Attendees: Bharath, Peter, Vignesh
Summary: Bharath gave a quick update on the situation with the forums. The forums crashed while he was working to update them to the latest version. The forums were previously running on an AWS EC2 instance, but he’s planning to migrate to a digital ocean droplet to keep maintenance simpler. He hopes to have the forums up and restored from the backup by this weekend.
Bharath’s spent most of his development time over the last week overhauling the DeepChem tutorials and put up a few PRs (see here and here). We also had a community PR that added in missing metrics merged in. He also merged in a number of old PRs that improved the tutorials.
Bharath also mentioned some of his early plans to start an overhaul of MoleculeNet. The original benchmarks are a couple years out of date and the community has introduced a number of new models and benchmarks. To facilitate this effort, the MoleculeNet code will be split off into a new repo which will host the benchmarks. Karl mentioned that he’d be interested in helping with this effort and there are a few other community members who said they’d be interested too.
Peter’s near completion of the TensorFlow 2.X update. There are only a couple of major sticking points. The first is that the DAG model doesn’t work in eager model. The second is that the GraphConv model doesn’t work either. The DAG model isn’t widely used by the community, so we’re tentatively planning to drop support for it. The GraphConv model is widely used though, so we need to get it up and running before we release DeepChem 2.4. Bharath asked if we could merge in the TensorFlow 2.X PR as-is and address the remaining issues in subsequent PRs. Peter said we’d be good to do this once the last changes with removing TensorGraph are completed. Once merged, we can make sure all the tutorials run on TensorFlow 2.X
Vignesh is continuing work on his thesis. He’s going to be mostly preoccupied with that effort for the next few weeks, but thinks he’ll have more time to get involved with development efforts in April once he’s got a little more time.
Bharath mentioned that we haven’t had a strong GSoC candidate come up yet, so that’s something to keep an eye out for. He’s also been taking some looks at Jax to figure out whether we can start using Jax to implement some of the next-gen models out there in the literature. Since Jax and TensorFlow share infrastructure, it might be a better next-gen target for DeepChem than PyTorch.