I used RaspiOS Lite as base and did all compiling on a Raspberry Pi 8GB (running also works with less memory, but for compiling PyTorch you would need to reduce parallelism and this would increase the time it takes).
Note: I always install wheels system-wide (so sudo pip3
install <wheel>).
Install additional packages:
sudo apt-get install python3-numpy python3-wheel python3-setuptools
python3-future python3-yaml python3-six python3-requests python3-pip
python3-pillow cmake jed libnuma-dev ninja-build ccache python3-typing-extensions
screen git python3-protobuf python3-numpy python3-scipy
python3-matplotlib llvm-dev python3-psutil cimg-dev jupyter-notebook
ipython3 build-essential
PyTorch: The wheel below has been compiled using my usual instructions from the master branch (git rev 963f7629b591dc9750476faf1513bc7f1fb4d6de). It contains quantization fixes over PyTorch 1.7, so we want this. I also added a diff (see below) to enable NNPACK for batch size one. This is has been proposed in PyTorch PR 26711.
TorchVision: The wheel below has been compiled from git tag v0.8.1.
Audio: apt-get install libsox-dev python3-sklearn python3-numba
python3-cffi
Compile torchaudio from master or use whl below (compiled from git rev
cf11427643edc4bf7681c0e425f0409eb21a7d28).
sudo pip3 install librosa*whl
Librosa has been rebuilt to not depend on a specific version of numba
(from git master rev a53fa56bdb6695a994008d5b6ccd0100870a6036) plus
the patch below.
Recompile libarm-compute20 and libarm-compute-dev on Buster from the Debian Bullseye sources after installing the build-dependencies or use the debs below.
ONNXRuntime install whl below. It has been built from a checkout of
the v1.6.0 tag with the patch for acl 20.08 below.
The command was ./build.sh --config RelWithDebInfo
--build_shared_lib --parallel --enable_pybind --build_wheel --use_acl
ACL_2002 --build_wheel --skip_tests
TVM has been built from the main branch
(rev b8ac8d94dec93cdaaec436f9105ea9b2eef752d0) after changing the
config to set USE_ARM_COMPUTE_LIB and
USE_ARM_COMPUTE_LIB_GRAPH_RUNTIME to on.
TVM needs XGBoost at runtime, the whl was compiled from
the master branch (rev 60cfd14349a3ed4593de276525fba068adcaea53).
pip will install additional packages, I ended up with appdirs packaging SoundFile joblib pooch attrs synr audioread
resampy