| --- |
| license: cc-by-nc-sa-4.0 |
| library_name: pytorch |
| pipeline_tag: image-feature-extraction |
| datasets: |
| - fdsml/FineFractalDB |
| tags: |
| - fdsml |
| - formula-driven-supervised-learning |
| - metric-learning |
| - fractal |
| - synthetic-data |
| - pre-training |
| - vit |
| - arxiv:2610.39423 |
| --- |
| |
| # FDSML ViT-Base (DeiT-Base/16, 224) — `diff` strategy |
|
|
| ViT-Base (DeiT-Base/16, 224) pre-trained on [FineFractalDB](https://huggingface.co/datasets/fdsml/FineFractalDB) with **FDSML** (Formula-Driven Supervised Metric Learning) using the Parameter-Aware Triplet Loss and the **`diff`** triplet selection strategy. No natural images are used during pre-training. The main ViT-Base results in the paper use the [`each`](https://huggingface.co/fdsml/fdsml-finefractaldb-vitbase-each) strategy. |
|
|
| - **Paper:** https://arxiv.org/abs/2610.39423 |
| - **Project page:** https://shineiarakawa.github.io/fdsml/ |
| - **Code:** https://github.com/ShineiArakawa/fdsml |
| - **Dataset:** https://huggingface.co/datasets/fdsml/FineFractalDB |
|
|
| ## Model details |
|
|
| | Item | Value | |
| |---|---| |
| | Architecture | ViT-Base (DeiT-Base/16, 224) (86.6M parameters) | |
| | Pre-training data | FineFractalDB (1,000 IFS classes × 1,000 images, synthetic) | |
| | Objective | Parameter-Aware Triplet Loss | |
| | Triplet strategy | `diff` (diff_class): selects all pairs from different classes | |
| | Projection dim | 256 | |
| | Training images seen | 524,288k (524.29 epochs) | |
| | Checkpoint | `checkpoint-524288kimg-524.29ep_fdsml_deitbase_diff.pt` | |
| | Name in the official code | `deit_base_patch16_224_mocov3_fdsml_diff_pretrained` | |
|
|
| ## Usage |
|
|
| The checkpoint is a plain PyTorch file and is loaded with the [official code](https://github.com/ShineiArakawa/fdsml). |
|
|
| ```python |
| import torch |
| from huggingface_hub import hf_hub_download |
| |
| path = hf_hub_download("fdsml/fdsml-finefractaldb-vitbase-diff", "checkpoint-524288kimg-524.29ep_fdsml_deitbase_diff.pt") |
| ckpt = torch.load(path, map_location="cpu") |
| ``` |
|
|
| To fine-tune on a downstream task, pass the checkpoint to `finetune.py`: |
|
|
| ```bash |
| torchrun --standalone --nproc_per_node=4 finetune.py \ |
| --net=deit_base_patch16_224_mocov3 \ |
| --ckpt=/path/to/checkpoint-524288kimg-524.29ep_fdsml_deitbase_diff.pt \ |
| --dset=/path/to/imagenet |
| ``` |
|
|
| See the [README](https://github.com/ShineiArakawa/fdsml#fine-tuning) for the full list of options and the exact configurations used in the paper. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @inproceedings{Otake2027fdsml, |
| title = {FDSML: Formula-Driven Supervised Metric Learning with Parameter-Aware Triplet Loss}, |
| author = {Hina Otake and Shinei Arakawa and Keitaro Tanaka and Yoshihiro Fukuhara and Hirokatsu Kataoka and Shigeo Morishima}, |
| booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)}, |
| year = {2027}, |
| } |
| ``` |
|
|
| ## License |
|
|
| Released under [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/). Non-commercial use only. |
|
|
