cjy666/lmc-ckpt

ckpt/ — pretrained weights

All weights needed by scripts/infer_one.py. Total ~11 GB.

Contents

ckpt/
├── dinov3/
│   └── dinov3_vitl16_pretrain_lvd1689m-8aa4cbdd.pth         1.2 GB   ViT-L/16 LVD-1689M backbone
│
├── vessel_seg_nnunet/                                       Sec 4.2: DIAS-trained vessel mask
│   ├── plans.json
│   ├── dataset.json
│   └── fold_{0..4}/checkpoint_best.pth            5×354 MB  clDice loss, 2D nnU-Net
│
├── pixel_branch_nnunet/                                     Sec 4.4: collateral pixel branch
│   ├── plans.json
│   ├── dataset.json
│   └── fold_{0..4}/checkpoint_best_prauc.pth      5×350 MB  PR-AUC selection, masked input
│
└── graph_branch_gat/                                        Sec 4.3: DINOv3 + GAT
    ├── train_config.json
    └── fold{0..4}_best_prauc.pt                   5×1.2 GB  exp_00167 best-PRAUC selection

Provenance

Trained checkpoints are copied verbatim from:

File group	Source path
dinov3/	DINOv3 official LVD-1689M ViT-L/16 release
vessel_seg_nnunet/fold_*/checkpoint_best.pth	Dataset501_DIASVessel, trainer nnUNetTrainerClDice2D
pixel_branch_nnunet/fold_*/checkpoint_best_prauc.pth	Dataset708_ZurichCollateralCVMasked_EVEN, trainer nnUNetTrainerPRAUC2D
graph_branch_gat/fold*_best_prauc.pt	exp_00167_2045694, batch BATCH1

DINOv3 weights filename

The DINOv3 hub loader requires the weights file to be named with an 8-character hash suffix (e.g. *-8aa4cbdd.pth; 8aa4cbdd is the canonical LVD-1689M hash). We therefore ship the single real weights file already named dinov3_vitl16_pretrain_lvd1689m-8aa4cbdd.pth, and the inference config points directly at it — no symlink or alias is created at load time.

Public-release notes

• Total size (~11 GB) exceeds typical git limits. For public release, host these on a dedicated artefact store (HuggingFace Hub / Zenodo / GitHub Release Assets) and use scripts/download_ckpts.sh (currently a stub — fill in the artefact URL and SHA-256) to pull the same layout.