Attendees: Bharath, Peter, Vignesh, Seyone, Sean, Dilip, Raman
Summary: We had a few new attendees on this call today so we started with introductions. Sean and Dilip are PhD students at CMU who’ve been working on porting DeepChem to Julia. Raman is a data scientist at Maze Therapeutics who’s been experimenting with DeepChem.
Bharath’s still continuing work on the examples PR. This PR is starting to stabilize, so Bharath is starting to slice out small segments of this PR to merge into master. He merged in a couple small PRs pulled out of the bigger examples PR, one for adding some geometry utils and another for adding hash function utilities. He’s planning to continue to pull out small PRs to merge into master over the next week or two.
Vignesh is working on his transfer learning WIP PR. He’s adding a framework for generalizing chemnet transfer learning to other types of models. Bharath asked if it would be possible to add an example for the new transfer learning framework and Vignesh said he’s planning to add an IPython notebook walking through use of the new framework. Raman mentioned that he’d been able to get the transfer learning code working on his end for a project he’s working with.
Seyone’s PR for transfer learning has been merged in. He’s now working on benchmarking a few other models such as Graph attention networks. He’s been looking at dgl in particular as a way to potentially speed up some of his models. Bharath mentioned that he came across the spektral library which is trying to build a version of DGL/PyTorch-Geometric on top of Keras. This might be a useful resource for improving DeepChem’s graph convolution support. Seyone mentioned it might be useful to have graph attention networks implemented in DeepChem.
Peter has been working to get Windows support added to DeepChem. He was able to get the basic test suite running locally on Windows without too much effort, but when he started looking at the full test suite (including tests marked
slow) he started to find a number of failing tests. Fixing these tests would conflict with Bharath’s example PR, so he’s blocked there until the large PR is merged in. Bharath said he hoped that large chunks of the examples PR should be in by next week. Peter’s now looking into getting Travis CI support for Windows up and running and is looking through some of the docs there.
Sean and Dilip have been working to port DeepChem over to Julia. In particular, they’ve been working to port the Weave models over. Sean asked whether there were any plans to support Julia in DeepChem. Bharath said that not at present, but he’d be glad to help a bit personally. Sean pointed out that Julia has a different model for machine learning than TensorFlow/PyTorch which means that some of the library API and design might have to be different to follow Julia best practices.
As a quick reminder to anyone reading along, the DeepChem developer calls are open to the public! If you’re interested in attending, please send an email to X.Y@gmail.com, where X=bharath, Y=ramsundar.