IndoxGen-Torch: Advanced GAN-based Synthetic Data Generation Framework

Official Website • Documentation • Discord

NEW: Subscribe to our mailing list for updates and news!

Overview

IndoxGen-Torch is a cutting-edge framework for generating high-quality synthetic data using Generative Adversarial Networks (GANs) powered by PyTorch. This module extends the capabilities of IndoxGen by providing a robust, PyTorch-based solution for creating realistic tabular data, particularly suited for complex datasets with mixed data types.

Key Features

• GAN-based Generation: Utilizes advanced GAN architecture for high-fidelity synthetic data creation.
• PyTorch Integration: Built on PyTorch for efficient, GPU-accelerated training and generation.
• Flexible Data Handling: Supports categorical, mixed, and integer columns for versatile data modeling.
• Customizable Architecture: Easily configure generator and discriminator layers, learning rates, and other hyperparameters.
• Training Monitoring: Built-in patience-based early stopping for optimal model training.
• Scalable Generation: Efficiently generate large volumes of synthetic data post-training.

Installation

pip install IndoxGen-Torch

Quick Start Guide

Basic Usage

from indoxGen_pytorch import TabularGANConfig, TabularGANTrainer
import pandas as pd

# Load your data
data = pd.read_csv("data/Adult.csv")

# Define column types
categorical_columns = ["workclass", "education", "marital-status", "occupation",
                       "relationship", "race", "gender", "native-country", "income"]
mixed_columns = {"capital-gain": "positive", "capital-loss": "positive"}
integer_columns = ["age", "fnlwgt", "hours-per-week", "capital-gain", "capital-loss"]

# Set up the configuration
config = TabularGANConfig(
    input_dim=200,
    generator_layers=[128, 256, 512],
    discriminator_layers=[512, 256, 128],
    learning_rate=2e-4,
    beta_1=0.5,
    beta_2=0.9,
    batch_size=128,
    epochs=50,
    n_critic=5
)

# Initialize and train the model
trainer = TabularGANTrainer(
    config=config,
    categorical_columns=categorical_columns,
    mixed_columns=mixed_columns,
    integer_columns=integer_columns
)
history = trainer.train(data, patience=15)

# Generate synthetic data
synthetic_data = trainer.generate_samples(50000)

Advanced Techniques

Customizing the GAN Architecture

custom_config = TabularGANConfig(
    input_dim=300,
    generator_layers=[256, 512, 1024, 512],
    discriminator_layers=[512, 1024, 512, 256],
    learning_rate=1e-4,
    batch_size=256,
    epochs=100,
    n_critic=3
)

custom_trainer = TabularGANTrainer(config=custom_config, ...)

Handling Imbalanced Datasets

original_class_distribution = data['target_column'].value_counts(normalize=True)
synthetic_data = trainer.generate_samples(100000)
synthetic_class_distribution = synthetic_data['target_column'].value_counts(normalize=True)

Configuration and Customization

The TabularGANConfig class allows for extensive customization:

• input_dim: Dimension of the input noise vector
• generator_layers and discriminator_layers: List of layer sizes for the generator and discriminator
• learning_rate, beta_1, beta_2: Adam optimizer parameters
• batch_size, epochs: Training configuration
• n_critic: Number of discriminator updates per generator update

Refer to the API documentation for a comprehensive list of configuration options.

Best Practices

1. Data Preprocessing: Ensure your data is properly cleaned and normalized before training.
2. Hyperparameter Tuning: Experiment with different configurations to find the optimal setup for your dataset.
3. Validation: Regularly compare the distribution of synthetic data with the original dataset.
4. Privacy Considerations: Implement differential privacy techniques when dealing with sensitive data.
5. Scalability: For large datasets, consider using distributed training capabilities of TensorFlow.

Roadmap

• Implement basic GAN architecture for tabular data
• Add support for mixed data types (categorical, continuous, integer)
• Integrate early stopping and training history
• Implement more advanced GAN variants (WGAN, CGAN)
• Add built-in privacy preserving mechanisms
• Develop automated hyperparameter tuning
• Create visualization tools for synthetic data quality assessment
• Implement distributed training support for large-scale datasets

Contributing

We welcome contributions! Please see our CONTRIBUTING.md file for details on how to get started.

License

IndoxGen-Torch is released under the MIT License. See LICENSE.md for more details.