| # FrameSkip-adapted (+500) |
|
|
| This directory reproduces the Florence-Bridge FrameSkip comparison used by this |
| release. It starts from the same complete 20K state as the OPTR arms and runs 500 |
| additional optimizer updates with the same two-GPU global-batch-64 contract. |
|
|
| The scorer uses the official FrameSkip implementation vendored under |
| `third_party/frameskip/` at upstream commit |
| `965becf63e9cd5cc339fbf7fde40d1db47efa67a`. Bridge does not have a released GMM |
| TPI cache, so this comparison is named **FrameSkip-adapted (Gaussian-TPI |
| fallback)**. The fixed settings are retention ratio 0.2, AVI/VAC/TPI weights |
| 0.6/0.2/0.2, VAC enabled, first/last and gripper-transition preservation, |
| top-10% action-change preservation, and five pruned microbatches followed by one |
| full microbatch. No p/q correction is applied because FrameSkip remaps physical |
| timesteps rather than sampling from OPTR's q(t|e). |
|
|
| Prepare DINOv2-S/14 separately, build the required-draw list and cache with |
| `build_required_draws.py`, `build_cache_shard.py`, and `finalize_cache.py`, then: |
|
|
| ```bash |
| scripts/run_frameskip_plus500.sh frameskip-plus500 /path/to/frameskip-cache.json |
| ``` |
|
|
| The cache and visual weights are runtime assets and are intentionally not stored |
| in Git. Evaluation uses the same `scripts/evaluate_bridge_384_no_video.sh` entry |
| as Uniform, Matched-Random, and Residual-OPTR. |
|
|
