Distributions and estimators for binary latent variables in Pytorch.
The estimators only use the torch.distributions package, and can be used in any codebase without additional libraries. A simple experiment with a VAE on MNIST is included.
Available estimators:
- Pathwise estimator
- Score function estimator
- ARM estimator
- DisARM estimator
Future additions:
- More binary relaxations
- Control variates for SFE
- Multi-sample estimates
- Straight-through estimator