_argument("--seed", type=int, default=0) |
| ap.add_argument("--gif-fps", type=int, default=12) |
| ap.add_argument("--max-questions", type=int, default=8) |
| args = ap.parse_args() |
|
|
| video_dir = os.path.join(args.out, "videos") |
| os.makedirs(video_dir, exist_ok=True) |
|
|
| env = CLEVRERLiteEnv(num_frames=args.num_frames, frame_size=args.frame_size) |
| annotations = [] |
| t0 = time.time() |
|
|
| for v in range(args.num_videos): |
| obs, info = env.reset(seed=args.seed + v) |
| frames = [obs] |
| while True: |
| obs, reward, terminated, truncated, info = env.step(0) |
| frames.append(obs) |
| if truncated: |
| break |
|
|
| ev = env.extract_events() |
| qs = env.build_questions(ev, max_questions=args.max_questions) |
| ev_public = {k: v for k, v in ev.items() if k not in ("moving", "speeds")} |
|
|
| fname = f"video_{v:05d}.gif" |
| imgs = [Image.fromarray(f) for f in frames] |
| imgs[0].save(os.path.join(video_dir, fname), save_all=True, |
| append_images=imgs[1:], duration=int(1000 / args.gif_fps), |
| loop=0) |
|
|
| record = { |
| "video_index": v, |
| "video_filename": f"videos/{fname}", |
| "num_frames": args.num_frames, |
| "frame_size": args.frame_size, |
| "fps": env.metadata["render_fps"], |
| "scene": {"objects": env.scene}, |
| "events": describe_events(env.scene, ev), |
| "raw_events": { |
| "collisions": ev_public["collisions"], |
| "starts": {str(k): v for k, v in ev_public["starts"].items()}, |
| "stops": {str(k): v for k, v in ev_public["stops"].items()}, |
| "initially_moving": ev_public["initially_moving"], |
| "wall_hits": [list(map(int, w)) for w in ev_public["wall_hits"]], |
| }, |
| "questions": qs, |
| } |
| annotations.append(record) |
| n_col = len(ev["collisions"]) |
| print(f"[{v + 1}/{args.num_videos}] {fname}: " |
| f"{len(env.scene)} objects, {n_col} collisions, {len(qs)} questions") |
|
|
| with open(os.path.join(args.out, "annotations.json"), "w") as f: |
| json.dump(annotations, f, indent=2) |
|
|
| dt = time.time() - t0 |
| print(f"\nDone: {args.num_videos} videos in {dt:.1f}s " |
| f"({dt / args.num_videos:.2f}s/video) -> {args.out}/") |
| print(f"Total questions: {sum(len(a['questions']) for a in annotations)}") |
|
|
| |
| from collections import Counter |
| types = Counter(q["question_type"] for a in annotations for q in a["questions"]) |
| print("Question types:", dict(types)) |
| answers = Counter(q["answer"] for a in annotations for q in a["questions"]) |
| print("Top answers:", dict(answers.most_common(8))) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
