Reproduce “RPGAN: GANs Interpretability via Random Routing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Andrey Voynov, Artem Babenko
RPGAN makes GANs easier to inspect by fixing the input vector and randomly selecting among multiple instances of each generator layer. Researchers vary one layer’s instance while holding others fixed and compare outputs. Across CIFAR-10, Anime Faces, and LSUN bedrooms, intermediate buckets tend to affect semantic variation more and later buckets color variation; these roles depend on dataset and architecture.
Key insights
- Across several image datasets, early, intermediate, and later layers often affect different aspects of generated images, but these roles are not fixed: they depend on both the dataset and architecture.
- The route-based design can generate plausible comparisons while changing a layer, whereas perturbing a standard GAN’s weights often produces poor-quality outputs that are harder to interpret.
- The same structure can be extended by training added layer instances, and linear versions can permit layer fusion; these are exploratory applications rather than broad performance evaluations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.