This repository contains some experiments for Applying self pre-training method to GNN for quantum chemistry.

• Python 3.8

# Getting data
curl -sSL "https://www.dropbox.com/s/fifvs2gpdnocxxr/qm9.parquet?dl=1" > ./data/qm9.parquet

Move under src/.

Estimate epochs to train.

python run_pre_train.py params/pre_train/001.yaml

Train with all data. It will take 200 epochs.

python run_pre_train.py params/pre_train/002.yaml

Move under src/.

Train without self pre-training (baseline).

python run_train.py params/train/001.yaml

Edit config to use pre-trained weight.

# src/params/train/002.yaml
pretrained_ckpt_path: xxx

Train with self pre-training.

python run_train.py params/train/002.yaml

• RAdam is used as an optimizer, so that no warmup is used.
• EMA model is used for evaluation, so that no annealing is used.
• Is should be evaluated with full cross validations with 4 or 5 folds, but I used only 1/4 folds because of no time.