Transfer learning for moleculer property prediction

Datasets for molecular property prediction can be found here (This link for downloading).

pytorch >= 1.8.0

torch_geometric >= 2.0.3

rdkit >= 2019.03.1.0

tqdm >= 4.31

1. pre-training

python pretraining.py

2. fine-tuning

python finetune.py --input_model_file <model_path> --dataset <dataset_name>

Results in the paper can be reproduced by running sh finetune.sh <dataset_name> using the pre-trained model in ./init_weights/pretrained.pth. Most hyper-parameters are shared across datasets. The differences can be found in finetuning.py.

The implementation is based on the codes in Hu et al: Strategies for Pre-training Graph Neural Networks