A pytorch implementation of NeuraSAT(github, paper)

In this implementation, we use SR(U(10, 40)) for training and SR(40) for testing, achieving the same accuracy 85% as in the original paper. The model was trained on a single K40 gpu for ~3 days following the parameters in the original paper.

Before running any experiment scripts, you need to generate the SAT problem instances using the src/data_maker.py script:

python src/data_maker.py <out_dir> <gen_log> <n_pairs> <max_nodes_per_batch> [other options]

For example, to generate the SR(40) dataset for testing, you could run something like:

python src/data_maker.py data/val/sr40.pkl data/val/sr40.log 1000 50000 --min_n 40 --max_n 40

To generate the SR(U(10, 40)) dataset for training:

python src/data_maker.py data/train/sr10-40.pkl data/train/sr10-40.log 10000 50000 --min_n 10 --max_n 40

• <out_dir>: Output path for the generated pickle file containing the SAT problems
• <gen_log>: Path for the generation log file
• <n_pairs>: Number of SAT problem pairs to generate
• <max_nodes_per_batch>: Maximum number of nodes per batch
• --min_n and --max_n: Range of variables in the SAT problems (SR(40) means min_n=max_n=40)
• --p_k_2 and --p_geo: Distribution parameters for clause generation (defaults: 0.3 and 0.4)

You'll also need to generate validation datasets for training scripts. For example:

# Generate validation dataset for SR(40)
python src/data_maker.py data/val/sr40.pkl data/val/sr40.log 1000 50000 --min_n 40 --max_n 40

The repository includes several shell scripts for training and evaluating models on different dataset configurations:

• run_sr3to10.sh: Trains on SR(3-10) dataset
• run_sr10to40.sh: Trains on SR(10-40) dataset
• run_sr40to100.sh: Trains on SR(40-100) dataset
• run_sr200to500.sh: Trains on SR(200-500) dataset

Each training script performs three steps:

1. Generates the required training dataset using data_maker.py
2. Generates a validation dataset if it doesn't exist yet
3. Runs the training using src/train.py

The scripts have been configured to automatically create all necessary directories and datasets.

To run a training script (example):

./run_sr10to40.sh

To evaluate a trained model on SR(40) dataset, use the run_eval.sh script:

./run_eval.sh

This script will first generate an SR(40) dataset specifically for evaluation and then run the evaluation using src/eval.py.