torch_rbp_install.rst

Install PyTorch on Raspberry Pi

Platform:	Raspberry Pi 3 B+
OS:	Raspbian Stretch Lite
Torch:	PyTorch 1.1.0
Python:	Berryconda

Contents

• Prerequisities
• Berryconda
• PyTorch
• Reference

This documentation serves as a memo for installing PyTorch (torch + caffe2) on Raspberry Pi. Unlike many online tutorials that start with the native python, I choose to compile it against the Miniconda python. Since the Miniconda is only available up to Python 3.5, Berryconda is used instead for Python 3.6.

Prerequisities

• Install Raspbian Stretch Lite on SD card:

  1. Download the Raspbian image: https://www.raspberrypi.org/downloads/raspbian/
  2. Follow the oneline tutorial: https://www.raspberrypi.org/documentation/installation/installing-images/

• Increase swap to at least 2GB:

  # edit the swap file setup
  sudo nano /etc/dphys-swapfile
  # change the CONF_SWAPSIZE = 2048
  # restart swap
  sudo /etc/init.d/dphys-swapfile stop
  sudo /etc/init.d/dphys-swapfile start
  

• Install build tools

  sudo apt install git cmake ninja gfortran
  

Berryconda

1. Download Berryconda 3: https://github.com/jjhelmus/berryconda/releases/download/v2.0.0/Berryconda3-2.0.0-Linux-armv7l.sh

2. Install via the sh script

   bash Berryconda3-2.0.0-Linux-armv7l.sh
   

3. Make conda available in normal mode:

   echo ". ${HOME}/conda/etc/profile.d/conda.sh" >> ${HOME}/.bashrc
   

4. Install Python dependencies:

   conda activate
   conda install cython m4 cmake pyyaml numpy
   conda deactivate
   

PyTorch

Now we can install PyTorch from source, which contains both torch and caffe2 packages.

1. Clone the most current repository from GitHub:

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

2. Disable component not available on Raspberry Pi by setting the build environment variables:

   export NO_CUDA=1
   export USE_FBGEMM=0
   export BUILD_TEST=0
   export USE_MIOPEN=0
   export USE_DISTRIBUTED=0
   export USE_MKLDNN=0
   export USE_NNPACK=0
   export USE_QNNPACK=0
   export BUILD_CAFFE2_OPS=0
   

3. Build from source

   conda activate
   python setup.py build
   conda deactivate
   

Reference

This section shows the useful online tutorial/links I follow.