HyperGAN/HyperGAN

What it solves

HyperGAN simplifies the process of building, training, and sharing Generative Adversarial Networks (GANs) using PyTorch. It provides a composable framework that removes the complexity of manual network architecture design and data preparation for developers, researchers, and artists.

How it works

The project provides a high-level API and a Command Line Interface (CLI) that allow users to create models based on JSON configuration templates. It handles data loading from folders of images (including unclean data) and supports accelerated parallel training backends. Models can be exported as ONNX models for use in other projects.

Who it’s for

Developers, AI researchers, and digital artists who want to generate synthetic data or images without needing to write extensive PyTorch code from scratch.

Highlights

• Composable Architecture: Define custom architectures via JSON or replace components with custom PyTorch modules.
• Flexible Training: Supports unsupervised learning, transfer learning, and online learning.
• Cross-Platform: Works on Windows, OSX, and Linux.
• Data Resilience: Automatically handles image resizing and cropping to manage unclean datasets.
• Model Sharing: Built-in support for sharing and using trained models across different environments.