udit-rawat / gan-n

Gan-n

Project Summary

Overview

This project implements a Generative Adversarial Network (GAN) to generate synthetic data for imbalanced datasets. The generated data is intended to augment the minority class in binary classification tasks, improving model performance.

Key Components

1. Data Preprocessing

   • Load and preprocess data from a CSV file.
   • Separate minority and majority classes.
   • Use SMOTEEN to generate additional samples for the minority class.

2. GAN Architecture

   • Generator: Maps random noise vectors (latent space) to synthetic data samples.
   • Discriminator: Distinguishes between real and synthetic data samples.

3. Training the GAN

   • Train the generator and discriminator iteratively.
   • Use the Binary Cross-Entropy loss function.
   • Optimize using Adam optimizer.

4. Main Script

   • Load and preprocess data.
   • Train the GAN.
   • Save the trained generator model.

Files and Directories

• src/: Contains modules for GAN training and data preprocessing.
• models/: Directory to save trained GAN models.
• Datasets/: Contains the CSV data files for training.

How to Run

1. Set Up Environment

   • Install dependencies: pip install -r requirements.txt

2. Train the GAN

   • Run the main script: dvc repro, to track the chnages.
   • Specify the target column and data file path in the script.

Example Usage

python3 main.py

Dependencies

• PyTorch
• NumPy
• SMOTEEN (Imbalanced-learn)

Goals

• Address class imbalance in datasets.
• Improve classification performance by augmenting training data with synthetic samples.
• Utilizes docker and DVC for data version controlling, re-experimenation and containerization