1.1 KiB
1.1 KiB
2026-04-06 LEWM ViT Transfer Notes
Root-cause fix
The first LEWM runs were stopped because the data path still resized each camera view to 224x224 before multiview fusion. That preserved the final tensor shape but broke the original LEWM geometry.
Corrected path now is:
- Training dataset: keep stored per-view
256x256images (data.image_resize_shape=nullat launch; dataset instantiate override isNonefor LEWM) - Eval rollout input: resize live MuJoCo
480x640camera images to256x256per view - Backbone: fuse
front, top, r_vison the LEWM axis, then resize fused short side to224
Verification
- Local tests passed (
38 passedacross the focused suite) - Remote check:
- dataset sample image shape:
(2, 3, 256, 256) - eval-prepared live frame shape:
(3, 256, 256)
- dataset sample image shape:
- Remote smoke passed with real checkpoint:
smoke-lewm-imf-rawpath-emb384-20260406-002002
Current runs
lewm-vit-imf-raw256fix-sim-transfer-emb384-l12-ph16-ex08-step50k-roll10-5880g0-20260406-002124lewm-vit-imf-raw256fix-sim-transfer-emb256-l12-ph16-ex08-step50k-roll10-5880g1-20260406-002124