lupalab/ace

Official code for the paper "Arbitrary Conditional Distributions with Energy".

ace's Issues

Dataset classes for MNIST and UCI Adult

Hi there! I'd like to reproduce the results regarding high dimensional data (MNIST) and mixed continuous and discrete data (UCI Adult dataset). I couldn't find the corresponding tfds classes in the code. Could you provide those, please? That would be really helpful, thanks!

The DEformer as an alternative approach to order-agnostic distribution estimation

Congratulations on your paper being accepted to NeurIPS! I wanted to bring my DEformer project to your attention because I thought you might find it interesting as an alternative approach to order-agnostic distribution estimation. The main idea behind the DEformer is to include each feature's identity alongside its value in the input, which allows neural architectures designed for sequential data (like Transformers) to be used for order-agnostic distribution estimation without modification. On the POWER dataset, the DEformer achieves a NLL of -0.68, while, like ACE, it can effectively fill-in arbitrary missing pixels of MNIST images.