A Merged Molecular Representation Learning
Pre-training BERT with molecular data for learning merged molecular representation.
Demo Video: https://youtu.be/efDLkbdXvRg
Usage for pre-training
Make a directory for saving models.
After preparing data with SMILES format (e.g. ZINC) and modifying SmilesDataset in data_utils.py, run the pretraining.py or double_pretraining.py.
python double_pretraining.py --path {dataset_path} --save_path {model_path} --adjacency True --batch 256 --epoch 20 --seq 256 --layers 6 --nhead 16 --seed 7
Fine-tuning with a pre-trained model
Download the pre-trained model using git lfs pull or here.
For using the fine-tuning service, prepare an input and output directory.
A dataset with CSV format and input.json should be located in the input directory.
Example input file for the fine-tuning service
input.json - {"split-ratio":0.8, "task":classification, "time":30}
dataset.csv - 2 columns (smiles, label)