scripts/check_pick_box_dataset.py

#!/usr/bin/env python3
import argparse
import json
import random
import sys
from pathlib import Path

import numpy as np
import pyarrow.parquet as pq

PROJECT_ROOT = Path(__file__).resolve().parents[1]
if str(PROJECT_ROOT) not in sys.path:
    sys.path.insert(0, str(PROJECT_ROOT))


STATE_COLS = [
    "observation.arm_joints",
    "observation.hand_joints",
    "observation.prev_height",
    "observation.prev_rpy",
    "observation.prev_vx",
    "observation.prev_vy",
    "observation.prev_vyaw",
    "observation.prev_dyaw",
]


def sample_rows(parquet_path: Path, max_rows: int, rng: random.Random):
    table = pq.read_table(parquet_path, columns=["action", *STATE_COLS])
    total = table.num_rows
    if total == 0:
        return None
    if max_rows >= total:
        indices = list(range(total))
    else:
        indices = rng.sample(range(total), max_rows)
    sub = table.take(indices)
    actions = np.array(sub.column("action").to_pylist(), dtype=np.float32)
    arms = np.array(sub.column("observation.arm_joints").to_pylist(), dtype=np.float32)
    hands = np.array(sub.column("observation.hand_joints").to_pylist(), dtype=np.float32)
    prev_h = np.array(sub.column("observation.prev_height").to_pylist(), dtype=np.float32).reshape(-1, 1)
    prev_rpy = np.array(sub.column("observation.prev_rpy").to_pylist(), dtype=np.float32)
    prev_vx = np.array(sub.column("observation.prev_vx").to_pylist(), dtype=np.float32).reshape(-1, 1)
    prev_vy = np.array(sub.column("observation.prev_vy").to_pylist(), dtype=np.float32).reshape(-1, 1)
    prev_vyaw = np.array(sub.column("observation.prev_vyaw").to_pylist(), dtype=np.float32).reshape(-1, 1)
    prev_dyaw = np.array(sub.column("observation.prev_dyaw").to_pylist(), dtype=np.float32).reshape(-1, 1)
    state = np.concatenate(
        [arms, hands, prev_h, prev_rpy, prev_vx, prev_vy, prev_vyaw, prev_dyaw], axis=1
    )
    return actions, state


def main() -> None:
    parser = argparse.ArgumentParser(description="Quick sanity checks for pick_box dataset.")
    parser.add_argument("--data_root", default="/hfm/data/pick_box")
    parser.add_argument("--max_episodes", type=int, default=10)
    parser.add_argument("--rows_per_episode", type=int, default=200)
    parser.add_argument("--seed", type=int, default=0)
    args = parser.parse_args()

    rng = random.Random(args.seed)
    data_root = Path(args.data_root)
    meta = data_root / "meta"
    info_path = meta / "info.json"
    episodes_path = meta / "episodes.jsonl"

    if not info_path.exists() or not episodes_path.exists():
        raise FileNotFoundError("Missing meta/info.json or meta/episodes.jsonl.")

    info = json.loads(info_path.read_text())
    episodes = [json.loads(line) for line in episodes_path.read_text().splitlines() if line.strip()]

    print("Dataset root:", data_root)
    print("Total episodes (meta):", info.get("total_episodes"))
    print("Episodes listed:", len(episodes))
    lengths = [int(e.get("length", 0)) for e in episodes]
    print("Episode length: min", min(lengths), "max", max(lengths), "avg", np.mean(lengths))

    instructions = [str(e.get("instruction", "") or "").strip() for e in episodes]
    empty_instr = sum(1 for i in instructions if not i)
    print("Unique instructions:", len(set(instructions)))
    print("Empty instructions:", empty_instr)

    data_path_tpl = info["data_path"]
    chunk_size = int(info["chunks_size"])
    actions_all = []
    states_all = []

    for ep in episodes[: args.max_episodes]:
        ep_index = int(ep["episode_index"])
        chunk = ep_index // chunk_size
        parquet_path = data_root / data_path_tpl.format(
            episode_chunk=chunk, episode_index=ep_index
        )
        if not parquet_path.exists():
            print("Missing parquet:", parquet_path)
            continue
        sample = sample_rows(parquet_path, args.rows_per_episode, rng)
        if sample is None:
            continue
        actions, states = sample
        actions_all.append(actions)
        states_all.append(states)

    if not actions_all:
        print("No action samples collected. Check dataset paths.")
        return

    actions = np.concatenate(actions_all, axis=0)
    states = np.concatenate(states_all, axis=0)

    print("Sampled actions shape:", actions.shape)
    print("Sampled states shape:", states.shape)
    print("Action dim:", actions.shape[1])
    print("State dim:", states.shape[1])
    print("Action NaN:", np.isnan(actions).any(), "Inf:", np.isinf(actions).any())
    print("State NaN:", np.isnan(states).any(), "Inf:", np.isinf(states).any())

    action_std = actions.std(axis=0)
    state_std = states.std(axis=0)
    print("Action std (min/median/max):", action_std.min(), np.median(action_std), action_std.max())
    print("State std (min/median/max):", state_std.min(), np.median(state_std), state_std.max())

    near_zero = np.mean(np.abs(actions) < 1e-4, axis=0)
    print("Action near-zero ratio (min/median/max):", near_zero.min(), np.median(near_zero), near_zero.max())

    stats_path = meta / "action_stats.json"
    if stats_path.exists():
        stats = json.loads(stats_path.read_text())
        min_stat = np.array(stats["min"], dtype=np.float32)
        max_stat = np.array(stats["max"], dtype=np.float32)
        print("Stats file min/max (first 8 dims):")
        print("  min:", min_stat[:8])
        print("  max:", max_stat[:8])


if __name__ == "__main__":
    main()