Build instructions for PyTorch on Raspberry Pi 3 B+ running the 32-bit Raspbian Buster using a chroot on Ubuntu.
The include .whl file was built on Ubuntu 18.04 using the below approach, with Python 3.7.3 and PyTorch v1.3.0.
Rather than build PyTorch from scratch on the device itself, we can instead build on a more powerful machine running Linux using a combination of chroot and QEMU. A chroot allows us to setup an isolated environment by changing the apparent root directory, while QEMU allows us to emulate the target ARM processor on a non-ARM machine. These instructions were adapted from here, here, and here.
First, we will create a basic Debian system using debootstrap:
sudo apt-get install debootstrap
sudo mkdir -p <path-to-chroot>
sudo debootstrap --foreign --variant=buildd --arch=armhf buster <path-to-chroot>
where is the path to our new file system. Now let's break down the arguments here a bit. The --foreign argument indicates that we should only perform the initial unpacking phase of bootstrapping. The second stage installs build-essential packages and must be done using the QEMU, since our target architecture is different than the host (e.g. ARM vs AMD64). The --variant=buildd installs the build-essential packages during the second stage and --arch=armhf sets the target architecture to armhf, which is for 32-bit architectures that support hardware floating point units (e.g., arm7l). Note that while the CPU on the Pi is technically 64 bits, Raspbian is a 32-bit OS. Finally, buster is the version of Debian we are installing.
Next, we need to install QEMU:
sudo apt-get install qemu-user-static qemu-system-arm virt-manager
and copy the QEMU static binary into the chroot file system:
sudo cp /usr/bin/qemu-arm-static <path-to-chroot>/usr/bin/
We also need to copy over the DNS information as the second stage of debootstrap requires network access:
sudo cp /etc/resolv.conf <path-to-chroot>/etc/
Now we enter the chroot and complete the second stage of debootstrap:
sudo chroot <path-to-chroot>
/debootstrap/debootstrap --second-stage
Again, this second stage should install all of the build-essential packages for Debian Buster.
Since we have created a new file system, we need to install all our dependencies to build PyTorch, regardless if they have been installed on the original file system:
apt install libopenblas-dev libblas-dev m4 cmake cython python3-dev python3-yaml python3-setuptools python3-wheel python3-pillow python3-numpy git
Make sure that you install the version of Python you intend to use with PyTorch on the Pi! Otherwise, you will run into
ModuleNotFoundError: No module named 'torch._C'
errors when you try to import torch. Look at this GitHub issue for more information regarding this issue.
We now need to clone the PyTorch repository and all of its submodules:
git clone https://github.com/pytorch/pytorch --recursive && cd pytorch
git checkout v<pytorch-version>
git submodule update --init --recursive
where is the version of PyTorch we wish to build. Next, we set the following environment variables for the build:
export NO_CUDA=1
export NO_DISTRIBUTED=1
export NO_MKLDNN=1
export BUILD_TEST=0
export MAX_JOBS=8
These are fairly self-explanatory, preventing the build script from building with CUDA, distributed support, use of the MKL-DNN library (which does not support ARM processors), and build tests. The last environment variable, MAX_JOBS, indicates the number of processors can be used to compile.
To build PyTorch, we simply run
python3 setup.py bdist_wheel
The entire build process should only take few hours (compared to the >12 hour time it would take on the Pi). The wheel will then be in /pytorch/dist/ folder, which you can install by copying to the Pi and installing with
python -m pip install <path-to-wheel>