zhen-he / ibp_vae

mean-field and structured VAEs for the IBP

Indian Buffet Process VAEs

This repository contains the code for the paper "Structured Variational Autoencoders for the Beta-Bernoulli Process", by Jeffrey Ling, Rachit Singh, and Finale Doshi-Velez. The paper is itself forthcoming on the ArXiv, but a workshop version is available at the AABI NIPS workshop page.

This codebase is somewhat incomplete, missing some code that we haven't had time to clean up. We're working on putting all of the code up as soon as possible, and will do so over the next few weeks. Please let us know if you run into problems by making an issue.

We can also be contacted at [email protected], and [email protected], and feel free to send us an email.

Quickstart

To run the code with GPU support, navigate to src/lgamma and do ./make.sh.

Navigate to base directory and run

python scripts/run_s_ibp_concrete.py --savefile testsave

--savefile is a required argument. Models will be saved under models/ (last, and best over epochs) and train, valid, test curves and timings are saved under runs/.

Use --help to see arguments.

Note about compilation

We're working on incorporating all of the required CUDA functionality in PyTorch master, but until then the compilation process is likely a little tricky. For example, you'll almost definitely need to change sm_35 in make.sh to your card's compute capability and the CUDA_PATH must be customized. Finally, one likely change you'll need to make is to fix build.py to have the correct cuda_path and include_dirs. Please feel free to send me an email if you have trouble compiling - I'm happy to help.