Instructions to use aronvandepol/visual-feature-discovery with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- timm
How to use aronvandepol/visual-feature-discovery with timm:
import timm model = timm.create_model("hf-hub:aronvandepol/visual-feature-discovery", pretrained=True) - Notebooks
- Google Colab
- Kaggle
Visual Feature Discovery for Colonial Korean Print: trained weights
These are the trained models from the article below. They classify a scanned page from a colonial-era Korean periodical by the printshop that printed it, over four classes: ๋๋์ธ์์, ์ ๋ฌธ๊ด, ์กฐ์ ์ธ์ ์ฃผ์ํ์ฌ and ํ์ฑ๋์ ์ฃผ์ํ์ฌ.
van de Pol, A., Prokic, J., & Mol, A. (2026). Interpretable visual feature discovery using multiple instance learning: A case study of colonial Korean print. Computational Humanities Research, 2, e22. https://doi.org/10.1017/chr.2026.10036
This repository is a mirror of the weights. The citable archive is the Zenodo record 10.5281/zenodo.20015055, which also holds the dataset metadata, the fold definitions and the patch-embedding and cluster files. The code is at aronvandepol/visual-feature-discovery.
Files
| Files | Model | Evaluation protocol | Size each |
|---|---|---|---|
attrimil_p16s8_flash_fold{0-4}.pth |
AttriMIL, 16x16 patches, stride 8 | held-out book (GroupKFold) | 12 MB |
attrimil_p8s4_flash_fold{0-4}.pth |
AttriMIL, 8x8 patches, stride 4 | held-out book (GroupKFold) | 12 MB |
convnextv2_base_book_split_fold{0-4}.pth |
ConvNeXtV2 Base, full page at 896x896 | held-out book (GroupKFold) | 351 MB |
convnextv2_base_page_split_fold{0-4}.pth |
ConvNeXtV2 Base, full page at 896x896 | page-random (KFold) | 351 MB |
swin_s3_base_224_book_split_fold{0-4}.pth |
Swin S3 Base 224, text-only page cutouts | held-out book (GroupKFold) | 282 MB |
swin_s3_base_224_page_split_fold{0-4}.pth |
Swin S3 Base 224, text-only page cutouts | page-random (KFold) | 282 MB |
Each .pth file is the best checkpoint of its fold by validation loss, with the optimizer state removed. It holds state_dict, hyper_parameters, epoch and pytorch-lightning_version.
eval/ has the per-fold results behind the table below. ZENODO_ASSETS.md lists the SHA-256 checksum and best epoch of every weight file.
Results
Five-fold mean accuracy with sample standard deviation, computed from the per-fold files in eval/. The means are those of the article's comparison table. The article prints different standard deviations for the two baselines.
| Model | Held-out book | Page-random |
|---|---|---|
| AttriMIL 16x16, stride 8 | 84.52 ยฑ 2.19% | |
| AttriMIL 8x8, stride 4 | 83.60 ยฑ 3.24% | |
| ConvNeXtV2 Base | 81.98 ยฑ 2.40% | 98.05 ยฑ 0.24% |
| Swin S3 Base 224 | 84.11 ยฑ 1.81% | 98.15 ยฑ 0.17% |
The held-out-book column is the one to read. Under page-random splits, pages of one book fall on both sides of the split, and the baselines score 14 to 16 points higher.
Loading
The checkpoints need the model classes from the code repository.
# run from a checkout of aronvandepol/visual-feature-discovery
import torch
from huggingface_hub import hf_hub_download
from src.models.lightning_attrimil import LightningAttriMIL
path = hf_hub_download("aronvandepol/visual-feature-discovery", "attrimil_p8s4_flash_fold0.pth")
ckpt = torch.load(path, map_location="cpu", weights_only=True)
model = LightningAttriMIL(**ckpt["hyper_parameters"])
model.load_state_dict(ckpt["state_dict"])
model.eval()
The baselines are plain timm models. Their keys carry a model. prefix from the Lightning wrapper (LightningTIMM in src/models/lightning_timm.py).
import timm, torch
from huggingface_hub import hf_hub_download
path = hf_hub_download("aronvandepol/visual-feature-discovery", "convnextv2_base_book_split_fold0.pth")
ckpt = torch.load(path, map_location="cpu", weights_only=True)
hp = ckpt["hyper_parameters"]
net = timm.create_model(hp["model_name"], pretrained=False, num_classes=hp["num_classes"])
net.load_state_dict({k.removeprefix("model."): v for k, v in ckpt["state_dict"].items()})
net.eval()
Data
The page images come from the Hyundam Mun'go collection (formerly Adan Mungo) and the corpus is not redistributed here or on Zenodo. Ten sample pages for testing ship with the code. The Zenodo record's dataset_metadata.parquet gives the catalog link of each of the 57,583 pages.
License
The weights are released under CC-BY-4.0. The code is under the MIT License.