After forcing the Jetson OS (based on the ubuntu 18.04) to have Python 3.8 running. After having the JetPack 4.6.3 installed in the Jetson Nano Jetson OS.

When attempting to install TensorFlow 2+ manually in the Jetson Nano based on Official TensorFlow for Jetson Nano! – Jetson & Embedded Systems / Jetson Nano – NVIDIA Developer Forums

Or running the pip3 install command:

sudo pip3 install --pre --extra-index-url https://developer.download.nvidia.com/compute/redist/jp/v46 tensorflow

After a long time installing all the dependency for TensorFlow. A dreaded error message will occur.

ERROR: Failed to build installable wheels for some pyproject.toml based projects (h5py)

To understand the error, scroll up to read what are the error.

The first error message scrolling up suggest the version of hd5 are not supported.

Scroll higher, the error was triggered when pip is building the h5py dependency/library.

Hence, forming the problem statement, h5py library for Python 3.8 cannot be installed due to the binary hdf5 of Jetson OS is version 1.10.0.

First approach will be updating the Jetson OS using apt-get to update the Jetson OS hdf5 library to at least version 1.10.0

First attempt used based on command:

sudo apt-get install libhdf5-serial-dev hdf5-tools libhdf5-dev

First attempt does not work, as minor version was not added to the repositories of dated Jetson OS (based on Ubuntu 18.04).

Second attempt, build the hd5 python library via source. Based on idea offered at Installation — h5py 3.13.0 documentation

After numpy is installed, proceed to build the h5py, but encountered a different error.

The second attempt, the error encounter suggesting that the gcc compiler was not able to find header file, the Python.h during its compilation. Hence error.

After taking some time to research, installing python without the development tool kit will not able to help compilation due to missing header.

To fix this error, install the proper development library into Jetson OS.

sudo apt install libpython3.8-dev

H5PY_SETUP_REQUIRES=0 pip3 install . --no-deps --no-build-isolation

Lets rerun the installation attempt to reinstall the TensorFlow.

sudo pip3 install --pre --extra-index-url https://developer.download.nvidia.com/compute/redist/jp/v46 tensorflow

Real test to validate TensorFlow successful installation works.

And installing TensorFlow on dated Jetson Nano for the Jetson OS is challenging. It will help if a person has knowledge of Ubuntu, Python library management. The part that was left out was how to install Python 3.8.