I am glad to hear from you.
1. I raised this issue about Deepchem lock-in 5 months ago on Github:
https://github.com/deepchem/deepchem/issues/708
I wanted to use the default graph in Tensorflow (which is the recommended use of Tensorflow, by the way), in order to combine a DeepChem discriminator and a non-deepchem generator (because graph generation is not possible now, it’s an open problem).
I received no answer, I moved on, and the solution is in my code.
2. Regarding char-CNN, I want to thank you again for this invitation to build a critical part of your product for free.
I have to decline it again, because as I wrote in this blog, I see no practical value in the DeepChem product at this point. Char-CNN are already available in plain Tensorflow, and that’s enough.
I already shared my little theory (a16z…) about why you didn’t have time to do this benchmark in the last 8 months. I wish I am wrong.