#!/usr/bin/env python3 """ LeRobot Dataset Visualizer — Web Dashboard 运行: python visualize_dataset.py -d 浏览器访问 http://localhost:7777 """ import sys, json, argparse, base64, threading, webbrowser, time from urllib.parse import urlparse, parse_qs from http.server import HTTPServer, BaseHTTPRequestHandler try: from scipy.ndimage import median_filter as _scipy_medfilt from scipy.ndimage import gaussian_filter as _scipy_gauss _HAS_SCIPY = True except ImportError: _HAS_SCIPY = False def _filter_tactile(vals, shape): """ 两步滤波： 1. 两遍 3×3 中值滤波：彻底消除孤立单点和双点噪声（数学保证：孤立点被邻域中值替换）。 2. 轻度高斯平滑（sigma=0.7）。 """ if not _HAS_SCIPY or not vals: return vals try: arr = np.array(vals, dtype=np.float32).reshape(shape) # 两遍中值滤波：第一遍消单点，第二遍消残余双点 arr = _scipy_medfilt(arr, size=3, mode='nearest') arr = _scipy_medfilt(arr, size=3, mode='nearest') # 高斯平滑 arr = _scipy_gauss(arr, sigma=0.7, mode='nearest') return arr.flatten().tolist() except Exception: return vals from pathlib import Path import numpy as np try: import pyarrow as pa import pyarrow.parquet as pq except ImportError: sys.exit("缺少 pyarrow") try: import cv2 _HAS_CV2 = True except ImportError: _HAS_CV2 = False PORT = 7777 _DS = None # 触觉传感器分组（基于 id 映射表） # 拇指 3 片(0-2)；食/中/无名/小指各 4 片（指尖→中→尾→掌内关节）；掌心 1 片(19) SENSOR_GROUPS = [ {"id":"thumb", "label":"拇指", "color":"#ff6b6b", "type":"finger", "sensors":["tactile_0","tactile_1","tactile_2"]}, {"id":"index", "label":"食指", "color":"#ffd93d", "type":"finger", "sensors":["tactile_3","tactile_4","tactile_5","tactile_6"]}, {"id":"middle", "label":"中指", "color":"#6bcb77", "type":"finger", "sensors":["tactile_7","tactile_8","tactile_9","tactile_10"]}, {"id":"ring", "label":"无名", "color":"#4d96ff", "type":"finger", "sensors":["tactile_11","tactile_12","tactile_13","tactile_14"]}, {"id":"pinky", "label":"小指", "color":"#c77dff", "type":"finger", "sensors":["tactile_15","tactile_16","tactile_17","tactile_18"]}, {"id":"palm", "label":"掌心", "color":"#9ecbff", "type":"palm", "sensors":["tactile_19"]}, ] # ══════════════════════════════════════════════════════════════════════════════ class DataServer: def __init__(self, root: str): self.root = Path(root) self._load_info(); self._load_episodes() self._load_frames(); self._load_tasks(); self._load_cmap() self._compute_tactile_baselines() def _load_info(self): with open(self.root/"meta"/"info.json", encoding="utf-8") as f: self.info = json.load(f) self.fps = self.info["fps"] self.tactile_2d_shapes = self.info.get("tactile_2d_shapes", {}) print(f"OK {self.root.name} fps={self.fps} frames={self.info['total_frames']}") def _load_episodes(self): files = sorted((self.root/"meta"/"episodes").rglob("*.parquet")) ep_dict = pa.concat_tables([pq.read_table(f) for f in files]).to_pydict() n = len(ep_dict["episode_index"]) self.episodes = [{k: ep_dict[k][i] for k in ep_dict} for i in range(n)] self.n_episodes = n def _load_frames(self): files = sorted((self.root/"data").rglob("*.parquet")) fd = pa.concat_tables([pq.read_table(f) for f in files]).to_pydict() self.frame_dict = fd self._idx_map = {int(v): i for i, v in enumerate(fd["index"])} def _load_tasks(self): try: tp = next((self.root/"meta").rglob("tasks.parquet"), None) self.task_desc = pq.read_table(tp).to_pydict().get("task",["Unknown"])[0] if tp else "Unknown" except Exception: self.task_desc = "Unknown" def _load_cmap(self): cfg = (self.root.parent.parent / "hand"/"Usb_API_stable_Tujian_"/"spliter"/"config_mapping_zyhand.json") if cfg.exists(): with open(cfg, encoding='utf-8') as f: data = json.load(f) self.cmap_colors = [ f"#{int(c[0]*255):02x}{int(c[1]*255):02x}{int(c[2]*255):02x}" for c in data["color_map"]] else: self.cmap_colors = [ "#262626","#0a2a5e","#0d4f8b","#0d7ab5","#18a0c8", "#1dc9d0","#22e0be","#39e87d","#8aef3f","#d4e830", "#f0c020","#f0901a","#e05015","#c02510","#8b0d08","#5e0505","#3d0202"] def _compute_tactile_baselines(self): """每个 episode 前 10 帧取均值作为触觉零点基准。""" self.tactile_baselines = {} for ep_id, ep in enumerate(self.episodes): start = int(ep["dataset_from_index"]) end = int(ep["dataset_to_index"]) - 1 sums, count = {}, 0 for fidx in range(start, min(start + 10, end + 1)): i = self._idx_map.get(fidx) if i is None: continue for key in self.tactile_2d_shapes: raw = self.frame_dict.get(f"tactile_{key}") if raw is None or i >= len(raw) or raw[i] is None: continue arr = [float(v) for v in raw[i]] if key not in sums: sums[key] = [0.0] * len(arr) for j, v in enumerate(arr): sums[key][j] += v count += 1 self.tactile_baselines[ep_id] = ( {k: [v / count for v in sums[k]] for k in sums} if count else {} ) print(f"触觉基准已计算 episodes={len(self.tactile_baselines)}") def get_tactile_timeline(self, ep_id: int) -> dict: ep = self.episodes[ep_id] start = int(ep["dataset_from_index"]) end = int(ep["dataset_to_index"]) - 1 baseline = self.tactile_baselines.get(ep_id, {}) sensors_data = [] for grp in SENSOR_GROUPS: for sensor_key in grp["sensors"]: # parquet列名是 "tactile_tactile_0"，sensor_key是 "tactile_0" col_key = f"tactile_{sensor_key}" sums = [] for fidx in range(start, end + 1): i = self._idx_map.get(fidx) if i is None: sums.append(0.0); continue raw_col = self.frame_dict.get(col_key) if raw_col is None or i >= len(raw_col) or raw_col[i] is None: sums.append(0.0); continue base = baseline.get(sensor_key, []) vals = [float(v) for v in raw_col[i]] if base: vals = [max(0.0, v - b) for v, b in zip(vals, base)] shape = self.tactile_2d_shapes.get(sensor_key) if shape: vals = _filter_tactile(vals, shape) sums.append(round(sum(vals), 2)) sensors_data.append({ "key": sensor_key, "color": grp["color"], "sums": sums, }) return {"sensors": sensors_data, "n_frames": end - start + 1, "start": start} def get_frame(self, idx: int) -> dict: i = self._idx_map.get(idx, idx) return {k: self.frame_dict[k][i] for k in self.frame_dict} def get_video_frame_b64(self, ep_idx, frame_in_ep, stream): if not _HAS_CV2: return None ep = self.episodes[ep_idx] vk = f"observation.images.{stream}" chunk = int(ep.get(f"videos/{vk}/chunk_index", 0)) fidx = int(ep.get(f"videos/{vk}/file_index", 0)) from_ts = float(ep.get(f"videos/{vk}/from_timestamp", 0.0)) vpath = (self.root/f"videos/{vk}/chunk-{chunk:03d}/file-{fidx:04d}.mp4") if not vpath.exists(): return None cap = cv2.VideoCapture(str(vpath)) cap.set(cv2.CAP_PROP_POS_FRAMES, int(round(from_ts*self.fps)) + frame_in_ep) ret, frame = cap.read(); cap.release() if not ret: return None if stream == "depth": r, g, b = frame[:,:,2], frame[:,:,1], frame[:,:,0] if np.abs(r.astype(int)-g.astype(int)).mean() < 4: frame = cv2.applyColorMap(r, cv2.COLORMAP_TURBO) _, buf = cv2.imencode(".jpg", frame, [cv2.IMWRITE_JPEG_QUALITY, 88]) return base64.b64encode(buf).decode() # ══════════════════════════════════════════════════════════════════════════════ class Handler(BaseHTTPRequestHandler): def log_message(self, *_): pass def _json(self, data): body = json.dumps(data, ensure_ascii=False).encode() self.send_response(200) self.send_header("Content-Type", "application/json") self.send_header("Access-Control-Allow-Origin", "*") self.end_headers() self.wfile.write(body) def _html(self, body): enc = body.encode() self.send_response(200) self.send_header("Content-Type", "text/html; charset=utf-8") self.end_headers() self.wfile.write(enc) def do_GET(self): ds = _DS url = urlparse(self.path) qs = parse_qs(url.query) if url.path == "/": self._html(build_html(ds)) elif url.path == "/api/meta": self._json({ "name": ds.root.name, "fps": ds.fps, "n_episodes": ds.n_episodes, "task": ds.task_desc, "tactile_2d_shapes": ds.tactile_2d_shapes, "cmap_colors": ds.cmap_colors, "sensor_groups": SENSOR_GROUPS, "episodes": [{"start": int(ep["dataset_from_index"]), "end": int(ep["dataset_to_index"]) - 1, "length":int(ep["length"]), "task": (ep.get("tasks") or ["—"])[0]} for ep in ds.episodes], }) elif url.path == "/api/frame": n = int(qs.get("n", [0])[0]) ep = int(qs.get("ep", [0])[0]) d = ds.get_frame(n) pos = d.get("observation_motion_positions", []) quat = d.get("observation_motion_quaternions", []) tactile = {} baseline = ds.tactile_baselines.get(ep, {}) for key in ds.tactile_2d_shapes: raw = d.get(f"tactile_{key}") if raw is not None: base = baseline.get(key, []) if base: vals = [max(0.0, float(v) - b) for v, b in zip(raw, base)] else: vals = [float(v) for v in raw] tactile[key] = _filter_tactile(vals, ds.tactile_2d_shapes[key]) self._json({ "timestamp": float(d.get("timestamp", 0)), "positions": [float(v) for v in pos] if pos else [], "quaternions": [float(v) for v in quat] if quat else [], "tactile": tactile, }) elif url.path == "/api/tactile_timeline": ep = int(qs.get("ep", [0])[0]) self._json(ds.get_tactile_timeline(ep)) elif url.path == "/api/image": ep = int(qs.get("ep", [0])[0]) f = int(qs.get("f", [0])[0]) typ = qs.get("type", ["color"])[0] self._json({"data": ds.get_video_frame_b64(ep, f, typ)}) else: self.send_response(404); self.end_headers() # ══════════════════════════════════════════════════════════════════════════════ def build_html(ds: DataServer) -> str: return r""" LeRobot Dataset Visualizer
— — —
Ep 0 Description: —
Color —
Motion Skeleton 20 markers · 5 fingers
Tactile Timeline —
Tactile —
Palm
t = 0.000 s Frame 0 / 0
""" # ══════════════════════════════════════════════════════════════════════════════ def _pick_folder(): try: import tkinter as tk from tkinter import filedialog r=tk.Tk(); r.withdraw() p=filedialog.askdirectory(title="选择 LeRobot 数据集文件夹") r.destroy(); return p or None except Exception: return None def main(): global _DS parser=argparse.ArgumentParser() parser.add_argument("--dataset","-d",default=None) parser.add_argument("--port","-p",type=int,default=PORT) args=parser.parse_args() path=args.dataset or _pick_folder() if not path: sys.exit("未指定数据集路径") _DS=DataServer(path) class _Srv(HTTPServer): allow_reuse_address=True srv=_Srv(("127.0.0.1",args.port),Handler) url=f"http://127.0.0.1:{args.port}" print(f"\n可视化服务已启动 → {url}\n按 Ctrl+C 停止\n") threading.Thread(target=lambda:(time.sleep(0.8),webbrowser.open(url)),daemon=True).start() try: srv.serve_forever() except KeyboardInterrupt: print("\n已停止") if __name__=="__main__": main()