Support for chainer?

[cmpute]

🚀 Feature

Support chainer as backend

Motivation

I'm a chainer user and it would be great if chainer is supported as backend. Usage of Chainer is quite similar to Pytorch, so I think it would not be that hard to port it.

Additional context

Chainer Repo
Cupy Repo

[jermainewang]

Hi, thank you for the suggestion. We do love to have more backends so more users can use DGL. One question, does Chainer support DLPack protocol? It allows framework-specific tensors to expose their raw memory so DGL can perform computation on it in C.

[cmpute]

Sorry, I'm not sure about this. Chainer is built based on numpy and cupy. I think this could be answered by chainer developers.

@kmaehashi Sorry to quoting you. I hope this won't makes you too many troubles. Could you kindly help with this issue? Thanks!

[BarclayII]

Seems that Chainer tensors are just numpy and cupy arrays, both of which support DLPack (for cupy, see https://github.com/cupy/cupy/blob/master/docs/source/reference/interoperability.rst#dlpack).

@cmpute Could you try implementing the Chainer backend? Since DGL is platform-agnostic, all you need to do to support a new backend is to create a directory python/dgl/backend/chainer and fill in the implementations. You could refer to python/dgl/backend/pytorch/tensor.py to see what are the needed functions and provide the Chainer equivalents accordingly.

For unit-testing the new backend, creating a directory tests/backend/chainer and filling in the implementations should also do most (if not all) of the job.

[cmpute]

I can have a try if time permitted. However I'm not working on graph learning recently, the progress may not be much

[BarclayII]

I can have a try if time permitted. However I'm not working on graph learning recently, the progress may not be much

No problem. I am currently working on Chainer support now. It should take a few days.

[yzh119]

@BarclayII for pooling layer we need more backend functions, please check https://github.com/dmlc/dgl/pull/669/files, thanks