SparsePC

Code of the NeurIPS 2022 paper Sparse Probabilistic Circuits via Pruning and Growing.

Environment

1. Install Juilia version 1.8.
2. Run the following command to install required packages.

   julia env.jl
   

Datasets

Run the following command to download all datasets. And enter y each time when hinted to.

julia --project download_data.jl

Experiments

Run the following commands to generate runnable scripts.

1. Learning hidden Chow-Liu trees (HCLT).

   bash sh/gen_hclt.sh
   

   After learning, postprocess checkpoints to pick the best HCLTs as initial structures and save them in circuits\.

   julia --project post_process.jl
   

2. Learning sparse PC structure via pruning and growing (Section 6.1).

   bash sh/gen_loop.sh
   

3. Distilling large PC by iteratively pruning (Section 6.2).

   bash sh/gen_compress.sh
   

Citation

To cite this paper, please use

@inproceedings{DangNeurIPS22,
  author    = {Dang, Meihua and Liu, Anji and Van den Broeck, Guy},
  title     = {Sparse Probabilistic Circuits via Pruning and Growing},
  booktitle = {Advances in Neural Information Processing Systems 35 (NeurIPS)},
  month     = {dec},
  year      = {2022},
}

Contact

Please contact [email protected] if you have any questions.