AutoGAN provides a novel neural architecture search (NAS) framework specifically for Generative Adversarial Networks (GANs). It aims to automate the discovery of high-performing GAN architectures for unconditional image generation tasks, targeting researchers and practitioners in computer vision and deep learning. The framework has demonstrated state-of-the-art results on datasets like CIFAR-10.

How It Works

AutoGAN employs an RNN controller to search for optimal GAN architectures within a defined search space. This approach allows for the automated exploration and selection of network components and configurations, leading to architectures that achieve competitive performance metrics like FID and Inception Score. The framework is designed to be efficient in discovering effective GAN designs without extensive manual tuning.

Quick Start & Requirements

• Install: pip install -r requirements.txt
• Prerequisites: Python >= 3.6, PyTorch >= 1.1.0. Requires downloading FID statistics to ./fid_stat.
• Setup: Requires downloading pre-calculated FID statistics.
• Search & Train: Run sh exps/autogan_search.sh for search, then python train_derived.py with the discovered architecture vector.
• Paper Results: Run sh exps/autogan_cifar10_a.sh for CIFAR-10.
• Testing: Use python test.py with a specified model path.
• Pre-trained Models: Available via Google Drive.

Highlighted Details

• Achieved a record FID score of 12.42 and Inception score of 8.55 on CIFAR-10.
• Supports unconditional image generation on CIFAR-10 and STL-10.
• Utilizes an RNN controller for the neural architecture search process.
• Code released for the search component in October 2019.

Maintenance & Community

The project is associated with the VITA Group and the paper was presented at ICCV 2019. No specific community channels or active maintenance signals are evident from the README.

Licensing & Compatibility

The README does not explicitly state a license. The code relies on components from OpenAI's Improved GAN and TTUR, which have their own licenses. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The project appears to be a research artifact from 2019, and its current maintenance status or compatibility with newer PyTorch versions is not indicated. The setup requires manual downloading of statistics files.