Autoencoding Probabilistic Circuits

Code release for the paper Tractable Representation Learning with Probabilistic Circuits. The project implements autoencoding probabilistic circuits (APCs): a tractable PC encoder paired with a neural decoder, differentiable sampling (SIMPLE) for end-to-end training, and evaluation tooling for missing-data robustness, downstream tasks, and knowledge distillation from neural autoencoders.

The repository is organized around the apc/ package and Hydra configs in conf/.

Installation

• Python >= 3.10 and a recent PyTorch build (CUDA recommended for images).
• Clone the repo and install:

  pip install .

• Set data/result roots or rely on defaults (~/data and ~/results):

  export DATA_DIR=/path/to/datasets
  export RESULTS_DIR=/path/to/outputs

Quickstart

Train and evaluate an APC on MNIST-32 with the convolutional PC encoder and neural decoder:

python apc/main.py data=mnist-32 model=apc group=test

Configuration Notes

• Hydra configs: default config is conf/config.yaml; datasets and model presets live under conf/data/ and conf/model/.
• Logging: WandB is off by default (wandb=false); enable with your project name if desired. Logs and checkpoints are stored in ${RESULTS_DIR}/${dataset}/${group}/${subgroup}/${timestamp}_${tag}.
• Environment variables: DATA_DIR and RESULTS_DIR override storage roots;

Citation

If you use this code, please cite:

TODO