This folder contains materials to reproduce the submissions files that contributed to the final solution which finished in 5th place of the kaggle competition Predicting Molecular Properties

See the complete details of our solution here

Idea

The architecture is based on the paper Graph Networks as a Universal Machine Learning Framework for Molecules and Crystals with some changes to adapt to the use cases.

Requirement

1. Hardware

• CPU: 12 cores
• Memory: 32 GB
• GPU: 1 GTX 2080 Ti

2. Software:

• CUDA 10.0 CuDNN 7.4
• Anaconda Python 3.7 2019.03
• pytorch==1.1.0
• pytorch_geometric
• openbabel, rdkit

How to train?

• Generate data graph for every molecule:
  python data/data.py --data_dir=kaggle/data/csv --split_dir=kaggle/data/split --graph_dir=kaggle/data/graph

• Training:

  • first model: python train.py --out_dir=kaggle/output --model=model1 --optim=adam
  • second model: python train.py --out_dir=kaggle/output --model=model2 --optim=adam
  • third model: python train.py --out_dir=kaggle/output --model=model1 --optim=ranger

How to make prediction?

• python submit.py --out_dir=kaggle/output/submit --model=model1 --checkpoint=kaggle/output/checkpoints/checkpoint_01.pth

Credit

Credit to Cher Keng Heng for his starter kit code.