This package provides tools for training graph convolutional models for molecules using transfer learning techniques.
Use conda to install requirements presented in nb_deepchem_gpu.yml
conda env create -f nb_deepchem_gpu.ymlTraining process carried out with run_train.py through TransferTrainer object. All training features are controlled by several flags in trainer object. Transfer training carried out by adding path to folder of pretrained model. Input data samples are presented in Datasets
Flags for transfer training
restore_folder=source_fold_folder, layers_to_freeze=["graph_conv"]Flags for hyperopt optimization:
hyperopt=True, hyperopt_evals=100
There are several mandatory input for TransferTrainer class init
path_to_sdf - path to sdf file for molecules or to folder with cif for materials
valuename - name of target property in sdf file, optional for materials
source_fold_folder - path to pretrained model to use it as donor for transfer, only models trained with
TransferTrainer are allowed
output_folder - path to folder where TransferTrainer will create dir with all training results
mode - classification or regression based on target property, multiclass not implemented
If you use this code in your research, please cite this paper Size Doesn’t Matter: Predicting Physico- or Biochemical Properties Based on Dozens of Molecules