/GRL-SVO

This project proposes two Reinforcement Learning (RL) approaches combined with Graph Neural Networks (GNN) for Suggesting Variable Order (SVO).

Primary LanguagePython

GRL-SVO

Setup

  • GRL-SVO: source code and datasets;
  • cad.tar.gz: conda environment;
  • requirements.txt: required python packages.

Auto pip install

  1. conda create -n cad python=3.7.0

  2. conda activate cad

  3. pip install -r requirements.txt

  4. cd GRL-SVO/cad_order_gym

  5. pip install -e .

Notes

We are unable to provide the dataset anonymously due to its large size (> 1GB). Training and testing cannot run, we are very sorry for that.

Training

Position: GRL-SVO/

For GRL-SVO(NUP)

python train_grl_svo_nup.py

For GRL-SVO(UP)

python train_grl_svo_up.py

As GRL-SVO(UP) will interact with MAPLE, a new trajectory may make errors (we have stored most trajectories used in previous training, but not all). We have prepared some trained models in ./models/ diretory. One can still test the models.

Testing

Position: GRL-SVO/

For GRL-SVO(NUP)

python predict_grl_svo_nup.py

The result is stored in ./predict_result/nup_rand3/result_nup.log

cat ./predict_result/nup_rand3/result_nup.log

For GRL-SVO(UP)

python predict_grl_svo_up.py

GRL-SVO(UP) must interact with MAPLE during the testing process. If MAPLE is not installed, predict_grl_svo_up.py can only run one step, and the intermediate information stores in GRL_SVO/predict_result/up_rand3.

cat GRL_SVO/predict_result/up_rand3/action_1.log