Deep Explanation Penalization (CDEP)

This repository is a reimplementation of deep-explanation-penalization in Python 3.8 and TensorFlow 2.4 .

This work was born out of a submission to the ML reproducibility Challenge, Spring 2021 edition. Please check out the Wiki for more details on the experiments carried out and more.

Getting Started

Installation of Packages

Create a virtual environment and install all python packages using

pip install requirements.txt

Isic-skin-cancer

Navigate into the ISIC skin cancer directory

cd isic-skin-cancer/ISIC-skin-cancer/ 

Dataset Download and Preprocessing

Download The dataset along with metadata and preprocess

python 00_download_metadata.py
python 01_download_images_multiproc.py
python 02_sort_images.py

Calculate CD features after propagating through the main body of VGG-16

python 03_calculate_pretrained.py

Training and Validation

python train_CDEP.py

Stanford Sentiment Dataset

Navigate into the text directory

cd text/

Dataset Download and Preprocessing

Download the Glove embeddings

python download_glove.py

Create the random variant of the SST dataset

python 00_make_decoy.py 

Create the gender variant of the SST dataset

python 01_make_gender.py

Create the biased variant of the SST dataset

python 03_make_bias.py

Training and Validation

python train_all.py

Trains and records the results of all experiments

DecoyMNIST

Navigate to the DecoyMNIST directory

cd mnist/DecoyMNIST/

Download and process the dataset

python 00_make_data.py

Train on the dataset

python 01_train_all.py