zouxiaochuan / molnet_kddcup2021

This repository is used for submitting code of team no_free_lunch for pcqm4m track of ogb-lsc

this repository has two parts:

1. /pcqm4m is a clean version of MolNet
2. /reproduce is used for reproducing test score

how to reproduce test score

Due to the file size limitation of github, model files cannot be saved to github, you can download full package by https://kddcup2021.s3.ap-southeast-1.amazonaws.com/pcqm4m_fin.tar and then enter /reproduce directory and run "./run_predict.sh", it will take about 8 hours to run on single core single card machine. If you have any problems of downloading the full package, please contact me.

how to do training

1. first download pcqm4m dataset from the official site.
2. enter /pcqm4m and run "python create_dataset_new.py ../../datasets" to generate dataset. ../../datasets is where you put pcqm4m dataset folder.
3. enter /pcqm4m and run "python run_molnet.py", it will use default dataset path of ../../datasets. And model is saved in lightning_logs directory.