LaMolT5

Training

• You need to install all the dependencies from the requirements.txt.
• You need to login to the huggingface-cli with a write token and that you have set up the correct paths for the train/finetune.py file either change the default values or give them as arguments.
• Use the bash script: bash training/train.sh from the root directory to train and save intermediate steps on huggingface.
• To load the tensorboard use the following commands:
  • tensorboard --logdir <LOG_DIR> --port 6006 --bind_all
  • ssh -N -f -L localhost:16006:localhost:6006 <USER>@<SERVER>

Models

The models for this project can be found at huggingface here (update path to new huggingface repo). For now there are a checkpoint for each 100 epochs.

Predict

• You can generate predictions with a trained model using the prediction/predict.py script.

Evaluation

• You can evaluate trained models with the two scripts evaluation/evaluate_caption2smiles.py and evaluation/evaluate_smiles2caption based on predictions generated on the previous step.

Results

The results from the evaluation step will be saved as a file under the evalauation folder.