Single-Step-Grasp-Refinement / metadata /release_decisions.md
yiju-1213's picture
Publish single-step grasp refinement evaluation checkpoint
4e5de6c verified
|
Raw
History Blame Contribute Delete
1.37 kB
# Release Decisions
## Default Checkpoint
The default public checkpoint is `full_sga_gsn_seed8_best.pt`.
Reasons:
- It is the only uncommented Full SGA-GSN entry in the current paper-style manifest.
- Its best validation metric is `0.109375`, higher than the seed-7 Full checkpoint value observed during audit.
- Its formal unseen-test lift is effectively aligned with seed 7 while showing smaller tactile degradation in the current summary table.
## Checkpoint Scope
The release checkpoint is evaluation-only:
- Kept: `actor_critic`, `calibrator`, `experiment_cfg`, `object_split`, best-metric metadata.
- Omitted: `optimizer`, full `history`.
If exact training resume is required later, publish a separate `training_resume/` artifact and document it as a larger, non-default file.
## Ablations
The minimal model repo does not include ablation weights. Learned ablation weights can be added later under `ablations/` if the paper release promises full ablation reproduction.
`no-action` and `rand-action` are policy modes and do not require checkpoint files.
## Pending Release Gates
- Freeze the final public RL code repository and Dockerfile.
- Freeze the public SGA-GSN perception checkpoint path and checksums.
- Freeze the 3DA-VTG dataset repo and VT-Grasp simulation asset repo.
- Decide whether paper-exact seed 7 should be included alongside seed 8.