Attendees: Bharath, Peter, Seyone
Summary: Bharath spent most of the last week working on the DeepChem upgrade to TensorFlow 2.X (see PR). This PR required a number of changes for TensorFlow 2.X support. Notably, the graph convolutions had to be rewritten in TensorFlow’s imperative style (in which you subclass
tf.keras.Model directly). The DAG graph convolutions had to be tweaked as well to run to completion in TensorFlow 2.X. This PR hasn’t been merged in yet and needs review. Peter said he’d do a review pass after the call.
Peter’s been working on a pass through the book examples to upgrade for DeepChem 2.4. Most things worked, but a number of small changes and fixes, to both the book code and DeepChem itself were required. These changes are being done on a separate branch, but will be put up in a PR once ready. Peter’s been fixing some of the small DeepChem issues he’s been finding along the way (PR, PR).
Seyone’s been continuing work on exploring the use of huggingface-style pretrained NLP models for making predictions on smiles strings. He’s got a model training on Colab, but the model itself will take some time to train to completion. Once the model is functional, Seyone said he’d be interested in putting together a tutorial explaining the technique.