Collaborating with TVM and NNVM from DMLC

Hey there,

As someone who has been in the Python data-science community for a long time, I love the energy around Machine Learning frameworks, but I’m also alarmed by the amount of fracturing, siloing, and lack of coordination that is taking place. As I’ve said in several talks at this point, the division in Python machine-learning that exists right now makes my fears over a Numeric and Numarray split that led me to sacrifice my tenure track to develop NumPy seem rather “quaint”.

How can we foster cooperation and collaboration. I’ve posted another message about Chainer specifically: https://chainer.org/

This message is to raise awareness to TVM and NNVM: http://tvmlang.org/2017/10/06/nnvm-compiler-announcement.html. Is ONNX the only way for PyTorch and these tools to be interoperable?

In my studies and research lately, I’ve found that TVM and NNVM to be fascinating projects. I’d like to know what this community thinks of that work and if anyone sees potential for integrating those projects more clearly with PyTorch.

I have some potential funding (and intern talent) I can use to help with these efforts.

Thanks for your feedback.

-Travis

On this topic, we are closely integrating with TVM now, with some engineering behind it. Please see https://github.com/pytorch/tvm

We are less inclined about NNVM – TVM provides Relay, which is a high-level IR.