Learning to Group Auxiliary Datasets for Molecule

This is our PyTorch implementation for the paper:

Tinglin Huang, Ziniu Hu, and Rex Ying (2023). Learning to Group Auxiliary Datasets for Molecule. Paper in arXiv. In NeurIPS'2023, New Orleans, USA, Dec 10-16, 2023.

Dataset preparation

cd dataset
wget https://raw.githubusercontent.com/snap-stanford/ogb/master/ogb/graphproppred/master.csv
mkdir qm8
cd qm8
wget https://deepchemdata.s3-us-west-1.amazonaws.com/datasets/qm8.csv
cd ..
mkdir qm9
cd qm9
wget https://deepchemdata.s3-us-west-1.amazonaws.com/datasets/qm9.csv

Environment Requirement

The code has been tested running under Python 3.9.16. The required packages are as follows:

• pytorch == 1.13.1+cu117
• torch_geometric == 2.3.1
• ogb == 1.3.6
• rdkit == 2023.03.2
• pandas == 1.3.1
• cython == 3.0.0

Once you finished these installation, please run pip install -e .

Run the code

• Run all combinations of custom dataset pairs (upsample/downsample training instances)
  • ./scripts/all_pairs.sh
• Pretrain graphormer on pcqm4mv2 dataset
  • ./scripts/pretrain_graphormer.sh
• Run molgroup example
  • ./scripts/dataset_grouping.sh
• Run an example of vanilla GIN with dataset combination:
  • ./scripts/example_gin.sh
• Run an example of pretrained Graphormer with dataset combination:
  • ./scripts/example_graphormer.sh

The description of the hyperparameters can be found in the utils/utils.py file. The hyperparameter datasets is a list of datasets' names (e.g., ogbg-molbbbp ogbg-molfreesolv), inciding the datasets to be used for dataset combination. Specifically, the first dataset in the list will be considered as the target dataset.

Citation

If you want to use our codes and datasets in your research, please cite:

@inproceedings{huang2023learning,
  author    = {Tinglin Huang and 
              Ziniu Hu and
              Rex Ying},
  title     = {Learning to Group Auxiliary Datasets for Molecule},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2023}
}