Download scripts/fivehundredshots.py from catachiii/FiveHundredShots: direct link, hf CLI and curl.
- Browser
- Download file 4.6 kB
-
https://huggingface.co/datasets/catachiii/FiveHundredShots/resolve/main/scripts/fivehundredshots.py
- Command line
-
hf download hf://datasets/catachiii/FiveHundredShots/scripts/fivehundredshots.py
-
curl -L -o fivehundredshots.py https://huggingface.co/datasets/catachiii/FiveHundredShots/resolve/main/scripts/fivehundredshots.py
4.6 kB
| """Loader for the FiveHundredShots dataset (needs only numpy). | |
| A clip is a float32 array of shape (T, 281) sampled at 50 Hz, world frame z-up: | |
| columns field shape notes | |
| 0:3 root_pos (T, 3) pelvis position [m] | |
| 3:7 root_quat (T, 4) pelvis orientation (w, x, y, z) | |
| 7:36 joint_pos (T, 29) [rad], order = JOINT_NAMES | |
| 36:39 root_lin_vel (T, 3) world frame [m/s] | |
| 39:42 root_ang_vel (T, 3) world frame [rad/s] | |
| 42:71 joint_vel (T, 29) [rad/s], order = JOINT_NAMES | |
| 71:281 body_pos/quat (T, 30, 7) per body: position (3) + quaternion wxyz (4), | |
| world frame, order = BODY_NAMES | |
| python scripts/fivehundredshots.py # dataset summary | |
| python scripts/fivehundredshots.py shot_0000 # one clip | |
| """ | |
| from __future__ import annotations | |
| import json | |
| import sys | |
| from dataclasses import dataclass | |
| from pathlib import Path | |
| import numpy as np | |
| ROOT = Path(__file__).resolve().parent.parent | |
| FPS = 50 | |
| FRAME_DIM = 281 | |
| # Breadth-first joint order (the one Isaac Lab reports for the G1); MuJoCo orders joints differently. | |
| JOINT_NAMES = [ | |
| "left_hip_pitch_joint", "right_hip_pitch_joint", "waist_yaw_joint", | |
| "left_hip_roll_joint", "right_hip_roll_joint", "waist_roll_joint", | |
| "left_hip_yaw_joint", "right_hip_yaw_joint", "waist_pitch_joint", | |
| "left_knee_joint", "right_knee_joint", | |
| "left_shoulder_pitch_joint", "right_shoulder_pitch_joint", | |
| "left_ankle_pitch_joint", "right_ankle_pitch_joint", | |
| "left_shoulder_roll_joint", "right_shoulder_roll_joint", | |
| "left_ankle_roll_joint", "right_ankle_roll_joint", | |
| "left_shoulder_yaw_joint", "right_shoulder_yaw_joint", | |
| "left_elbow_joint", "right_elbow_joint", | |
| "left_wrist_roll_joint", "right_wrist_roll_joint", | |
| "left_wrist_pitch_joint", "right_wrist_pitch_joint", | |
| "left_wrist_yaw_joint", "right_wrist_yaw_joint", | |
| ] | |
| BODY_NAMES = [ | |
| "pelvis", "left_hip_pitch_link", "right_hip_pitch_link", "waist_yaw_link", | |
| "left_hip_roll_link", "right_hip_roll_link", "waist_roll_link", | |
| "left_hip_yaw_link", "right_hip_yaw_link", "torso_link", | |
| "left_knee_link", "right_knee_link", | |
| "left_shoulder_pitch_link", "right_shoulder_pitch_link", | |
| "left_ankle_pitch_link", "right_ankle_pitch_link", | |
| "left_shoulder_roll_link", "right_shoulder_roll_link", | |
| "left_ankle_roll_link", "right_ankle_roll_link", | |
| "left_shoulder_yaw_link", "right_shoulder_yaw_link", | |
| "left_elbow_link", "right_elbow_link", | |
| "left_wrist_roll_link", "right_wrist_roll_link", | |
| "left_wrist_pitch_link", "right_wrist_pitch_link", | |
| "left_wrist_yaw_link", "right_wrist_yaw_link", | |
| ] | |
| class Clip: | |
| root_pos: np.ndarray # (T, 3) | |
| root_quat: np.ndarray # (T, 4) wxyz | |
| joint_pos: np.ndarray # (T, 29) | |
| root_lin_vel: np.ndarray # (T, 3) world | |
| root_ang_vel: np.ndarray # (T, 3) world | |
| joint_vel: np.ndarray # (T, 29) | |
| body_pos: np.ndarray # (T, 30, 3) world | |
| body_quat: np.ndarray # (T, 30, 4) wxyz, world | |
| def __len__(self) -> int: | |
| return len(self.root_pos) | |
| def load_clip(path) -> Clip: | |
| """Split a (T, 281) clip into named arrays (views of one array, no copies).""" | |
| raw = np.load(path) | |
| if raw.ndim != 2 or raw.shape[1] != FRAME_DIM: | |
| raise ValueError(f"{path}: expected shape (T, {FRAME_DIM}), got {raw.shape}") | |
| body = raw[:, 71:].reshape(len(raw), len(BODY_NAMES), 7) | |
| return Clip(raw[:, 0:3], raw[:, 3:7], raw[:, 7:36], raw[:, 36:39], raw[:, 39:42], raw[:, 42:71], | |
| body[..., :3], body[..., 3:]) | |
| def load_metadata(root=ROOT) -> list[dict]: | |
| """One dict per shot, in file order (see the dataset card for the fields).""" | |
| with open(Path(root) / "metadata.jsonl") as f: | |
| return [json.loads(line) for line in f] | |
| def main(): | |
| shots = load_metadata() | |
| if len(sys.argv) > 1: | |
| shot = next((s for s in shots if s["id"] == sys.argv[1]), None) | |
| if shot is None: | |
| raise SystemExit(f"unknown shot id: {sys.argv[1]}") | |
| clip = load_clip(ROOT / shot["file"]) | |
| print(json.dumps(shot, indent=1)) | |
| for name, value in vars(clip).items(): | |
| print(f"{name:13s} {value.shape}") | |
| return | |
| frames = sum(s["n_frames"] for s in shots) | |
| strokes = {k: sum(s["stroke"] == k for s in shots) for k in ("forehand", "backhand")} | |
| print(f"{len(shots)} shots, {frames} frames ({frames / FPS / 60:.1f} min at {FPS} Hz), {strokes}") | |
| if __name__ == "__main__": | |
| main() | |