from __future__ import annotations import argparse import json from pathlib import Path import numpy as np import pandas as pd from tqdm import tqdm from psi.config.transform import SeedPosttrainRepackTransform from psi.data.seed.motion_repr import build_seed_textop_action from psi.data.seed.seed_motion_dataset import SeedMotionDataset def stats_dict(array: np.ndarray) -> dict[str, list[float]]: return { "mean": np.mean(array, axis=0).tolist(), "std": np.std(array, axis=0).tolist(), "min": np.min(array, axis=0).tolist(), "max": np.max(array, axis=0).tolist(), "q01": np.quantile(array, 0.01, axis=0).tolist(), "q99": np.quantile(array, 0.99, axis=0).tolist(), } def stats_from_sample( *, count: int, total: np.ndarray, square_total: np.ndarray, min_value: np.ndarray, max_value: np.ndarray, quantile_sample: np.ndarray, ) -> dict[str, list[float]]: mean = total / float(count) variance = square_total / float(count) - mean * mean variance = np.maximum(variance, 0.0) return { "mean": mean.tolist(), "std": np.sqrt(variance).tolist(), "min": min_value.tolist(), "max": max_value.tolist(), "q01": np.quantile(quantile_sample, 0.01, axis=0).tolist(), "q99": np.quantile(quantile_sample, 0.99, axis=0).tolist(), } def take_sample_rows(array: np.ndarray, sample_indices: np.ndarray, start: int, end: int) -> np.ndarray: left = np.searchsorted(sample_indices, start, side="left") right = np.searchsorted(sample_indices, end, side="left") if right <= left: return np.zeros((0, array.shape[1]), dtype=np.float32) local_indices = sample_indices[left:right] - start return array[local_indices].astype(np.float32) def map_seed_motion_dir(root_dir: Path, move_g1_path: str) -> Path: rel = str(move_g1_path).strip().replace("\\", "/") rel = rel.replace("g1/csv/", "g1/csv_sonic/") rel = rel.replace("/csv/", "/csv_sonic/") if rel.endswith(".csv"): rel = rel[:-4] if rel.startswith("dataset/seed/"): rel = rel[len("dataset/seed/") :] if rel.startswith("seed/"): rel = rel[len("seed/") :] p = Path(rel) if p.is_absolute(): return p return root_dir / p def load_seed_motion_array(motion_dir: Path) -> np.ndarray: joint = pd.read_csv(motion_dir / "joint_pos.csv").to_numpy(dtype=np.float32) body_pos = pd.read_csv(motion_dir / "body_pos.csv").to_numpy(dtype=np.float32) body_quat = pd.read_csv(motion_dir / "body_quat.csv").to_numpy(dtype=np.float32) length = min(joint.shape[0], body_pos.shape[0], body_quat.shape[0]) return np.concatenate([joint[:length], body_pos[:length], body_quat[:length]], axis=1).astype(np.float32) def count_csv_data_rows(path: Path) -> int: with path.open("r", encoding="utf-8", errors="ignore") as f: return max(0, sum(1 for _ in f) - 1) def calculate_frame_level_stats_from_metadata( *, root_dir: str, metadata_csv: str, action_chunk_size: int, num_past_frames: int, quantile_sample_size: int, seed: int, ) -> dict: root = Path(root_dir) metadata_path = Path(metadata_csv) if not metadata_path.is_absolute(): metadata_path = root / metadata_path metadata = pd.read_csv(metadata_path, usecols=["move_g1_path"]) motion_dirs = [map_seed_motion_dir(root, p) for p in metadata["move_g1_path"].tolist()] lengths = [] valid_motion_dirs = [] for motion_dir in tqdm(motion_dirs, desc="Counting SEED frame rows"): joint_path = motion_dir / "joint_pos.csv" if not joint_path.exists(): continue n = count_csv_data_rows(joint_path) if n <= 0: continue lengths.append(n) valid_motion_dirs.append(motion_dir) total_action_rows = int(sum(max(0, n - 1) for n in lengths)) total_state_rows = int(sum(lengths)) rng = np.random.RandomState(seed) action_sample_size = min(int(quantile_sample_size), total_action_rows) state_sample_size = min(int(quantile_sample_size), total_state_rows) action_sample_indices = np.sort(rng.choice(total_action_rows, size=action_sample_size, replace=False)) state_sample_indices = np.sort(rng.choice(total_state_rows, size=state_sample_size, replace=False)) action_sum = np.zeros((70,), dtype=np.float64) action_sumsq = np.zeros((70,), dtype=np.float64) action_min = np.full((70,), np.inf, dtype=np.float64) action_max = np.full((70,), -np.inf, dtype=np.float64) state_sum = np.zeros((36,), dtype=np.float64) state_sumsq = np.zeros((36,), dtype=np.float64) state_min = np.full((36,), np.inf, dtype=np.float64) state_max = np.full((36,), -np.inf, dtype=np.float64) action_samples = [] state_samples = [] action_offset = 0 state_offset = 0 for motion_dir in tqdm(valid_motion_dirs, desc="Collecting frame-level SEED TextOp stats"): motion = load_seed_motion_array(motion_dir) if motion.shape[0] > 