example_usage.py

# -*- coding: utf-8 -*-
"""
Fixed_Viewpoint_Tactile_Dataset 使用示例（可直接运行）

环境要求: Python >= 3.10, 已安装 lerobot
    conda activate zyhand        # 你机器上有 lerobot 0.4.3 的环境
    python example_usage.py                      # 从 HF 自动下载
    python example_usage.py --root E:/tachintech/dataset/lerobot_tactile_hand_20fps  # 用本地, 不下载
"""
import argparse
import numpy as np

REPO = "Tachintech/Fixed_Viewpoint_Tactile_Dataset"


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--root", default=None, help="本地数据集路径(给了就不从HF下载)")
    ap.add_argument("--save-img", default="sample_color.png", help="保存一帧RGB的路径")
    args = ap.parse_args()

    # ── 1) 加载数据集 ────────────────────────────────────────────────
    from lerobot.datasets.lerobot_dataset import LeRobotDataset
    print("加载数据集中 ...")
    ds = LeRobotDataset(REPO, root=args.root)
    print(f"  总帧数   : {ds.num_frames}")
    print(f"  episodes : {ds.num_episodes}")
    print(f"  fps      : {ds.fps}")

    # ── 2) 列出所有键(字段) ──────────────────────────────────────────
    print("\n===== 数据集所有键 (来自 meta/info.json) =====")
    feats = ds.meta.info["features"]
    for name, spec in feats.items():
        print(f"  {name:35s} dtype={spec['dtype']:8s} shape={spec['shape']}")

    # 触觉每路的 2D 形状
    shapes = ds.meta.info.get("tactile_2d_shapes", {})
    print("\n触觉传感器 2D 形状 (tactile_2d_shapes):")
    for k, v in shapes.items():
        print(f"  {k:12s} -> {v[0]} x {v[1]}")

    # ── 3) 取一帧, 看实际内容 ────────────────────────────────────────
    print("\n===== 取第 0 帧, 各键的形状 =====")
    s = ds[0]
    for k in sorted(s.keys()):
        v = s[k]
        shp = tuple(v.shape) if hasattr(v, "shape") else f"{type(v).__name__}={v}"
        print(f"  {k:35s} {shp}")

    # ── 4) 动捕: 还原成 20 个点 ──────────────────────────────────────
    pos = s["observation_motion_positions"].numpy()        # [60]
    quat = s["observation_motion_quaternions"].numpy()     # [80]
    pos20 = pos.reshape(20, 3)      # 20 个点 x (x,y,z)
    quat20 = quat.reshape(20, 4)    # 20 个点 x (w,x,y,z)
    print("\n===== 动捕 =====")
    print(f"  20 个标记点位置 reshape -> {pos20.shape}, 第0点 xyz = {np.round(pos20[0],4)}")
    print(f"  20 个标记点姿态 reshape -> {quat20.shape}, 第0点 quat= {np.round(quat20[0],4)}")

    # ── 5) 触觉: 还原成 2D 阵列 ──────────────────────────────────────
    print("\n===== 触觉 (还原成 2D) =====")
    for i in [0, 19]:
        key = f"tactile_{i}"
        flat = s[f"tactile_tactile_{i}"].numpy()
        grid = flat.reshape(shapes[key])
        print(f"  tactile_tactile_{i}: {flat.shape} -> 2D {grid.shape}, 和={grid.sum():.3f}")

    # ── 6) action 是什么 ─────────────────────────────────────────────
    act = s["action"].numpy()                              # [140]
    obs_concat = np.concatenate([pos, quat])              # 60+80=140
    print("\n===== action =====")
    print(f"  action shape = {act.shape}")
    print(f"  action 是否 == (positions ⊕ quaternions): "
          f"{'是(完全相等)' if np.allclose(act, obs_concat) else '否'}")
    print("  即 action = 当前帧 20 个动捕点的 (位置60 + 四元数80)")

    # ── 7) 视频帧 (已自动从 mp4 解码) ───────────────────────────────
    color = s["observation.images.color"]                  # [3,480,848] float[0,1]
    depth = s["observation.images.depth"]
    print("\n===== 视频 =====")
    print(f"  color: {tuple(color.shape)} dtype={color.dtype} 值域[{color.min():.2f},{color.max():.2f}]")
    print(f"  depth: {tuple(depth.shape)}")
    try:
        from PIL import Image
        img = (color.permute(1, 2, 0).numpy() * 255).astype(np.uint8)
        Image.fromarray(img).save(args.save_img)
        print(f"  已保存一帧 RGB -> {args.save_img}")
    except Exception as e:
        print(f"  (保存图片跳过: {e})")

    # ── 8) 按 episode 取 + DataLoader 批训练 ────────────────────────
    print("\n===== 遍历 / 训练 =====")
    # episode 边界
    try:
        froms = ds.meta.episodes["dataset_from_index"]
        tos = ds.meta.episodes["dataset_to_index"]
        print(f"  episode 0: 帧 [{froms[0]}, {tos[0]}) 共 {tos[0]-froms[0]} 帧")
    except Exception:
        print("  (episode 边界 API 视版本而定, 可从 meta/episodes parquet 读 dataset_from/to_index)")

    from torch.utils.data import DataLoader
    dl = DataLoader(ds, batch_size=8, shuffle=True, num_workers=0)
    batch = next(iter(dl))
    print(f"  一个 batch: action={tuple(batch['action'].shape)} "
          f"color={tuple(batch['observation.images.color'].shape)} "
          f"tactile_0={tuple(batch['tactile_tactile_0'].shape)}")

    print("\n===== 完成: 数据集可正常加载和使用 =====")


if __name__ == "__main__":
    main()