Training and Inference on Any-Order Autoregressive Models the Right Way

This repository contains code for the paper:

Training and Inference on Any-Order Autoregressive Models the Right Way
by Andy Shih, Dorsa Sadigh, Stefano Ermon

<br>

Any-Order Autoregressive Models (AO-ARMs) are a powerful model family that can compute arbitrary conditionals and marginals. Broadly defined, some examples of AO-ARMs are:

• A Deep and Tractable Density Estimator
• BERT
• XLNet
• Arbitrary Conditioning with Energy
• Autoregressive Diffusion Models

We introduce MAC: Mask-Tuned Arbitrary Conditional Models, which improve AO-ARMs by training on a smaller set of univariate conditionals while still maintaining support for efficient arbitrary conditional and marginal inference. In short, MAC improves model performance without sacrificing tractability.

---

Installation

pip install -r requirements.txt

Commands

The current batch sizes assume a GPU with 48GB memory.

ARDM

python image_main.py dataset=CIFAR10 mask.strategy=none mask.order=random batch_size=24

python image_main.py dataset=IMAGENET32 mask.strategy=none mask.order=random batch_size=24

python lang_main.py dataset=TEXT8 mask.strategy=none mask.order=random batch_size=180

MAC

python image_main.py dataset=CIFAR10 mask.strategy=marginal mask.order=spaced mask.normalize_cardinality=True batch_size=24

python image_main.py dataset=IMAGENET32 mask.strategy=marginal mask.order=spaced mask.normalize_cardinality=True batch_size=24

python lang_main.py dataset=TEXT8 mask.strategy=marginal mask.order=spaced mask.normalize_cardinality=True batch_size=180

Citation

"Training and Inference on Any-Order Autoregressive Models the Right Way"
Andy Shih, Dorsa Sadigh, Stefano Ermon
In Proceedings of the 36th Conference on Neural Information Processing Systems (NeurIPS), 2022

@inproceedings{shih2022aoarm,
  author    = {Andy Shih and Dorsa Sadigh and Stefano Ermon},
  title     = {Training and Inference on Any-Order Autoregressive Models the Right Way},
  booktitle = {Advances in Neural Information Processing Systems 35 (NeurIPS)},
  month     = {december},
  year      = {2022},
}