It is recommmended using conda for creating a virtual environment. If you want to (re)train the models, your system needs to have CUDA dependencies, please use the environment.yaml file for the installation.
conda env create -f environment.yaml
conda activate molbind
The simulated spectra data have been compiled from IBM's Multimodal Spectroscopic Dataset.
(WIP 🏗️) Run molbind-get-datasets from the command line to download the data.
Your environment file should look like this:
WANDB_PROJECT="<your-wandb-project-name>"
WANDB_ENTITY="<your-wandb-account-name>"
TOKENIZERS_PARALLELISM=False
After you have defined your system variables in .env, it is read into the script as following:
load_dotenv("path/to/.env")The experiment configs can be found at config. For example, to run the train.py
python train.py 'experiment="train/ir_simulated_large_dataset"'The training scripts outputs the checkpoints at experiments/checkpoints/<run-code-name>/<checkpoint-file-name>.ckpt
To find all three checkpoints used in this work, please access the supplementary information on Zenodo.
To run the metrics on these experiments:
python retrieval.py 'experiment="metrics/ir_simulated_large_dataset"'For the training script 4 NVIDIA A100-40GB GPUs have been used. For the retrieval script 1 NVIDIA A100-40GB GPUs has been used.
This work was funded by the Carl-Zeiss Foundation. In addition, this work was partly funded by the SOL-AI project funded as part of the Helmholtz Foundation Model Initiative of the Helmholtz Association. Moreover, this work was supported by Helmholtz AI computing resources (HAICORE) of the Helmholtz Association’s Initiative and Networking Fund through Helmholtz AI.