lupalab/posterior-matching

Official code for the NeurIPS 2022 paper "Posterior Matching for Arbitrary Conditioning".

posterior-matching's Issues

Mismatch between config variable names and network constructor arguments

Hi @rystrauss, thank you for such an amazing repository. I've found it very easy to begin working with this library after cloning it, so I really appreciate the time and effort spent in making it very usable (if anything getting JAX to work correctly with my GPU setup was the hardest part). While trying to recreate some of the paper's experiments, I noticed a bit of a strange behaviour when changing the network's encoder or decoder parameters in some of the configs in the repo. In short, it seemed like some of the results did not change when varying the configuration of the encoders and decoders for some of the configs.

After some digging, I noticed that in vae.py (and similarly for vade.py) the encoder's parameters are passed via the encoder_net_config argument and the decoder's parameters are passed via the decoder_net_config argument. However, several of the configs in this repository actually use encoder_config and decoder_config to pass these parameters, leading to the network constructors pretty much ignoring these parameters and using the default values instead. Because every argument in these constructors has a default value, it is hard to notice this is happening as no error would be thrown.

The configs that seem to be affected are:

$ grep -r -E '(encoder_config)|(decoder_config)' configs/ -l
configs/pm_vae_bsds.py
configs/pm_vae_gas.py
configs/pm_vae_hepmass.py
configs/pm_vae_miniboone.py
configs/pm_vae_power.py

Fixing this would just require a simple parameter name change in those configs so this is not a big issue. I am happy to fix it myself and make a PR but I wasn't sure what the contribution policy was for this repository. In the meantime, I thought that just raising a ticket may save time for others using this repository in case they stumble upon this issue. Thanks so much again for such a wonderful library!