ibm / materials.selfies-ted

selfies-ted

selfies-ted is a project for encoding SMILES (Simplified Molecular Input Line Entry System) into SELFIES (SELF-referencing Embedded Strings) and generating embeddings for molecular representations.

Model Architecture

Configuration details

Encoder and Decoder FFN dimensions: 256 Number of attention heads: 4 Number of encoder and decoder layers: 2 Total number of hidden layers: 6 Maximum position embeddings: 128 Model dimension (d_model): 256

Pretrained Models and Training Logs

We provide checkpoints of the selfies-ted model pre-trained on a dataset of molecules curated from PubChem. The pre-trained model shows competitive performance on molecular representation tasks. For model weights: "HuggingFace link".

To install and use the pre-trained model:

Download the selfies_ted_model.pkl file from the "HuggingFace link". Add the selfies-ted selfies_ted_model.pkl to the models/ directory. The directory structure should look like the following:

models/
└── selfies_ted_model.pkl

Installation

To use this project, you'll need to install the required dependencies. We recommend using a virtual environment:

python -m venv venv
source venv/bin/activate  # On Windows use `venv\Scripts\activate`

Install the required dependencies

pip install -r requirements.txt

Usage

Import

import load

Training the Model

To train the model, use the train.py script:

python train.py -f <path_to_your_data_file>

Note: The actual usage may depend on the specific implementation in load.py. Please refer to the source code for detailed functionality.

Load the model and tokenizer

load.load("path/to/checkpoint.pkl")

Encode SMILES strings

smiles_list = ["COC", "CCO"]

embeddings = load.encode(smiles_list)

Example Notebook

Example notebook of this project is selfies-ted-example.ipynb .