learn_BP

This code accompanies the paper Belief Propagation Neural Networks

This figure shows the computaion graph of an iterative BPNN-D layer run until convergence:

Dependencies

PyTorch
PyTorch Geometric
mrftools (not required, can comment out imports)
libDAI used with Ising model experiments:

junction tree algorithm for obtaining the exact partition function
loopy belief propagation for comparison

See environment.yml for appropriate versions. Instructions to build a conda environment from a .yml file.

Training the BPNN on Ising models

Run learn_BP_spinGlass.py with mode="train"

Training the baseline GNN on Ising models

Run learn_GNN_spinGlass.py with mode="train"

Making figures/tables

Compiling spin glass tables into the figure in the paper

learn_BP/data/experiments/make_figure_from_table.py (save as .eps, as shown here for high quality figures)

Run learn_BP_spinGlass.py with mode="test"

To make the table call create_many_ising_model_figures()
To make a single figure, call create_ising_model_figure()
Set BPNN_trained_model_path and GNN_trained_model_path to the appropriate trained model paths
create latex table with data/experiments/make_latex_table.py

Ising Model convergence plots (see commit e543da652af9ffa4ac0174c48246fbaa53ff4118/e00674bdb2962133e402dfe6a690dab81fcd20b2 on branch bpnn_d_experiments):

set PLOT_CONVERGENCE=True in nn_models.py (and set whether BP or BPNN data is recorded)
run learn_BP_spinGlass.py (loading a trained BPNN model, or pre-training for BP)
run plot_convergence/make_plots.py
get BPNN estimate improvements on converged BP models by manually entering estimates to plot_convergence/check_errors.py

SAT table/figures

Make the SAT tables in latex form from wandb results using data/wandbCSV_to_latexTable.py
run data/compare_BPNNvsHashing_runtimes.py (for the figuring comparing approximation of BPNN vs. F2 vs. ApproxMC and for the time ratio comparison)

Make Fig 1, aggregating BPNN/GNN/BP comparison with data/experiments/make_figure_from_table.py

SAT data info