A guide to install and use the PyTorch C++ API with Anaconda.md

PyTorch C++ API Ubuntu Installation Guide

The best way to get a clean installation of PyTorch, is to install the pre-compiled binaries from the Anaconda distribution. Therefore, we need to setup Anaconda first.

Step 1: Install Anaconda

• Go to the download section and download your desired Anaconda version for Linux

• Run the downloaded shell script and follow the install instruction, do

cd ~/Downloads
sh Anaconda3-2021.11-Linux-x86_64.sh # actual name depends on your download

Step 2: Setup an Environment

• To keep things clean, now setup an environment for Anaconda, do

conda create --name py39_torch python=3.9 # you can choose the name of the environment and the python version

• Activate the environment

conda activate py39_torch

Step 3: Install PyTorch

There are two different ways on how to proceed.

Install pre-compiled Binaries

• Go to pytorch.org and choose your desired settings like so

• Run the proposed command, and make sure your environment is activated

conda install pytorch-cpu torchvision-cpu -c pytorch

Install from Source

Sometimes you want to compile from source, due to specific dependencies for example. You can look up what to do here. The main steps are

• Get dependencies

conda install numpy ninja pyyaml mkl mkl-include setuptools cmake cffi typing

• In case you want to use CUDA on a GPU, otherwise you can skip

# Add LAPACK support for the GPU if needed
conda install -c pytorch magma-cuda90 # or [magma-cuda92 | magma-cuda100 ] depending on your cuda version

• Clone PyTorch from GitHub

git clone --recursive https://github.com/pytorch/pytorch
cd pytorch

• Install PyTorch, make sure your environment is activated

export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
python setup.py install

Step 4: Link against libtorch.so

Great you came this far, now lets see if everything worked well. Here is a minimal example, on how to link against the C++ API of Pytorch.

• Build the example, the path depends on your installation. If you followed the previous steps, it should be

mkdir build
cd build
cmake -DCMAKE_PREFIX_PATH="${HOME}/anaconda3/envs/py39_torch/lib/python3.9/site-packages/torch" ..
make

• Run the example

./main

Thats it! Hope it helped you.

CMakeLists.txt

cmake_minimum_required(VERSION 3.10 FATAL_ERROR)
project(example)

find_package(Torch REQUIRED)

add_executable(main main.cpp)
target_link_libraries(main ${TORCH_LIBRARIES})
set_property(TARGET main PROPERTY CXX_STANDARD 14)

main.cpp

#include <torch/torch.h>

int main()
{
    torch::Tensor tensor = torch::zeros({2, 2});
    std::cout << tensor << std::endl;

    return 0;
}

where to do mkdir build?

see Step 4 @slulura, folder structure should look like this:

/some_folder
  CMakeLists.txt
  main.cpp
  ./build

Nice instructions! Thank you!

Great job. please add the following code as a dependency of PyTorch
pip install typing-extensions