Repository Overview

This repository is a comprehensive collection of code, models, and scripts focused on fairness and compression in machine learning. It contains implementations and experiments on advanced techniques such as adversarial debiasing, fairness constraints, model pruning, quantization, and knowledge distillation.

Directory Structure

• data/
  Contains raw data files (e.g., adult.csv, adult_test.csv, compas.csv) used for training and evaluation.

• models/
  Stores pre-trained models, including baseline, debiased, pruned, quantized, and distilled variants.

• scripts/
  Contains training, evaluation, pruning, quantization, and fine-tuning scripts. For example, scripts/basicFCN/train.py and scripts/basicFCN/train_fairness.py offer different training strategies.

• src/
  Provides core logic, model definitions, and data loaders (e.g., src/simpleFCNN.py and src/adultIncome.py).

• notebooks/
  Contains Jupyter notebooks for experimentation, visualization, and analysis.

• LICENSE
  Licensed under the MIT License.

• CONVENTIONS.md
  Documents repository-specific guidelines and conventions.

Getting Started

1. Installation
   Install the required dependencies:

   pip install -r requirements.txt

2. Data Preparation
   Place your raw data in the data/raw/ directory (e.g., adult.csv, adult_test.csv, compas.csv).

3. Training

   • For baseline training:

     python scripts/basicFCN/train.py --dataset adult

   • For adversarial debiasing:

     python scripts/basicFCN/train_fairness.py --dataset adult --debiasing adversarial --protected_attr sex --adv_weight 0.5

   • For training with fairness constraints:

     python scripts/basicFCN/train_fairness.py --dataset adult --debiasing fairness_constraint --protected_attr sex --fairness_type demographic_parity --lambda_fair 0.3

4. Evaluation
   Use the provided scripts and notebooks to evaluate model performance and fairness metrics.

Contributing

Please review the guidelines in CONVENTIONS.md before contributing to ensure consistency with the project's conventions.

License

This project is licensed under the MIT License. See the LICENSE file for details.