Odin bindings for the C++ api of PyTorch. The goal is to provide a thin wrapper around the C++ PyTorch api (a.k.a. libtorch). Staying as close as possible to the original.
Downloads required:
LibTorch (C++): Download the cxx11 ABI version (Linux) or standard version (Windows) from pytorch.org. Unzip it to libtorch in the root of this project.
Use odin run etc to download libtorch C api headers from tch-rs to ffi/ and
automagically compile the shared lib to atg/
Alternatively, you can also run the command to compile libtorch abi manually:
macos:
clang++ -std=c++17 -dynamiclib \
-I libtorch/include \
-I libtorch/include/torch/csrc/api/include \
-L libtorch/lib \
-ltorch -ltorch_cpu -lc10 \
-Wno-deprecated-declarations \
-o atg/torch_wrapper.dylib \
ffi/torch_api.h ffi/torch_api_gen.h ffi/stubs.cpp \
-Wl,-rpath,$(pwd)/libtorch/liblinux:
clang++ -std=c++17 -shared -fPIC \
-I libtorch/include \
-I libtorch/include/torch/csrc/api/include \
-L libtorch/lib \
-ltorch -ltorch_cpu -lc10 \
-Wno-deprecated-declarations \
-o atg/torch_wrapper.so \
ffi/torch_api.h ffi/torch_api_gen.h ffi/stubs.cpp \
-Wl,-rpath,'$ORIGIN/libtorch/lib'windows:
TODO should not be all to hard just don't have windows personally.
Contribution welcome!
Check the demo folder:
$ odin run demo/basic
Lib Torch via Odin
Creating tensor of shape [2, 3] with data: [1, 2, 3, 4, 5, 6]
Tensor created successfully!
[LibTorch Output]:
1 2 3
4 5 6
[ CPUFloatType{2,3} ]
[Filling tensor with 99's]
99 99 99
99 99 99
[ CPUFloatType{2,3} ]