KataGo icon indicating copy to clipboard operation
KataGo copied to clipboard

Looking for help implementing ONNXRuntime backend

Open isty2e opened this issue 5 years ago • 2 comments

I am currently trying to implement an ONNXRuntime backend. Main motivations are:

  • Being available to use .onnx file as the network file, which enables us to experiment with different network architectures or details without implementing everything in CUDA/OpenCL backend.
  • Graph optimization shipped in ONNXRuntime, which might be helpful for performance.
  • Easier access to some platform-specific backends such as TensorRT, DirectML, and MIGraphX.

I have done some work in my branch, but it is of course not functional yet. My main concerns are:

  • I have little to no experience with C++, so I'm pretty sure that I messed up somewhere and everything is broken here and there.
  • I have totally no idea how to configure CMakeList.txt, especially considering that find_package(onnxruntime) does not work. This comment is probably helpful for this purpose, but I'm not getting how to use this in a proper way.

For those who might be interested in, b20 and b40 weight files converted into .onnx format are available here.

isty2e avatar Oct 25 '20 05:10 isty2e