NewMoses

Dependency

torch <= 1.13

Environment setup

create and activate conda environment named moses with python=3.8

conda create -n moses python=3.8 -y \
conda activate moses \
pip install -r requirements.txt \
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch

if you use MacBook (or DO NOT USE cuda), use this code

conda create -n moses python=3.8 -y \
conda activate moses \
pip install -r requirements.txt \
pip install torch==1.12.0+cpu torchvision==0.13.0+cpu torchaudio==0.12.0 --extra-index-url https://download.pytorch.org/whl/cpu

Running the benchmark data and model

• Only train one model on each process b/c of wandb tracking
• run the below code for training the model
• You can use benchmark model: aae, char_rnn, latentgan, organ and vae
• You can use benchmark dataset: QM9, ZINC, ZINC250K
• Before running the code, make sure to unzip the train.zip and test_stats.zip files of ZINC, which were compressed due to upload limitations
• If you use cuda, add '--device cuda:{device_idx}', else --device cpu
• If you want to train model with selfies format, you add the '--use_selfies 1' when you run the scripts/run.py
  • if you don't add it, the model is trained by smiles format automatically
  • (!!Caution!!) if you use --use_selfies with any kind of format (ex: 0,1, ..., True, False...), the model is trained by selfies

Example:

python scripts/run.py --device cuda:0 —model vae --use_selfies 1 --n_batch 2048

For use the wandb, you need to setup below file:

python scripts/run.py --device cuda:0 —model vae --use_selfies 1 --n_batch 2048 --wandb_entity {wandb_id} --wandb_project {project_name} --nowandb 0

• How to run VAE with property predictor
  • When you run the vae with property predictor, you can choose certain properties or all.

Example:

python scripts/run.py --device cuda:0 —model vae_property --reg_prop_tasks logP qed --n_batch 2048

Adding the Dataset

If you train model using your model, add the splited dataset named train.csv, test.csv in moses > dataset > data > {datasetname} > files For example, we have already make the directory for ZINC and QM9 dataset

Sampling (Generate Sample using trained model)

• n_samples: how many samples do you want to generate
• model_save_time: the time of the model folder
• load_epoch: what epoch do you want to use

python scripts/run_samples.py --model_save_time 20240515_021753 --model vae --data ZINC --load_epoch 080 --n_samples 1000

Evaluation

• device: choose cpu or cuda:{index}
• n_jobs: How many workers for evaluating the models

python scripts/run_eval.py --data ZINC --model vae --model_save_time 20240515_021753 --device cpu --n_jobs 8

Reference code

We re-generate the code from https://github.com/molecularsets/moses for our project.