rrn-17.csv.xz is a compressed csv file with the hardest 17 hints RRN test set problems.

train-sufficient-statistics/ contains precomputed sufficient statistics built from 8,000 samples from the RRN training set. These can be computed in linear time in the sample size. A is for exact data, B is for LeNet decoded images.

lambdas/ contains precomputed values of lambda (using a dedicated grid search in lambda log-space).

PEMRF.py is the implementation of PE_MRF/ADMM with L1, L2 and L1/L2 norms.

The main programs are Sudoku-train-validate.py and Visual-Sudoku-train-validate.py. They will train the CFN parameters from the sufficient statistics and the value of lambda, produce a CFN and solve test set instances using it (either from exact hints or from Visual hints).

Learned-CFN-13000-3.5624615290574217.cfn.gz contains a CFN learned with 13,000 samples. We will have a look into it.

MNIST_test_marginals contains the 'pickled' output of LeNet on every test digit in MNIST test set (organized as a per digit list of outputs per instance of the digit). An output itself is a numpy array of negated logits, that can directly be used as scores to minimize).

MNIST_test_indices is just a mapping to find the image index that corresponds to evey output.

These 2 files have been produced using MNIST_train.py (PyTorch).

The program itself is in MNIST_sudoku_CP.py. It takes one argument: a grid number in SAT-Net test set (0 to 999).