README.md

Uncertainty-Aware Pseudo-labeling for Quantum Calculations

This repository hosts the code for

Uncertainty-Aware Pseudo-labeling for Quantum Calculations

By Kexin Huang, Mykola Bordyuh, Vishnu Sresht, Brajesh Rai.

Install

0. Create a conda environment:

conda create -n Pseudo python=3.8
conda activate Pseudo

1. Git clone the repo:

git clone https://github.com/PfizerRD/pseudo.git 
cd pseudo

2. Install PyTorch & Pytorch-Geometric

3. Install other libraries via

pip install -r requirements.txt

Dataset

PSEUDO relies on QM9 and PC9 datasets. Please download the processed dataset through this link. Unzip it and place the folder "dataset" in the repository.

Train

python train.py --label homo \ # molecule target
                --setting low_data \ # low data setting or standard fully supervised setting
                --training_fraction 0.01 \ # fraction of training QM9, rest is used as unlabeled
                --pseudo_label True \ # whether or not to use pseudo-label or standard
                --model dimenet \ # model backbone, select from schnet/dimenet
                --iteration 10 \ # outer loop number of episodes
                --epoch 100 \ # inner loop number of epochs for each episode
                --initial_train_epoch 100 \ # Number of training epochs for the first episode on labeled data
                --batch_size 128 \ # batch size
                --lr 0.001 \ # learning rate
                --lr_decay_factor 0.5 \ # step LR decay factor
                --lr_decay_step_size 25 \ # steps to decay LR
                --uncertainty_type epistemic \ # uncertainty type select from epistemic and aleatoric
                --evi_lambda 0.5 \ # evidental loss regularization coefficient                

For more options, checkout the parser in train.py. Also checkout a demo test in the notebook demo.ipynb.

To reproduce full data setting:

Click here for the code!

python train.py --label homo \
                --model dimenet \
                --pseudo_label True \
                --setting standard \
                --pseudo_ensemble True \
                --lr 0.001 \
                --lr_decay_factor 0.5 \
                --lr_decay_step_size 25 \
                --iteration 15 \
                --evi_lambda 0.5 \
                --epoch 75 \
                --batch_size 128 \
                --uncertainty_type epistemic

To reproduce low data setting, replace the training_fraction value with the low data fraction you consider:

Click here for the code!

python train.py --label homo \
                --model dimenet \
                --pseudo_label True \
                --setting low_data \
                --training_fraction 0.1 \
                --initial_train_epoch 300 \
                --pseudo_ensemble True \
                --lr 0.001 \
                --lr_decay_factor 0.5 \
                --lr_decay_step_size 15 \
                --iteration 15 \
                --evi_lambda 0.5 \
                --epoch 50 \
                --batch_size 128 \
                --uncertainty_type epistemic

Citation

@inproceedings{huang2022uncertainty,
  title={Uncertainty-Aware Pseudo-labeling for Quantum Calculations},
  author={Huang, Kexin and Sresht, Vishnu and Rai, Brajesh and Bordyuh, Mykola},
  booktitle={The 38th Conference on Uncertainty in Artificial Intelligence},
  year={2022}
}

Acknowledgement

We use DIG library as the backbone for SchNet and DimeNet++ implementation.