This work was done as a class project for XCS224N
This repo uses Pytorch Lightning and HuggingFace Transformer to fine-tune T5 for Chemical Reaction Prediciton. See https://github.com/pschwllr/MolecularTransformer for training script and model training a Transformer from scratch using OpenNMT and https://github.com/blender-nlp/MolT5 for details on pre-training T5.
-
fine_tune_pl.py trains molT5
-
test_beam_search.py evaluates model checkpoint using a beam search decoder
-
test_top_p.py evaluates the model checkpoint using a top-p decoder.
-
pre_training_w_mlm contains a script for pre-training T5 on a new corpus, for example if one desires to train only on SMILES, rather than a mixed corpus which molT5 used