Should work with Python 3.6, fast.ai 1.0 on armv6 (tested) or armv7 (not tested).

NOTE: I didn't manage to build a working wheel for spacy - however it's only required for the fastai text package - vision works fine without it.

These have been tested on a Raspberry Pi Zero W, running Raspbian Buster and Berryconda 3.2.

PyTorch was built with the following options:

NO_CUDA=1
NO_DISTRIBUTED=1
NO_MKLDNN=1 

Make sure you get Python 3.6 on the raspberry first, I used Berryconda: https://github.com/jjhelmus/berryconda

Then install the wheels:

pip install numpy-1.17.2-cp36-cp36m-linux_armv6l.whl
pip install numexpr-2.7.0-cp36-cp36m-linux_armv6l.whl
pip install Bottleneck-1.2.1-cp36-cp36m-linux_armv6l.whl
pip install Pillow-6.1.0-cp36-cp36m-linux_armv6l.whl
pip install torch-1.2.0a0+8554416-cp36-cp36m-linux_armv6l.whl
pip install torchvision-0.4.0a0+d31eafa-cp36-cp36m-linux_armv6l.whl

And finally pip install fastai --no-deps

Inference time with a resnet34 model is around 52s per image on a Raspberry Pi zero - not exactly fast, but enough for my use-case :)

Thanks to @nmilosev for the very useful guide of getting cross-compilation of PyTorch working on a desktop: https://nmilosev.svbtle.com/compling-arm-stuff-without-an-arm-board-build-pytorch-for-the-raspberry-pi

(I used Ubuntu and debootstrap with qemu)