1: textop_action = build_seed_textop_action(motion[1:], motion[:1]) action_sum += textop_action.sum(axis=0, dtype=np.float64) action_sumsq += np.square(textop_action.astype(np.float64)).sum(axis=0) action_min = np.minimum(action_min, textop_action.min(axis=0)) action_max = np.maximum(action_max, textop_action.max(axis=0)) action_end = action_offset + textop_action.shape[0] sampled = take_sample_rows(textop_action, action_sample_indices, action_offset, action_end) if sampled.shape[0] > 0: action_samples.append(sampled) action_offset = action_end state_sum += motion.sum(axis=0, dtype=np.float64) state_sumsq += np.square(motion.astype(np.float64)).sum(axis=0) state_min = np.minimum(state_min, motion.min(axis=0)) state_max = np.maximum(state_max, motion.max(axis=0)) state_end = state_offset + motion.shape[0] sampled = take_sample_rows(motion, state_sample_indices, state_offset, state_end) if sampled.shape[0] > 0: state_samples.append(sampled) state_offset = state_end action_sample = np.concatenate(action_samples, axis=0).astype(np.float32) state_sample = np.concatenate(state_samples, axis=0).astype(np.float32) return { "action": stats_from_sample( count=total_action_rows, total=action_sum, square_total=action_sumsq, min_value=action_min, max_value=action_max, quantile_sample=action_sample, ), "states": stats_from_sample( count=total_state_rows, total=state_sum, square_total=state_sumsq, min_value=state_min, max_value=state_max, quantile_sample=state_sample, ), "meta": { "representation": "joint_abs29_joint_delta29_root_delta_local3_root_rot6d_delta6_root_height1_foot_contact2", "stat_mode": "frame_level_direct_metadata", "action_dim": 70, "state_dim": 36, "action_chunk_size": int(action_chunk_size), "num_past_frames": int(num_past_frames), "num_metadata_rows": int(len(metadata)), "num_motions": int(len(valid_motion_dirs)), "num_action_rows": int(total_action_rows), "num_state_rows": int(total_state_rows), "action_quantile_sample_rows": int(action_sample.shape[0]), "state_quantile_sample_rows": int(state_sample.shape[0]), "quantile_sample_seed": int(seed), }, } def calculate_frame_level_stats( dataset: SeedMotionDataset, *, action_chunk_size: int, num_past_frames: int, quantile_sample_size: int, seed: int, ) -> dict: total_action_rows = int(sum(max(0, int(r["n_frames"]) - 1) for r in dataset.motion_records)) total_state_rows = int(sum(int(r["n_frames"]) for r in dataset.motion_records)) rng = np.random.RandomState(seed) action_sample_size = min(int(quantile_sample_size), total_action_rows) state_sample_size = min(int(quantile_sample_size), total_state_rows) action_sample_indices = np.sort(rng.choice(total_action_rows, size=action_sample_size, replace=False)) state_sample_indices = np.sort(rng.choice(total_state_rows, size=state_sample_size, replace=False)) action_sum = np.zeros((70,), dtype=np.float64) action_sumsq = np.zeros((70,), dtype=np.float64) action_min = np.full((70,), np.inf, dtype=np.float64) action_max = np.full((70,), -np.inf, dtype=np.float64) state_sum = np.zeros((36,), dtype=np.float64) state_sumsq = np.zeros((36,), dtype=np.float64) state_min = np.full((36,), np.inf, dtype=np.float64) state_max = np.full((36,), -np.inf, dtype=np.float64) action_samples = [] state_samples = [] action_offset = 0 state_offset = 0 for record in tqdm(dataset.motion_records, desc="Collecting frame-level SEED TextOp stats"): motion = dataset._load_motion_array(record["motion_dir"]) motion = np.asarray(motion[:, :36], dtype=np.float32) if motion.shape[0] > 1: textop_action = build_seed_textop_action(motion[1:], motion[:1]) action_sum += textop_action.sum(axis=0, dtype=np.float64) action_sumsq += np.square(textop_action.astype(np.float64)).sum(axis=0) action_min = np.minimum(action_min, textop_action.min(axis=0)) action_max = np.maximum(action_max, textop_action.max(axis=0)) action_end = action_offset + textop_action.shape[0] sampled = take_sample_rows(textop_action, action_sample_indices, action_offset, action_end) if sampled.shape[0] > 0: action_samples.append(sampled) action_offset = action_end state_sum += motion.sum(axis=0, dtype=np.float64) state_sumsq += np.square(motion.astype(np.float64)).sum(axis=0) state_min = np.minimum(state_min, motion.min(axis=0)) state_max = np.maximum(state_max, motion.max(axis=0)) state_end = state_offset + motion.shape[0] sampled = take_sample_rows(motion, state_sample_indices, state_offset, state_end) if sampled.shape[0] > 0: state_samples.append(sampled) state_offset = state_end action_sample = np.concatenate(action_samples, axis=0).astype(np.float32) state_sample = np.concatenate(state_samples, axis=0).astype(np.float32) return { "action": stats_from_sample( count=total_action_rows, total=action_sum, square_total=action_sumsq, min_value=action_min, max_value=action_max, quantile_sample=action_sample, ), "states": stats_from_sample( count=total_state_rows, total=state_sum, square_total=state_sumsq, min_value=state_min, max_value=state_max, quantile_sample=state_sample, ), "meta": { "representation": "joint_abs29_joint_delta29_root_delta_local3_root_rot6d_delta6_root_height1_foot_contact2", "stat_mode": "frame_level", "action_dim": 70, "state_dim": 36, "action_chunk_size": int(action_chunk_size), "num_past_frames": int(num_past_frames), "num_motions": int(len(dataset.motion_records)), "num_action_rows": int(total_action_rows), "num_state_rows": int(total_state_rows), "action_quantile_sample_rows": int(action_sample.shape[0]), "state_quantile_sample_rows": int(state_sample.shape[0]), "quantile_sample_seed": int(seed), }, } def main() -> None: parser = argparse.ArgumentParser(description="Calculate SEED TextOp-style action statistics.") parser.add_argument("--root-dir", type=str, required=True) parser.add_argument("--metadata-csv", type=str, default="metadata/seed_metadata_v004_100.csv") parser.add_argument("--output-path", type=str, required=True) parser.add_argument("--action-chunk-size", type=int, default=30) parser.add_argument("--num-past-frames", type=int, default=3) parser.add_argument("--frame-level", action="store_true", default=False) parser.add_argument("--direct-metadata", action="store_true", default=False) parser.add_argument("--quantile-sample-size", type=int, default=1000000) parser.add_argument("--seed", type=int, default=292285) args = parser.parse_args() if args.frame_level and args.direct_metadata: stats = calculate_frame_level_stats_from_metadata( root_dir=args.root_dir, metadata_csv=args.metadata_csv, action_chunk_size=args.action_chunk_size, num_past_frames=args.num_past_frames, quantile_sample_size=args.quantile_sample_size, seed=args.seed, ) output_path = Path(args.output_path) output_path.parent.mkdir(parents=True, exist_ok=True) output_path.write_text(json.dumps(stats, indent=2), encoding="utf-8") print(f"Wrote {output_path}") print(json.dumps(stats["meta"], indent=2)) return dataset = SeedMotionDataset( root_dir=args.root_dir, metadata_csv=args.metadata_csv, split="all", action_chunk_size=args.action_chunk_size, num_past_frames=args.num_past_frames, ) if args.frame_level: stats = calculate_frame_level_stats( dataset, action_chunk_size=args.action_chunk_size, num_past_frames=args.num_past_frames, quantile_sample_size=args.quantile_sample_size, seed=args.seed, ) output_path = Path(args.output_path) output_path.parent.mkdir(parents=True, exist_ok=True) output_path.write_text(json.dumps(stats, indent=2), encoding="utf-8") print(f"Wrote {output_path}") print(json.dumps(stats["meta"], indent=2)) return repack = SeedPosttrainRepackTransform( textop_actions=True, action_chunk_size=args.action_chunk_size, num_past_frames=args.num_past_frames, pad_action_dim=70, pad_state_dim=36, ) actions = [] states = [] for i in tqdm(range(len(dataset)), desc="Collecting SEED TextOp stats"): item = repack(dataset[i]) # 只统计有效action帧,避免chunk尾部padding重复最后一帧影响q01/q99。 action_mask = np.asarray(item["actions_mask"], dtype=np.float32) valid_rows = np.any(action_mask > 0.5, axis=1) actions.append(item["actions"][valid_rows].reshape(-1, item["actions"].shape[-1])) states.append(item["states"].reshape(-1, item["states"].shape[-1])) action_array = np.concatenate(actions, axis=0).astype(np.float32) state_array = np.concatenate(states, axis=0).astype(np.float32) stats = { "action": stats_dict(action_array), "states": stats_dict(state_array), "meta": { "representation": "joint_abs29_joint_delta29_root_delta_local3_root_rot6d_delta6_root_height1_foot_contact2", "stat_mode": "chunk_window", "action_dim": 70, "state_dim": 36, "action_chunk_size": int(args.action_chunk_size), "num_past_frames": int(args.num_past_frames), "num_samples": int(len(dataset)), "num_action_rows": int(action_array.shape[0]), "num_state_rows": int(state_array.shape[0]), }, } output_path = Path(args.output_path) output_path.parent.mkdir(parents=True, exist_ok=True) output_path.write_text(json.dumps(stats, indent=2), encoding="utf-8") print(f"Wrote {output_path}") print(f"action shape: {action_array.shape}, states shape: {state_array.shape}") if __name__ == "__main__": main()