Solar filament instance segmentation: weights (Filament Segmentation Challenge 2026)

Weights of our two final submissions to the Solar Filament Segmentation Challenge 2026. The code, configs and full documentation are in the GitHub repository xiehuanyi/filament-segmentation-2026.

Input: 2048 x 2048 GONG H-alpha frames. Output: one pixel-disjoint mask per filament.

pick pipeline 5-fold OOF PQ
s15 6-model U-Net teacher ensemble -> LightGBM scorer -> crop refiner -> CNN rescorers 0.4678
s16 3 distilled student U-Nets -> LightGBM scorer -> crop refiner -> CNN rescorers 0.4669

Files

dense/<run>/model_f{0..4}.pt, args_f{0..4}.json   9 runs x 5 folds (EMA state dicts + training arguments)
scorer/{ensB,eS}_lgb{0,1}.txt, {ensB,eS}.json      LightGBM boosters
refiner/{ensB,eS}_refiner.pt                       full-data refiners (+ _f{0..4}.pt fold refiners, OOF check only)
rescorer/{ensB,eS}_<name>_full.pt                  13 full-data rescorers

The inference bundle is 7.33 GB (s15 alone 6.80 GB, s16 alone 2.54 GB); the fold refiners add 0.98 GB. final/weights_manifest.csv in the GitHub repository lists every file with its size and sha256.

Usage

git clone https://github.com/xiehuanyi/filament-segmentation-2026 && cd filament-segmentation-2026
pip install -r requirements.txt
python scripts/get_weights.py --out WEIGHTS --hf_repo NightNightUp/filament-segmentation-2026               # download + sha256 check
python -m filseg.pipeline --config final/s15.json --split test --data DATA --work WORK --weights WEIGHTS \
    --out submission_s15.csv --gpus 0

Add --bundle inference --pick s16 to get_weights.py to fetch only what one pick needs.

Training data

The models were trained on the MAGFiLO annotations of the competition training frames. The students of s16 were also trained on 1314 unlabelled GONG H-alpha frames from the public NSO archive: other days than any train or test frame, and at least 6 h from every test timestamp. Those frames were used for training only, never at inference.

License

MIT. The pretrained encoder weights we started from (timm / segmentation_models_pytorch), the MAGFiLO annotations and the GONG frames keep their own terms.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support