Attendees: Bharath, Peter
Summary: Bharath started by summarizing some more conversations he’d had with DeepChem users. One of them mentioned that while they made extensive use of DeepChem for basic models, they found it hard to extend and modify core models. There was also some feedback that having PyTorch Geometric support would make a difference here.
Besides conversations, Bharath mentioned that he’d gotten DeepChem test suite running on his new machine and was starting work on figuring out what the causes of failing tests for the TensorFlow 2.X conversion were. Peter suggested looking at the graph convolutional test failures and reporting back progress.
Peter is working on converting the reinforcement learning code to TensorFlow 2.X. There is unfortunately a complicated issue arising from the invocation of a Keras model returning a Tensor instead of an EagerTensor object. There has been some discussion on the TensorFlow github of this issue, but no fix at present. Peter is going to continue looking for a workaround, which might involve changing the RL code from A3C to A2C or PPO.
Bharath and Peter are both primarily planning to continue debugging and conversion efforts this week and hope to have more progress next week.