Mimicry is a lightweight PyTorch library aimed towards the reproducibility of GAN research.
Comparing GANs is often difficult - mild differences in implementations and evaluation methodologies can result in huge performance differences. Mimicry aims to resolve this by providing:
- Standardized implementations of popular GANs that closely reproduce reported scores
- Baseline scores of GANs trained and evaluated under the same conditions
- A framework for researchers to focus on implementation of GANs without rewriting most of GAN training boilerplate code, with support for multiple GAN evaluation metrics.