无法在树莓派 4 上安装 Tensorflow 2.3

Question

I can't currently install TensorFlow 2.3 on a Raspberry Pi 4. Unfortunately, the pip doesn't return any TensorFlow version:

pip install tensorflow==
Looking in indexes: https://pypi.org/simple, https://www.piwheels.org/simple
ERROR: Could not find a version that satisfies the requirement tensorflow==
ERROR: No matching distribution found for tensorflow==

I read somewhere that TensorFlow only supports the 64-bit version of Python and only the versions between 3.5 to 3.8 so I made sure that I have these versions installed as well.

The current Linux distro that I have is the following aarch64 .

pip and python have the following versions:

pip -V python
pip 20.3.1 from /home/pi/envs/awe/lib/python3.7/site-packages/pip (python 3.7)

python -V
Python 3.7.3

And I also have a 64-bit version of Python:

platform.architecture()
('64bit', 'ELF')

Why can't pip find a compatible TensorFlow of version 2.3?

Answer 1

For the Raspberry Pi 4 ARM64 platform TensorFlow is not available as pre-built version for pip and must be compiled manually. Follow these instructions and cross compile it on a different more powerful environment, not on the Pi 4. The instructions are no up to date anymore, I have completed them here. These are the steps to cross compile TensorFlow 2.3.1 on an Ubuntu Linux:

git clone git@github.com:tensorflow/tensorflow.git
cd tensorflow

The old version 2.3.1 does not know Python 3.8. When using Python 3.7 then this might not be necessary.

wget https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/ci_build/Dockerfile.pi-python38 -o tensorflow/tools/ci_build/Dockerfile.pi-python38

CMake is now also required, I'm using here the master branch, which is of version 2.5.0 (in case this does not work anymore in the future with the master version):

wget -O tensorflow/tools/ci_build/install/install_cmake.sh https://raw.githubusercontent.com/tensorflow/tensorflow/master/tensorflow/tools/ci_build/install/install_cmake.sh
chmod +x tensorflow/tools/ci_build/install/install_cmake.sh

To support Python 3.8 add support for it, again taken from the master branch (version 2.5.0):

wget -O tensorflow/tools/ci_build/install/install_pi_python3x_toolchain.sh https://raw.githubusercontent.com/tensorflow/tensorflow/master/tensorflow/tools/ci_build/install/install_pi_python3x_toolchain.sh
chmod +x tensorflow/tools/ci_build/install/install_pi_python3x_toolchain.sh

You also need these file for the build:

wget -O tensorflow/tools/ci_build/pi/build_raspberry_pi.sh https://raw.githubusercontent.com/tensorflow/tensorflow/master/tensorflow/tools/ci_build/pi/build_raspberry_pi.sh
chmod +x tensorflow/tools/ci_build/pi/build_raspberry_pi.sh

There is an issue with newer numpy version. Use this cherry-pick from git:

git cherry-pick 75ea0b31477d6ba9e990e296bbbd8ca4e7eebadf

Now finally this can be compiled. This takes several hours most likely.

tensorflow/tools/ci_build/ci_build.sh PI-PYTHON38 tensorflow/tools/ci_build/pi/build_raspberry_pi.sh AARCH64

If something fails and some files have to be patched or added keep in mind to clean the Docker container to prevent using the Docker cache when restarting the build process. docker images , docker rmi <imageid> and docker rm <container> are your friends.