Source code of the paper Multi-granularity Part Sampling Attention for Fine-grained Visual Classification

conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
conda install numpy matplotlib pandas tensorboard scipy tqdm
pip install yacs opencv-python opencv-contrib-python timm einops gpustat

You may specific your dataset path in the /config/{settting}.yaml under data/datasets. Then please check if there is any conflict with the ip in line 103 of settings/setup_functions.py.

1. Put the pre-trained model (22k 1k for Stanford Dog, and 22k for others) in /pretrained/, and rename it to Swin Base 1k.pth for Dog and Swin Base.pth for others.
2. Change the log name and cuda visible by modifing line 13,14 in /setup.py.
3. Running the following code according to you pytorch version:

python -m main.py

python -m torch.distributed.launch --nproc_per_node 2 main.py

torchrun --nproc_per_node 2 main.py

You need to change the number behind the -nproc_per_node to your number of GPUs.

if this code is helpful to you, please cite as the following format

@ARTICLE{10638479,
  author={Wang, Jiahui and Xu, Qin and Jiang, Bo and Luo, Bin and Tang, Jinhui},
  journal={IEEE Transactions on Image Processing}, 
  title={Multi-Granularity Part Sampling Attention for Fine-Grained Visual Classification}, 
  year={2024},
  volume={33},
  number={},
  pages={4529-4542},
  keywords={Feature extraction;Semantics;Visualization;Shape;Location awareness;Attention mechanisms;Transformers;Fine-grained visual classification;multi-granularity;part sampling;attention mechanism},
  doi={10.1109/TIP.2024.3441813}
}