FDSML/fdsml-finefractaldb-vittiny-all

FDSML ViT-Tiny (DeiT-Tiny/16, 224) — all strategy

ViT-Tiny (DeiT-Tiny/16, 224) pre-trained on FineFractalDB with FDSML (Formula-Driven Supervised Metric Learning) using the Parameter-Aware Triplet Loss and the all triplet selection strategy. No natural images are used during pre-training. The main ViT-Tiny results in the paper use the each strategy.

• Paper: arXiv link coming soon
• 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-Tiny (DeiT-Tiny/16, 224) (5.7M parameters)
Pre-training data	FineFractalDB (1,000 IFS classes × 1,000 images, synthetic)
Objective	Parameter-Aware Triplet Loss
Triplet strategy	all (all): uses all possible triplets within a batch
Projection dim	128
Training images seen	536,870k (536.87 epochs)
Checkpoint	checkpoint-536870kimg-536.87ep_fdsml_deittiny_all.pt
Name in the official code	deit_tiny_patch16_224_mocov3_fdsml_all_pretrained

Usage

The checkpoint is a plain PyTorch file and is loaded with the official code.

import torch
from huggingface_hub import hf_hub_download

path = hf_hub_download("fdsml/fdsml-finefractaldb-vittiny-all", "checkpoint-536870kimg-536.87ep_fdsml_deittiny_all.pt")
ckpt = torch.load(path, map_location="cpu")

To fine-tune on a downstream task, pass the checkpoint to finetune.py:

torchrun --standalone --nproc_per_node=4 finetune.py \
    --net=deit_tiny_patch16_224_mocov3 \
    --ckpt=/path/to/checkpoint-536870kimg-536.87ep_fdsml_deittiny_all.pt \
    --dset=/path/to/imagenet

See the README for the full list of options and the exact configurations used in the paper.

Citation

@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. Non-commercial use only.