MaybeRichard/rae-fm-generation-pipeline

RAE FM Generation Pipeline Snapshot

Public lightweight snapshot for continuing the FM + generation-model experiments on ISIC and Kermany2018.

Included

• code/RAEv2/: source, configs, and launch/evaluation scripts used for ISIC and Kermany SD-VAE arms.
• code/OCT_RAE_main/: OCT foundation-model RAE-main source/configs/scripts for Kermany RETFound, VisionFM, DINOv2-L, and MAE-L arms.
• code/kermany_pipeline/: current Kermany preparation, sampling, downstream, aggregation, and monitor scripts.
• code/local_repo/: report repo files, helper scripts, and the latest HTML report.
• configs/: high-signal config subsets for ISIC and Kermany.
• results/: compact JSON/CSV/eval outputs, including fixed-decoder Kermany RETFound eval and completed downstream JSONs.
• manifests/: upload and resume status manifests.

Excluded

Raw medical images, full generated image folders, model caches, and most large checkpoints are excluded from this lightweight repo. Large checkpoints live in MaybeRichard/rae-fm-generation-artifacts.

Current Kermany Status

Kermany2018 uses the same experimental route as ISIC. Core arms are SD-VAE, RETFound, VisionFM, DINOv2-L, and MAE-L. The earlier FM sampling bug was fixed by using decoder-only decoder_ema.pt weights instead of full stage1 checkpoints. RETFound fixed-decoder quality eval is included. VisionFM/DINOv2-L/MAE-L have corrected sampling CSVs and uploaded weights; their quality/downstream can be resumed on a new machine after preparing the Kermany dataset.

Resume Notes

1. Download Kermany2018 with kagglehub.dataset_download("paultimothymooney/kermany2018").
2. Recreate labels/ImageFolder views with code/kermany_pipeline/ scripts.
3. Download large weights from MaybeRichard/rae-fm-generation-artifacts.
4. Replace h800 absolute paths in configs with the new machine paths, especially data roots and checkpoint paths.