CARP-Net stack: solar filament instance segmentation
The 30 trained fold models behind our selected submission (public PQ 0.40) to the Kaggle challenge
Solar Filament Instance Segmentation 2026. Code and demo notebook: https://github.com/cazaroni/filament-carp
(notebooks/demo.ipynb downloads this repo automatically with huggingface_hub.snapshot_download).
| Path | Contents |
|---|---|
<family>/fold<k>/model.pth |
PyTorch state_dict of one fold model (fold k held out in training) |
<family>/fold<k>/config.json |
training arguments and that fold's in-training validation scores |
stack.json |
members, blend weights, test-time augmentation and decoder settings |
SHA256SUMS |
checksums of the 30 .pth files |
| Family | Architecture | Blend weight |
|---|---|---|
carp |
CARP-Net: ConvNeXt-T HiRes U-Net + stride-4 affinity-steered propagation | 1/4 |
carpw |
CARP-Net, images sampled by annotator-record count | 1/4 |
mitb2 |
MiT-B2 U-Net | 1/8 |
rmi |
MiT-B2 U-Net + RMI loss | 1/8 |
cnx |
ConvNeXt-T U-Net | 1/8 |
cnxhr |
ConvNeXt-T HiRes U-Net | 1/8 |
Load one model with filseg.stack.load_member(root, "carp", 0) from the code repo. Input: one 2048 ร 2048
grayscale image normalised over the solar disk (filseg.model.normalise). Output: one logit map.