Document the bundled in-domain eval scenes
Browse files
README.md
CHANGED
|
@@ -113,6 +113,27 @@ pipeline (5.8 h for 1016 episodes).
|
|
| 113 |
> LeRobotDataset(repo_id, root=root)
|
| 114 |
> ```
|
| 115 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 116 |
### Reproducing
|
| 117 |
|
| 118 |
Generation, conversion, training and evaluation scripts, plus the measured results and the
|
|
|
|
| 113 |
> LeRobotDataset(repo_id, root=root)
|
| 114 |
> ```
|
| 115 |
|
| 116 |
+
### In-domain eval scenes (`eval_scenes/`)
|
| 117 |
+
|
| 118 |
+
Reproducing the in-domain numbers needs the scenes restored exactly as they were generated, which
|
| 119 |
+
`run_ckpt_eval*.sh` does through `--replay_h5_dir` on the source HDF5s. Those are 37 GB (25 Hz) /
|
| 120 |
+
267 GB (250 Hz) and are not published.
|
| 121 |
+
|
| 122 |
+
They do not have to be. `eval_server.load_h5_episode` reads only the root attrs (`env_config`,
|
| 123 |
+
`seed`, `instruction`, `init_*`) and the `action` dataset -- the camera frames, which are 99.99%
|
| 124 |
+
of the file, are re-rendered by the server. Stripped to that, a 250 Hz 20-scene set goes from
|
| 125 |
+
5.1 GB to 0.37 MB, and ships here under `eval_scenes/`. Restoring from the stripped files was
|
| 126 |
+
verified to reproduce the original scene exactly: object pose and velocity, bowl pose, arm pose
|
| 127 |
+
and action length all match.
|
| 128 |
+
|
| 129 |
+
```bash
|
| 130 |
+
huggingface-cli download mickeykang/dvla-can-250hz-events-250fps-delta-trim --repo-type dataset --include 'eval_scenes/*' --local-dir .
|
| 131 |
+
REPLAY_DIR=$PWD/eval_scenes/indomain_scenes_s20_250 ... bash overnight_delta_n3/run_policy_eval.sh
|
| 132 |
+
```
|
| 133 |
+
|
| 134 |
+
Out-of-distribution eval needs no files at all -- those scenes are generated from seeds
|
| 135 |
+
9000000-9000019 at run time, and both control rates draw the same ones.
|
| 136 |
+
|
| 137 |
### Reproducing
|
| 138 |
|
| 139 |
Generation, conversion, training and evaluation scripts, plus the measured results and the
|