ckpt_backup / README.md
chengenwu's picture
Add files using upload-large-folder tool
2558113 verified
|
Raw History Blame Contribute Delete
735 Bytes
# ckpt_backup
Backup of `robopolicy` training checkpoints and evaluation rollouts.
## Layout
- `checkpoints/<run>/ckpt_final.pt` — 194 checkpoints.
Optimizer state has been stripped: keys are `model_state_dict`, `processor_stats`,
`run_config`, `scheduler_state_dict`, `step`. Inference/eval only — these cannot
be resumed for further training.
- `rollouts/<run>/<name>_rollouts/videos.tar` — per-run rollout videos (store-only tar).
- `rollouts/<run>/<name>_rollouts/results.json|txt` — eval metrics, uncompressed.
- `checkpoint_aliases.json` — 192 alias paths that were local symlinks
to `ckpt_final.pt` (same step, identical weights). Recreate with
`ln -s ckpt_final.pt <alias>` rather than downloading twice.