ibm-research/materials.selfies-ted2m

selfies-ted2m

selfies-ted is an transformer based encoder decoder model for molecular representations using SELFIES. This is a 2.2M parameter version of the model. For the full-sized version and more information on architecture, see selfies-ted.

This version also includes a projection layer to convert the last hidden state of the BART model (256-dimensional vector per token) to a single 128-dimension vector for the whole SELFIES sequence.