WAM-data / README.md
ziyuz's picture
Upload README.md with huggingface_hub
2a2b98b verified
|
Raw
History Blame Contribute Delete
1.34 kB
---
license: other
task_categories:
- image-to-text
- robotics
tags:
- autonomous-driving
- nuScenes
- nuReasoning
- trajectory-prediction
- world-model
---
# WAM-data (GeoWAM Stage-2 extracted data)
`qixuewei/nuReasoning` 抽取的 GeoWAM / MMaDA Stage-2 训练数据的**元数据**
## 仓库内容
| 文件 | 内容 |
|---|---|
| `WAM_data_meta.tar.gz` | `manifest.jsonl` + `ego_state.jsonl` + `done_clips.txt` + `ego_done_clips.txt` |
- **manifest.jsonl**(7785 行):每个决策关键帧一条样本,字段
`clip, cur(6视图图片名), fut, prompt, reasoning, spatial, geom, command`
- **ego_state.jsonl**(7920 行):每帧 ego 运动学状态 + 5s 未来轨迹
`clip, fidx, speed, vx, vy, ax, ay, px, py, yaw, traj_future[[x,y,yaw]...]`
## 图片未包含(55GB,5.5万张)
图片体量太大未上传。`manifest.jsonl` 里记录了每张图片的文件名,可用项目里的
`training/extract_nureasoning.py` 从原始 `qixuewei/nuReasoning` 重新抽取(断点续传):
```bash
python -m training.extract_nureasoning --n_clips 0 --workers 24
```
抽出的图片落在 `images/<clip>_<fidx>_cur_<view>.jpg` 和 `<clip>_<fidx>_fut.jpg`,
与 manifest 中的文件名一一对应。
## 配套代码 / 模型
- 代码 + 环境 + 对话历史 + Stage-1 backbone:`ziyuz/WAM`(model repo)