Pre-built LibTorch for ARM32/64 Systems

The pre-built package is https://github.com/wangkuiyi/pytorch-rpi and https://github.com/ljk53/pytorch-rpi.

NVIDIA Jetson Nano

Alternative 1

The steps to build it include:

1. Log in to an NVIDIA Jetson Nano.
2. Operating System: Ubuntu 18.04 for a recent Python.
3. Install Clang with sudo apt install -y clang.
4. Set environment variables:

   export USE_CUDA=1
   export USE_CUDNN=1
   export MAX_JOBS=4
   export USE_NNPACK=0
   export USE_QNNPACK=0
   export USE_PYTORCH_QNNPACK=0
   export BUILD_CAFFE2_OPS=0
   export USE_MKLDNN=0
   export USE_OPENCV=0

5. Run ./build_libtorch.sh to build and pack the zip file.

Alternative 2

The steps to get libtorch.so with GNU C++ compiler:

1. Visit Nvidia forum here. Download PyTorch what do you want, for example PyTorch v1.7.0 (Python 3.6 - torch-1.7.0-cp36-cp36m-linux_aarch64.whl)
2. Follow command below:

   $ sudo apt install python3-pip 
   $ sudo pip3 install torch-1.7.0-cp36-cp36m-linux_aarch64.whl
   $ pip3 show torch    # to check path of the installed torch 

3. After installation, check libtorch.so in path:

   $ cd /usr/local/lib/python3.6/dist-packages/torch/lib/      # that's consistent with output of command above
   $ ls

4. Sett CMakeLists.txt like below

   set(CMAKE_PREFIX_PATH "/usr/local/lib/python3.6/dist-packages/torch")
   find_package(Torch REQUIRED)
   if (Torch_FOUND)
       message(STATUS "Torch library found!")
       message(STATUS "    include path: ${TORCH_INCLUDE_DIRS}" \n)
   else ()
       message(FATAL_ERROR "Could not locate Torch" \n)
   endif()

Alternative 3

The steps to get libtorch.so with Clang C++ compiler:

1. Visit Qengineering repo here. Download PyTorch what do you want, for example PyTorch v1.7.0 (Python 3.6 - torch-1.7.0a0-cp36-cp36m-linux_aarch64.whl and torchvision-0.8.1a0+45f960c-cp36-cp36m-linux_aarch64.whl)

2. Follow command below:

   $ sudo apt install python3-pip 
   $ sudo pip3 install torch-1.7.0a0-cp36-cp36m-linux_aarch64.whl
   $ sudo pip3 install torchvision-0.8.1a0+45f960c-cp36-cp36m-linux_aarch64.whl
   $ pip3 show torch    # to check path of the installed torch 

3. After installation, check libtorch.so in path:

   $ cd /usr/local/lib/python3.6/dist-packages/torch/lib/      # that's consistent with output of command above
   $ ls
   $ export CC=/usr/bin/clang
   $ export CXX=/usr/bin/clang++

4. Comment script in TorchConfig.cmake in file torch/share/cmake/Torch/

   if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
      #set(TORCH_CXX_FLAGS "-D_GLIBCXX_USE_CXX11_ABI=0")
   endif()

5. Sett CMakeLists.txt like below

   set(CMAKE_CXX_FLAGS "-D_GLIBCXX_USE_CXX11_ABI=1")
   set(CMAKE_PREFIX_PATH "/usr/local/lib/python3.6/dist-packages/torch")
   find_package(Torch REQUIRED)
   if (Torch_FOUND)
       message(STATUS "Torch library found!")
       message(STATUS "    include path: ${TORCH_INCLUDE_DIRS}" \n)
   else ()
       message(FATAL_ERROR "Could not locate Torch" \n)
   endif()

NVIDIA Drive PX2

The steps to build it include:

1. Log in to an NVIDIA Drive PX2 computer.
2. Upgrade the system from Ubuntu 16.04 to Ubuntu 18.04 for a recent Python.
3. Install Clang with sudo apt install -y clang.
4. Set environment variables:

   export USE_CUDA=0
   export USE_QNNPACK=0
   export USE_PYTORCH_QNNPACK=0

5. Run ./build_libtorch.sh to build and pack the zip file.

Raspberry Pi (32-bit)

1. Install the latest Raspberry Pi OS (32-bit).
2. Run sudo apt update && sudo apt upgrade.
3. Run git clone git@github.com:ljk53/pytorch-rpi && cd pytorch-rpi.
4. Run LIBTORCH_VARIANT=armv7l-cxx11-abi-shared-without-deps ./build_libtorch.sh.

Raspberry Pi (64-bit)

1. Install the latest 64-bit Ubuntu 20.04 for Raspberry Pi.
2. Run sudo apt update && sudo apt upgrade.
3. Run git clone git@github.com:ljk53/pytorch-rpi && cd pytorch-rpi.
4. Run MAX_JOBS=2 LIBTORCH_VARIANT=aarch64-cxx11-abi-shared-without-deps ./build_libtorch.sh. Note: set the max number of concurrent build jobs to 2 to avoid running out of the RAM.

References

• Pre-built LibTorch for ARM32/64 Systems (torch and libtorch) Tested Ubuntu: https://github.com/ljk53/pytorch-rpi
• pytorch-arm-builds (torch and torchvision) Tested Fedora: https://github.com/nmilosev/pytorch-arm-builds
• Q-engineering: https://github.com/Qengineering
  • PyTorch-Raspberry-Pi-64-OS: https://github.com/Qengineering/PyTorch-Raspberry-Pi-64-OS
  • PyTorch-Jetson-Nano: https://github.com/Qengineering/PyTorch-Jetson-Nano
• Nvidia PyTorch for Jetson - version 1.11 now available: https://forums.developer.nvidia.com/t/pytorch-for-jetson-version-1-11-now-available/72048