TransFG: A Transformer Architecture for Fine-grained Recognition

This is the official PyTorch implementation of the paper "TransFG: A Transformer Architecture for Fine-grained Recognition" (Ju He, Jie-Neng Chen, Shuai Liu, Adam Kortylewski, Cheng Yang, Yutong Bai, Changhu Wang, Alan Yuille).

Official PyTorch code for the paper: TransFG: A Transformer Architecture for Fine-grained Recognition

Implementation based on DeiT pretrained on ImageNet-1K with distillation fine-tuning will be released soon.


  • Python 3.7.3
  • PyTorch 1.5.1
  • torchvision 0.6.1
  • ml_collections


1. Download Google pre-trained ViT models


2. Prepare data

In the paper, we use data from 5 publicly available datasets:

Please download them from the official websites and put them in the corresponding folders.

3. Install required packages

Install dependencies with the following command:

pip3 install -r requirements.txt

4. Train

To train TransFG on CUB-200-2011 dataset with 4 gpus in FP-16 mode for 10000 steps run:

CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m torch.distributed.launch --nproc_per_node=4 --dataset CUB_200_2011 --split overlap --num_steps 10000 --fp16 --name sample_run


If you find our work helpful in your research, please cite it as:

  title={TransFG: A Transformer Architecture for Fine-grained Recognition},
  author={He, Ju and Chen, Jieneng and Liu, Shuai and Kortylewski, Adam and Yang, Cheng and Bai, Yutong and Wang, Changhu and Yuille, Alan},
  journal={arXiv preprint arXiv:2103.07976},


Many thanks to ViT-pytorch for the PyTorch reimplementation of An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale