tensorflow-lite fixes

[vivien]

This PR makes it possible to compile libcamera-apps from meta-raspberrypi with tflite support.

[taos-ci]

📝 TAOS-CI Version: 1.5.20200925. Thank you for submitting PR #85. Please a submit 1commit/1PR (one commit per one PR) policy to get comments quickly from reviewers. Your PR must pass all verificiation processes of cibot before starting a review process from reviewers. If you are new member to join this project, please read manuals in documentation folder and wiki page. In order to monitor a progress status of your PR in more detail, visit http://ci.nnstreamer.ai/.

[taos-ci]

cibot: @vivien, A builder checker could not be completed because one of the checkers is not completed. In order to find out a reason, please go to http://ci.nnstreamer.ai/meta-neural-network/ci/repo-workers/pr-checker/85-0-664b/.

[myungjoo]

We have a few systems that install both tf-lite 1.x and tf2-lite 2.x, which forced us to install this file as tensorflow2-lite.pc.

For your needs, I'd recommend to make a "meta package" that creates symbolic links from tensorflow2-lite* to tensorflow-lite* for the compatibility of both tensorflow2* users and tensorflow* users.

[vivien]

I replied in the PR conversation for clarity ;)

[myungjoo]

This breaks packages using tensorflow2-lite.*

Please consider creating symbolic links from tensorflow2* to tensorflow* optionally or by another package.

[vivien]

For your needs, I'd recommend to make a "meta package" that creates symbolic links from tensorflow2-lite* to tensorflow-lite* for the compatibility of both tensorflow2* users and tensorflow* users.

I added the following to the concerned layer (here meta-raspberrypi) and this works fine:

$ cat meta-raspberrypi/dynamic-layers/neuralnetwork/recipes-tensorflow/tensorflow-lite/tensorflow-lite_2.%.bbappend 
do_install:append () {
	ln -sf libtensorflow2-lite.a ${D}${libdir}/libtensorflow-lite.a
	ln -sf tensorflow2-lite.pc ${D}${libdir}/pkgconfig/tensorflow-lite.pc
}

But I do not think that requiring this for all layers using tensorflow-lite is the correct packaging approach, see below.

We have a few systems that install both tf-lite 1.x and tf2-lite 2.x, which forced us to install this file as tensorflow2-lite.pc.

I understand your use case, however the package seems broken for a few reasons:

• the project builds libtensorflow-lite-bundled.a, not libtensorflow2-lite-bundled.a
• headers are installed into ${includedir}/tensorflow/lite, regardless of the version
• ${includedir}/tensorflow/core/public/version.h is installed regardless of the version
• ${includedir}/flatbuffers/include/flatbuffers/*.h are installed regardless of the version

The r2.* branches are just a version bump from the tensorflow-lite project, not a different project.

Because you have a special scenario, I would suggest that you install the various versions in different prefixes on your system (e.g. /opt/tflite/1.12, /opt/tflite/1.15, /opt/tflite/2.3, /opt/tflite/2.11, etc.) and configure your system or applications to look for the correct version there.

A way to do that is to have a dynamic tensorflow-lite_%.bbappend that is enabled only for your use case, which moves the installed files to /opt/tflite/${PV} for example.

What do you think?