This sub-repo contains code to train a conditional Determinstic and Structured-Decomposable probabilistic circuits on the 4x4 grid experiment from the semantic loss paper.

To run it, use

python -W ignore grid_net.py --iters 20000 --data test.data

where --iters can be set to any number of iterations

Most of the changes pertaining to the work in this repo can be found in cmpe.py, pypsdd/pypsdd/sdd.py as well as GatingFunction.py

(One?) Recipe for implementing Semprola:

• Create a PSDD who's logical base is true
• Create an SDD encoding the required logical constraint
• Overparameterize the PSDD
• Multiply the PSDD with the SDD [Missing]

TODOs

• Vectorize MPE computation -- other computations are vectorized.
• Overparameterize the PSDD, making it non-deterministic [Done -- needs more testing]