Summary

This repository contains python wheels where no pre-built version is available.

The binaries can be downloaded under relases.

Funding

The project was funded by the German Federal Ministry of Education and Research under grant number 01IS22S34 from September 2022 to February 2023. The authors are responsible for the content of this publication.

aarch64

These are platforms like the Raspberry Pi 4 or NVIDIA Xavier NX.

All compilations are targeting Python 3.8. Install a

git clone https://github.com/pyenv/pyenv.git ~/.pyenv
# Follow instructions on https://github.com/pyenv/pyenv
sudo apt-get install libbz2-dev libssl-dev libreadline-dev libsqlite3-dev libffi-dev
ldconfig
pyenv install 3.8.16

Tensorflow

A cross-platform build with tensorflow/tools/ci_build/ci_build.sh PI-PYTHON38 tensorflow/tools/ci_build/pi/build_raspberry_pi.sh AARCH64 does not work and fails currently, hence the compilation must be executed on a Raspberry 4 directly (takes 10+ hours)

Before python dependencies can be installed run:

sudo apt-get install gcc g++ python3-dev libhdf5-dev

Build TensorFlow directly on Raspberry Pi 4

NOTE: In case the build configuration is changed it might be necessary to delete the bazel cache with rm -rf ~/.cache/bazel/_bazel_ubuntu.

git clone [email protected]:tensorflow/tensorflow.git
cd tensorflow
git checkout v2.8.4
wget https://github.com/bazelbuild/bazel/releases/download/4.2.1/bazel-4.2.1-linux-arm64
chmod 755 bazel-4.2.1-linux-arm64
sudo mv bazel-4.2.1-linux-arm64 /usr/local/bin/bazel

The Python version used for the build is Python 3.8 which is the default on Ubuntu 20.04.

Specify the march flag listed in gcc -march=native -Q --help=target while running the interactive configure script. It should be -march=armv8-a+crc

Instructions are taken partly from qengineering.

./configure
pyenv local 3.8.16
pip3 install numpy==1.23.5 wheel 
pip3 install keras_preprocessing --no-deps
sudo apt install python-is-python3
tmux
bazel --host_jvm_args=-Xmx6000m build --local_cpu_resources=2 --config=nogcp --config=nohdfs --config=nonccl --config=v2 --config=noaws --copt=-ftree-vectorize --copt=-funsafe-math-optimizations --copt=-ftree-loop-vectorize --copt=-fomit-frame-pointer --config=opt //tensorflow/tools/pip_package:build_pip_package
./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
# copy the whl file from /tmp/tensorflow_pkg to aarch64 and zip it

Tensorflow Addons

For Raspberry Pi 4 TensorFlow Addons is needed for Rasa. There is no pre-built pip wheel.

Execute on the Raspberry Pi 4:

git clone https://github.com/tensorflow/addons.git
cd addons
git checkout v0.17.1
pyenv local 3.8.16
python3 -m venv venv
source venv/bin/activate
pip3 install ../python-wheels/aarch64/tensorflow-2.8.4-cp38-cp38-linux_aarch64.whl
python3 ./configure.py
bazel build build_pip_pkg
bazel-bin/build_pip_pkg artifacts
# relase the whl

Tensorflow Text

For Raspberry Pi 4 TensorFlow Text is needed.

Execute on the Raspberry Pi 4:

git clone [email protected]:tensorflow/text.git
cd text
git checkout v2.8.2
pyenv local 3.8.16
python3 -m venv venv
source venv/bin/activate
pip3 install aarch64/tensorflow-2.8.2-cp38-cp38-linux_aarch64.whl
./oss_scripts/run_build.sh
# relase the whl

confluent-kafka

The python module for confluent-kafka librdkafka is needed.

Compile the latest version of the librdkafka package:

git clone https://github.com/confluentinc/librdkafka.git
cd librdkafka/
git checkout v1.9.2
sudo apt-get install gcc make libevent-pthreads-2.1-7 
./configure
make
sudo make install

Build the Python wheel:

pyenv local 3.8.16
python3 -m venv confluent-kafka
source confluent-kafka/bin/activate
pip install confluent-kafka
find ~/.cache/pip/wheels/ -name *_kafka*
# copy wheel from cache folder
cp ~/.cache/pip/wheels/<path>/confluent_kafka-1.9.2-cp38-cp38-linux_aarch64.whl aarch64/
rm -rf confluent-kafka

Atom Based Processors

These are platforms like the Rock Pi X or MeLe Quieter2Q.

Tensorflow

The x86_64 wheels for TensorFlow are causing:

The TensorFlow library was compiled to use AVX instructions, but these aren't available on your machine.

Use a Docker image for building TensorFlow (a direct build was missing some dependencies) on a different Ubuntu based system. The direct compilation on the Rock Pi X is slow and fails due to memory problems.

docker pull tensorflow/tensorflow:devel
docker run -it -w /tensorflow_src -v $PWD:/mnt -e HOST_PERMS="$(id -u):$(id -g)" tensorflow/tensorflow:devel bash

Inside the container execute:

git pull
git config remote.origin.fetch '+refs/heads/*:refs/remotes/origin/*'
git fetch --all
git checkout v2.8.4

The Python version used for the build is Python 3.8, and the TensorFlow Docker image is only using version 3.6 on Ubuntu 18.04.

sudo apt-get install nano
git clone https://github.com/pyenv/pyenv.git ~/.pyenv
# Follow instructions on https://github.com/pyenv/pyenv
source ~/.profile 
source ~/.bashrc
sudo apt-get install libbz2-dev libssl-dev libreadline-dev libsqlite3-dev libffi-dev
ldconfig
pyenv install 3.8.16
pyenv local 3.8.16

Specify the march flag listed in gcc -march=native -Q --help=target while running configure. It should be

• Rock Pi X: -march=silvermont
• LeMe Quieter2Q: -march=goldmont-plus

./config
wget https://github.com/bazelbuild/bazel/releases/download/4.2.1/bazel-4.2.1-linux-x86_64
chmod 755 bazel-4.2.1-linux-x86_64
sudo mv bazel-4.2.1-linux-x86_64 /usr/local/bin/bazel
pip3 install numpy==1.23.5 wheel 
pip3 install keras_preprocessing --no-deps
bazel build --config=mkl --config=opt //tensorflow/tools/pip_package:build_pip_package
./bazel-bin/tensorflow/tools/pip_package/build_pip_package /mnt
chown $HOST_PERMS /mnt/tensorflow-2.8.4-cp38-cp38-linux_x86_64.whl
# copy files from host computer
# relase the whl from directory where Docker was started
# quit the docker image