Paper about a model that convert a chemical diagram drawing to its computer representation
Worked with Wilson Jusuf on this project. In the project, Wilson worked on the image segmentation model to extract the chemical diagram, removing any distracting backgrounds. I worked on implementing the model that convert that chemical diagram to its SMILES representation.
We implemented a bare-bones transformer from Harvard NLP group, which decodes diagrams into their SMILES representation. We also integrated a self-critical training scheme, which uses policy gradient to optimize the model to minimize the levehstein distance between target and prediction. All in all, we managed to get a SOTA performance of ~75+% accurancy (based on levehstein distance).
Check the paper out here