Retrosynthesis prediction using Transformers

The following instructions guide you in reproducing results from [1,2] using the powerful Transformers architecture [3].

[1] https://link.springer.com/chapter/10.1007/978-3-030-30493-5_78
[2] https://www.nature.com/articles/s41467-020-19266-y
[3] https://arxiv.org/abs/1706.03762

Requirements

This repository uses Python 3.8.11 and requires some libraries to run. Please ensure that you have rdkit pip install rdkit.
PyTorch 1.9 with cuda 11.1 was used, and the trainer writes to a tensorboard, and can be installed as follows:

pip3 install torch==1.9.1+cu111 torchvision==0.10.1+cu111 torchaudio==0.9.1 -f https://download.pytorch.org/whl/torch_stable.html
pip install tensorboard

pip install tqdm is also required.

Installation

git clone https://github.com/ArvidWartenberg/retrosynthesis.git
cd retrosynthesis/retrosynthesis_cluster_training/

Train your model

Choose any name for your run, if not it will default to a "hh/mm/ss".
The key model parameters can easily be passed as arguments. See run.py for detailed descriptions.

python3 run.py --device 0 --name my_run

Inference

Differnt weight checkpoints are made during training. These & the used model params lie in the models/chemformers folder.
The specific path to your model will be /yyyy/mm/dd/my_name. The key arguments for inference are listed below

parser.add_argument("--model", help="date/time model indicator", required=False)
parser.add_argument("--algorithm", help="search algorithm", default="greedy", required=False)
parser.add_argument("--device", help="choose device", default="0", required=False)
parser.add_argument("--weights", help="Choose checkpoint", default="best_acc", required=False)
parser.add_argument("--dataset", help="choose dataset", default="non-augmented", required=False)
parser.add_argument("--n_infer", help="num inference points", default="all", required=False)
parser.add_argument("--k", help="top k to be reported", default=5, required=False)
parser.add_argument("--beam_size", help="beam size", default=5, required=False)

Example: Run inference on entire test/val (using same xD) using greedy decoding. Model with best token ACC during training is used

python3 infer.py --model yyyy-mm-dd/my_run --weights best_acc --n_infer all --device=0 --algorithm greedy

Inferred reactants will be written to a pickle file (script prints exact path).
Also, each inference run writes to the RESULTS.txt file in root with a short summary of accuracies & which model was used.

License