This is a collection of wheels for PyTorch and Torchvision build on top of PYNQ images. As opposed to most other PyTorch wheels for ARM, ours provide support for torch.distributed
All wheels are built in a chroot environment on Ubuntu, using the PYNQ-provided Qemu static binary for aarch64. This flow has been tested on Ubuntu 16.04 and 18.04. Install host-side requirements with:
sudo apt-get install chroot unzip kpartx
Download and unzip PYNQ SD-card image and clone the PYNQ repo. The following tutorial assumes the image file is called pynq.img and the PYNQ repo was cloned in the same folder where the image file resides.
First we need to mount the PYNQ image to a folder:
mkdir img
./PYNQ/sdbuild/scripts/mount_image.sh pynq.img img/
You will be asked for your sudo password. After the script completes, img/ will be populated with the contents of the PYNQ SDcard image. We now need to mount a few host-side directories to the image: a workspace folder for more space, and /proc, /dev, /sys for various functionality we'll need during the build.
mkdir workspace
sudo mount --bind `pwd`/workspace `pwd`/img/workspace
sudo mount -o bind /proc img/proc
sudo mount -o bind /dev img/dev
sudo mount -o bind /sys img/sys
To enable networking in the chroot, we need to copy the config to the image mount (make sure it's copied and not symlinked!).
sudo cat /etc/resolv.conf > img/etc/resolv.conf
Next we need to register qemu as the default interpreter for ARM (aarch64) binaries. You'll need to be root to do this.
echo ':qemu-aarch64:M::\x7fELF\x02\x01\x01\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\xb7\x00:\xff\xff\xff\xff\xff\xff\xff\x00\xff\xff\xff\xff\xff\xff\xff\xff\xfe\xff\xff\xff:/usr/bin/qemu-aarch64-static:' > /proc/sys/fs/binfmt_misc/register
Set-up completed! Now enter the chroot to the PYNQ image.
sudo chroot `pwd`/img /bin/bash
First we need to install some prerequisites in the PYNQ image
apt-get -y update && apt-get -y install protobuf-compiler libopencv-dev libopenmpi-dev
pip3 install setuptools
Clone PyTorch and switch to the desired version.
cd /workspace
git clone --recursive https://github.com/pytorch/pytorch.git -b v1.1.0
Set up your environment to enable/disable components as desired. Note that PyTorch 1.5 fails during QNNPACK build due to a bug.
export USE_CUDA=OFF
export USE_MKLDNN=OFF
export USE_NNPACK=OFF
export USE_QNNPACK=OFF
export USE_PYTHON_QNNPACK=OFF
export USE_OPENCV=1
export BUILD_CAFFE2_MOBILE=OFF
export BUILD_CAFFE2_OPS=OFF
export BUILD_CUSTOM_PROTOBUF=OFF
export BUILD_TEST=0
Now build the binary wheel:
cd pytorch
python3 setup.py bdist_wheel
Build times will vary depending on the configuration you set and the number of cores available. The resulting wheel is in /workspace/pytorch/dist/. Install it with:
pip3 install /workspace/pytorch/dist/*.whl
Once PyTorch is installed, we can build Torchvision. Mind the PyTorch version installed when choosing a Torchvision version. The only other dependency is Pillow. Note that Pillow 7 is only supported in Torchvision 0.5 and above.
pip3 install "Pillow<7"
Clone the repo:
git clone https://github.com/pytorch/vision.git -b 0.4.0
And build the wheel:
cd vision
python3 setup.py bdist_wheel
Exit the chroot with exit, then unmount every host-side folder mounted to the image:
umount img/workspace
umount img/proc
umount img/dev
umount img/sys
Unmount the image itself using the PYNQ script:
./PYNQ/sdbuild/scripts/unmount_image.sh img/ pynq.img
The script will unmount the image file and loop partitions created for it. The workspace/ folder contains your wheels.
The build methodology is based on info gathered from tutorials here and here. Alternative PYNQ-compatible PyTorch wheel available here (built with NO_DISTRIBUTED=1). Additional PyTorch and Torchvision wheels are available here but are built against Python 3.7 and therefore incompatible with PYNQ.