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:

pip install git+https://github.com/arkel23/fgir-zoo.git

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:

import torch
from huggingface_hub import hf_hub_download
path = hf_hub_download('ERISLab/FGIRFT', 'ft_448/aircraft_vit_b16_avg_cls_rollout_3.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. For the parameter-efficient adapter works (ILA, SAW, AAA) on a frozen ViT-B/16, see ERISLab/FGIR-ViT. As a full fine-tuning line, FGIRFT sits close to the ERISLab FGIR-Backbones benchmark (repo ERISLab/FGIR-Backbones), which studies backbones and pretraining recipes for fine-grained recognition.