ProtoViT: Interpretable Image Classification with Adaptive Prototype-based Vision Transformers

This repository contains the official implementation of the paper "Interpretable Image Classification with Adaptive Prototype-based Vision Transformers". (NeurIPS 2024)

Table of Contents

• Overview
• Prerequisites
• Installation
• Dataset Preparation
• Training
• Analysis
• Model Zoo
• Citation
• Acknowledgments

Overview

ProtoViT is a novel approach that combines Vision Transformers with prototype-based learning to create interpretable image classification models. Our implementation provides both high accuracy and explainability through learned prototypes.

Prerequisites

Software Requirements

These packages should be enough to reproduce our results. We add requirement.txt based on our conda environment for reference just in case.

• Python 3.8+
• PyTorch with cuda
• NumPy
• OpenCV (cv2)
• Augmentor
• Timm==0.4.12 (Note: Higher versions may require modifications to the ViT encoder)

Hardware Requirements

Recommended GPU configurations:

• 1× NVIDIA Quadro RTX 6000 (24GB) or
• 1× NVIDIA GeForce RTX 4090 (24GB) or
• 1× NVIDIA RTX A6000 (48GB)

Installation

git clone https://github.com/Henrymachiyu/ProtoViT.git
cd ProtoViT
pip install -r requirements.txt

Dataset Preparation

CUB-200-2011 Dataset

1. Download CUB_200_2011.tgz

2. Extract the dataset:

   #Download the dataset CUB_200_2011.tgz from http://www.vision.caltech.edu/visipedia/CUB-200-2011.html
   tar -xzf CUB_200_2011.tgz

3. Process the dataset:

   For cropping data and training_test split, please carefully follow the instructions from the dataset. Sample code can be found in preprocess sample code that can crop and split data with Jupyter Notebook.

   # Create directory structure
   mkdir -p ./datasets/cub200_cropped/{train_cropped,test_cropped}
   
   # Crop and split images using provided scripts
   python your_own_scripts/crop_images.py  # Uses bounding_boxes.txt
   python your_own_scripts/split_dataset.py  # Uses train_test_split.txt
   #Put the cropped training images in the directory "./datasets/cub200_cropped/train_cropped/"
   #Put the cropped test images in the directory "./datasets/cub200_cropped/test_cropped/"
   
   # Augment training data 
   python img_aug.py
   #this will create an augmented training set in the following directory:
   #"./datasets/cub200_cropped/train_cropped_augmented/"

Stanford Cars Dataset

The official website for the dataset is:

• Official Stanford Cars Dataset

Alternative dataset option available from:

• Kaggle Mirror

Training

1. Configure settings in settings.py:

# Dataset paths
data_path = "./datasets/cub200_cropped/"
train_dir = data_path + "train_cropped_augmented/"
test_dir = data_path + "test_cropped/"
train_push_dir = data_path + "train_cropped/"

2. Start training:

python main.py

Analysis

Parameter settings

The corresponsing parameter settings for global and local analysis are saved in the analysis_settings.py

   load_model_dir = 'saved model path'#'./saved_models/vgg19/003/'
   load_model_name = 'model_name'#'14finetuned0.9230.pth'
   save_analysis_path = 'saved_dir_rt'
   img_name = 'prototype_vis_file'# 'img/'
   test_data = "test_dir"
   check_test_acc = False
   check_list =['list of test images'] #"163_Mercedes-Benz SL-Class Coupe 2009/03123.jpg", Could be a list of images

Local Analysis and reasoning process

To produce the reasoning plots:

We analyze nearest prototypes for specific test images and retrieve model reasoning process for predictions:

# this function provdes results for model's reasoning and local analysis

python local_analysis.py -gpuid 0

Global Analysis

To produce the global analysis plots:

This following file finds nearest patches for each prototype to ensure the prototypes are semantically consistent across samples in train and test data:

python global_analysis.py -gpuid 0

Location Misalignment

To run the experiment, you would also need cleverhans

pip install cleverhans

Parameter settings

All the parameters used for reproducing our results on location misalignment are stored in adv_settings.py

load_model_path = "."
test_dir = "./cub200_cropped/test_cropped"
model_output_dir = "." # dir for saving all the results 

To run the adversarial attack and retrieve the results

cd ./spatial_alignment_test
python run_adv_test.py # as default, we ran experiment over entire test set

Model Zoo

CUB-200-2011 Dataset Results

We provide checkpoints after projection and last layer finetuning on CUB-200-2011 dataset.

Model Version	Backbone	Resolution	Accuracy	Checkpoint
ProtoViT-T	DeiT-Tiny	224×224	83.36%	Download
ProtoViT-S	DeiT-Small	224×224	85.30%	Download
ProtoViT-CaiT	CaiT_xxs24	224×224	86.02%	Download

Acknowledgments

This implementation is based on the timm, ProtoPNet repository and its variations. We thank the authors for their valuable work.

Contact Info

If you have any questions regarding the paper or implementations, please don't hesitate to email us: [email protected]

Feel free to ⭐ the repo, contribute, or share it with others who might find it useful!

Citation

If you find this work helpful in your research, please also consider citing:

@inproceedings{ma2024interpretable,
  title={Interpretable Image Classification with Adaptive Prototype-based Vision Transformers},
  author={Ma, Chiyu and Donnelly, Jon and Liu, Wenjun and Vosoughi, Soroush and Rudin, Cynthia and Chen, Chaofan},
  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems}
}
}

License

This project is licensed under the MIT License.