sebastianjaszczur / deepsat

Solve SAT with neural networks

Generating data

dump_dataset.py -o <num_of_samples> -c <num_of_variables> -j <random_seed>

Training

Run on TPU like this:

python neurosat_tpu.py \
    --use_tpu=True \
    --tpu=$TPU_NAME \
    --train_file=$TRAINNIG_FILE \
    --test_file=$TEST_FILE \
    --train_steps=1200000 \
    --test_steps=80 \
    --model_dir=$MODEL_DIR \
    --export_dir=$EXPORT_DIR \
    --variable_number=30 \
    --clause_number=300 \   # 10 * variable_number
    --train_files_gzipped=False \
    --batch_size=128 \
    --export_model \
    --attention=$ATTENTION \
    --level_number=$LEVEL_NUMBER

Examples and hyperparameters can be read in notebooks/iclr2019/tpu_grid.sh.

Evaluation with DPLL or CDCL

• For DPLL with 1000 step limit see notebooks/iclr2019/dpll_1000_steps.ipynb.
• For DPLL without a step limit see notebooks/iclr2019/hybrid_dpll.ipynb.
• For CDCL without a step limit see notebooks/iclr2019/hybrid_cdcl.ipynb.