File size: 3,146 Bytes
0364d34 a9468ec 0364d34 b54bf45 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 | ---
pipeline_tag: image-classification
library_name: fgir-zoo
tags:
- fine-grained
- image-classification
- vision-transformer
- full-fine-tuning
---
# FGIRFT
A full fine-tuning benchmark for fine-grained image recognition (FGIR). Every checkpoint is an
ImageNet-21k **ViT-B/16 fine-tuned end to end** — the whole backbone is trained, not frozen — with a
fine-grained classification head, across 21 fine-grained datasets. FGIRFT is the full fine-tuning
counterpart to the ERIS Lab's parameter-efficient works (the ILA/SAW and AAA adapters in
`ERISLab/FGIR-ViT`, which keep the backbone frozen and train only small adapters).
242 checkpoints, one seed per configuration, last epoch only, grouped by input resolution: `ft_224`
(111) and `ft_448` (131). Everything is described in `manifest.csv` and loadable by name with the
standalone `fgir_zoo` library (no training repo needed).
## What it holds
ViT-B/16 fully fine-tuned at 224 or 448 px, each with one FGIR classification head: a plain linear
classifier (`cls`), GLSim, MAWS (the FFVT head), PSM (the TransFG head), CAL, MPN-COV, and
attention-rollout heads. The 21 datasets span general fine-grained sets (CUB, NABirds, Aircraft,
Stanford Cars, Stanford Dogs, Oxford Flowers, Oxford Pets, Food-101, VegFru, DAFB, Moe) and the
ultra-fine-grained leaf sets (Cotton, SoyAgeing and its five single-season subsets, SoyGene,
SoyGlobal, SoyLocal). `num_classes` ranges from 37 to 3263. `manifest.csv` maps every file to its
dataset, head, serial, resolution, `num_classes`, top-1 accuracy and sha256.
## Load a checkpoint
Standalone, no `fgir_vit` install — the model code is vendored inside `fgir_zoo`:
```bash
pip install git+https://github.com/arkel23/fgir-zoo.git
```
```python
import fgir_zoo
model = fgir_zoo.load('fgirvit/ft_448/aircraft_vit_b16_avg_cls_rollout_3')
model.eval()
# fgir_zoo.list_models(family='fgirvit', group='ft_448') lists every name; model.config holds the run's settings.
```
Or fetch the file directly and load it yourself:
```python
import torch
from huggingface_hub import hf_hub_download
path = hf_hub_download('ERISLab/FGIRFT', 'ft_448/aircraft_vit_b16_avg_cls_rollout_1003.pth')
ckpt = torch.load(path, map_location='cpu', weights_only=False) # dict: config, model, accuracy, epoch
```
## Requirements
- `torch>=2.5`, `timm==0.9.12` (the checkpoints were trained on it and the vendored model code forks
that version), `huggingface_hub`, `safetensors`, `einops`, `ml_collections`.
## Related
Part of the [ERISLab FGIRFT collection](https://huggingface.co/collections/ERISLab/fgirft-full-fine-tuning-benchmark-for-fgir-6ab2e8940f3a28ec7ba0bfbf).
For the parameter-efficient adapter works (ILA, SAW, AAA) on a frozen ViT-B/16, see
[`ERISLab/FGIR-ViT`](https://huggingface.co/ERISLab/FGIR-ViT). As a full fine-tuning line, FGIRFT sits
close to the [ERISLab FGIR-Backbones benchmark](https://huggingface.co/collections/ERISLab/fgir-backbones-fgvc13-cvpr-2026-6ab234cc19fcfd2de1f796a4)
(repo [`ERISLab/FGIR-Backbones`](https://huggingface.co/ERISLab/FGIR-Backbones)), which studies
backbones and pretraining recipes for fine-grained recognition.
|