This step assumes a pretrained AtomLenz model (on synthetic data) and will adapt to a target domain where only SMILES labels are available. Pretrained AtomLenz can be obtained here: pretrained atomlenz

trained models using below steps can be found here:

• Atom model: https://huggingface.co/spaces/moldenhof/atomlenz/resolve/main/models/atoms_model/real.ckpt
• Bond model: https://huggingface.co/spaces/moldenhof/atomlenz/resolve/main/models/bonds_model/real.ckpt
• Charge model: https://huggingface.co/spaces/moldenhof/atomlenz/resolve/main/models/charges_model/real.ckpt
• Stereo center model: https://huggingface.co/spaces/moldenhof/atomlenz/resolve/main/models/stereos_model/real.ckpt

Prepare dataset

AtomLenz can be trained on images with only SMILES labels. One such a dataset is available here: https://dx.doi.org/10.6084/m9.figshare.24599412 . This is a dataset with hand drawn images labeled with the corresponding SMILES.

Once downloaded, this dataset should be prepared for use by ProbKT. This can be done with parse smiles script.

python parse_smiles.py --datasetfolder hand_drawn_train/train/ --outputfolder atoms_dataset/

This should be repeated for bonds, charges and stereocenters:

python parse_smiles.py --datasetfolder hand_drawn_train/train/ --outputfolder bonds_dataset/ --parse_bonds 1

python parse_smiles.py --datasetfolder hand_drawn_train/train/ --outputfolder charges_dataset/ --parse_charges 1

python parse_smiles.py --datasetfolder hand_drawn_train/train/ --outputfolder stereos_dataset/ --parse_stereos 1

Train (with ProbKT) AtomLenz on each prepared dataset

This step assumes a pretrained AtomLenz model and some ProbKT prepared dataset (previous step). For each prepared dataset from target domain now the pretrained AtomLenz model can be finetuned.

python robust_detection/train/train_fine_tune.py --og_data_path path_to_synthetic_dataset --target_data_path atoms_dataset/ --fold 0 --experiment_path locations_of_pretrained_atom_model

The locations of the datasets and models need to be adapted to each type (atoms/bonds/charges/stereos). For the path_to_synthetic_dataset we recommend to create a random sample so that the number of samples is in the same order as the number of samples from the dataset in target domain.

Train (with self correcting mechanism) AtomLenz on target domain

python self_label.py --experiment_path_bonds ../ProbKT/robust_detection/train/logger/path_to_bonds_network  --experiment_path_charges ../ProbKT/robust_detection/train/logger/path_to_charge_network --experiment_path_stereo ../ProbKT/robust_detection/train/logger/path_to_stereos_network --experiment_path_atoms ../ProbKT/robust_detection/train/logger/path_to_atom_network  --data_path ../ProbKT/generate_data/path_to_dataset_target_domain --score_thresh 0.65