Attendees: Bharath, Peter
Summary: Bharath started off the meeting. Unfortunately his laptop suddenly stopped working that week so he wasn’t able to work on debugging the graph convs as planned. He did have a number of interesting conversations about use cases of graph convolutional methods in materials science which might be relevant to ongoing DeepChem 3 work.
Peter has been working on the continuing efforts to update to TensorFlow 2.X. The A3C RL codebase has been converted to A2C, and he’s now working on figuring out the graph convolutional errors. (The PPO code hasn’t been converted yet.) At present, he thinks that there might be a bug in TensorFlow that explains the issues he’s seeing. He’s still working on digging into what the cause for the changes in performance might be. Check out the discussion on Github here
A general point that came up in the discussion is that eager mode (default in TF 2.X) is considerably slower than graph mode. This will make all our models slower in TF 2.X by default. We also had a brief discussion of the new fast.ai API and the new paper they put out about layered API design for deep learning libraries. We’re planning on looking into it more as we continue DeepChem 3 work.
Peter is planning to continue debugging. Once Bharath is set up on his new machine, he’ll start pitching in as planned.