Acceleration of Interpretable Deep Neural Networks

This repository contains the proposed acceleration method – simplified calculation of the loss function – in the research "Acceleration of Interpretable Deep Neural Networks" (ISO).

Repository Overview

This repository includes code implementations for:

• ProtoPNet and its experiment results across different batch sizes (stored in the /saved_models directory).
• ProtoViT and TesNet, with experiment results only for batch size 80 (stored in their respective /saved_models directories).
• EvalProtoPNet, a folder containing the evaluator for ProtoPNet.

Setting Up the Environment

To install the required dependencies for each of these models, run in their subfolders:

pip install -r requirements.txt

---

Instructions for Preparing the Data

1. Download the dataset CUB_200_2011.tgz from Caltech Vision.
2. Unpack the dataset:

   tar -xvzf CUB_200_2011.tgz

3. Crop the images using crop_dataset.py and bounding box information from bounding_boxes.txt (included in the dataset):

   python crop_dataset.py

4. Split the cropped images into training and test sets using split_dataset.py and train_test_split.txt (included in the dataset):

   python split_dataset.py

5. Move the cropped images into appropriate directories:

   mv cropped_train ./datasets/cub200_cropped/train_cropped/
   mv cropped_test ./datasets/cub200_cropped/test_cropped/

6. Augment the training dataset using img_aug.py:

   python img_aug.py

   This will create an augmented training set in:

   ./datasets/cub200_cropped/train_cropped_augmented/
   

---

Instructions for Training the Models

Step 1: Settings

Edit settings.py to set appropriate paths for dataset and other hyperpameters

Step 2: Run Training Scripts

• For training the original models:

  bash main.sh

  or run directly:

  python main.py

• For training the accelerated models:

  bash main_custom.sh

  or run directly:

  python main_custom.py

---

Experiment Setup

All experiments were conducted on an NVIDIA Tesla T4 GPU with 16GB RAM.

---

Acknowledgments

This work builds upon the following repositories:

• ProtoPNet: https://github.com/cfchen-duke/ProtoPNet
• BetterProtoPNet - fork of ProtoPNet with data preprocessing scripts: https://github.com/KrystianJachna/BetterProtoPNet
• ProtoViT: https://github.com/Henrymachiyu/ProtoViT
• TesNet: https://github.com/JackeyWang96/TesNet
• EvalProtoPNet: https://github.com/hqhQAQ/EvalProtoPNet

---