Awesome

ProtoPFormer

Introduction

This is the official implementation of paper "ProtoPFormer: Concentrating on Prototypical Parts in Vision Transformers for Interpretable Image Recognition". [Paper]

The overview of ProtoPFormer is as below:

Updates

• 09/27/2022 The supplementary material has been added in the Arxiv preprint.
• 09/18/2022 Modify the default backbone to be deit_tiny_patch16_224.
• 08/29/2022 The citation of our paper has been added.
• 08/23/2022 Code has been uploaded.

Dataset

The CUB dataset is downloaded from https://www.vision.caltech.edu/datasets/cub_200_2011/.

The Stanford Dogs dataset is downloaded from http://vision.stanford.edu/aditya86/ImageNetDogs/.

The Stanford Cars dataset is downloaded from http://ai.stanford.edu/~jkrause/cars/car_dataset.html.

Decompress them and place them in the datasets/ folder as follows:

datasets
--CUB_200_2011
    |--attributes
    |--images
    |--parts
    |...
--stanford_dogs
    |--Annotation
    |--Images
    |...
--stanford_cars
    |--cars_test
    |--cars_train
    |...

Usage

First clone the repository locally:

git clone git@github.com:zju-vipa/ProtoPFormer.git

Then install the requried packages as below:

pytorch==1.8.1
torchvision==0.9.1
pillow==9.1.1
timm==0.5.4
opencv-python==4.6.0.66
tensorboard==2.9.1
scipy==1.8.1
pandas==1.4.3
matplotlib==3.5.2
scikit-learn==1.1.1

Newer versions of these packages may work.

Train

The training scripts for CUB, Dogs and Cars are scripts/train_cub.sh, scripts/train_dogs.sh, scripts/train_car.sh, respectively.

The format instruction of training is as below:

sh scripts/train_cub.sh $backbone $batch_size $num_gpus

Here, $backbone is chosen from [deit_tiny_patch16_224, deit_small_patch16_224, cait_xxs24_224], $batch_size denotes batch size, and $num_gpus denotes the number of gpus for training. An example for training a ProtoPFormer of backbone deit_tiny_patch16_224 on CUB with batch size 64 and 2 gpus is as below:

sh scripts/train_cub.sh deit_tiny_patch16_224 64 2

Visualize

The format instruction of training is as below:

sh scripts/visualize.sh $backbone $data_set $output_dir $use_gauss $modeldir $modelfile

An example for visualizing a ProtoPFormer of backbone deit_tiny_patch16_224 on CUB with a checkpoint file output_cosine/CUB2011U/deit_tiny_patch16_224/1028--adamw-0.05-200-protopformer/checkpoints/epoch-best.pth is as below:

sh scripts/visualize.sh deit_tiny_patch16_224 \
CUB2011U \
output_view \
False \
output_cosine/CUB2011U/deit_tiny_patch16_224/1028--adamw-0.05-200-protopformer/checkpoints \
epoch-best.pth

Noted that $use_gauss denotes whether visualizing gaussian shape, it will be much slower if $use_gauss is True.

Evaluating Interpretability

We borrow the evaluation code from EvalProtoPNet to evaluate the interpretability score of ProtoPFormer.

python eval_interpretability.py \
--gpuid=0 \
--base_architecture=deit_tiny_patch16_224 \
--reserve_layers=11 \
--data_path=datasets/CUB_200_2011/ \
--prototype_shape 2000 192 1 1 \
--use_global=True \
--global_proto_per_class=10 \
--resume=output_cosine/CUB2011U/deit_tiny_patch16_224/1028--adamw-0.05-200-protopformer/checkpoints/epoch-best.pth \
--global_coe=0.5 \
--reserve_token_nums=81 \
--use_ppc_loss=True \
--batch_size=32 \
--out_dir=output_view

The below table shows that ProtoPFormer significantly improves the interpretability score of ProtoPNet on vision transformers.

Citation

If you find our code is helpful for your research, please cite:

@article{xue2022protopformer,
  title={ProtoPFormer: Concentrating on Prototypical Parts in Vision Transformers for Interpretable Image Recognition},
  author={Xue, Mengqi and Huang, Qihan and Zhang, Haofei and Cheng, Lechao and Song, Jie and Wu, Minghui and Song, Mingli},
  journal={arXiv preprint arXiv:2208.10431},
  year={2022}
}

Statement

This project is for research purpose only. For any other questions please contact mqxue@zju.edu.cn.