libtorch C++ jni

[nazarblch]

The main changes:

1. Tensor methods (https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/core/Tensor.h) and torch Functions.h are partially wrapped into Scala API (in src/main/scala/torch_scala/api).
2. Tensor and Variable classes have been made generic (Tensor[T, TT] and Variable[T, TT]) with templates T corresponded to data type and TT corresponded to device type (example Tensor[Float, CUDA] or Tensor[Int, CPU])
3. Tensor has implicit wrappers in src/main/scala/torch_scala/api/aten/functions, including MathBackward.scala for backward methods.
4. Sbt plugin for JNI code generation JniGeneratorPlugin and sbt plugin for cmake so lib building JniBuildPlugin in project directory.
5. The old scorch code is partially adapted and completion are required.

[koen-dejonghe]

Thank you very much for your code. I will definitely have a look at it. Keep you posted.

[koen-dejonghe]

Is it possible to build without GPU support?

[nazarblch]

Yes, you can build it with https://download.pytorch.org/libtorch/cpu/libtorch-shared-with-deps-latest.zip Pull from the repo first and don't forget to export JAVA_HOME I would propose to discuss this project in skype, what do you think?

Nazar

чт, 4 апр. 2019 г. в 13:21, Koen Dejonghe [email protected]:

Is it possible to build without GPU support?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/botkop/scorch/pull/2#issuecomment-479839922, or mute the thread https://github.com/notifications/unsubscribe-auth/ABslRBWD1Y0CfguviXLxxewwWin_XOxiks5vddISgaJpZM4cI79k .

[koen-dejonghe]

I created a gitter room. The button is on the readme. We can discuss there.