Can't build with the Tensorflow Lite on Raspberry pi 4B, 64 bit

[DevelopDaily]

I am aware of the solution described in the issue #132.

But my problem is that I use Pi OS Bullseye (64-bit), the prebuilt lib does not seem to be compatible. I read the builder's website, which states " Binaries are compatible with Raspberry Pi OS Bullseye (32-bit)."

Does that mean the prebuilt lib for 64-bit does not exist anywhere?

Anyway, I tried to build the latest Tensorflow Lite on the 64-bit and the build succeeded. The trouble is that the libcamera-apps cannot link with it because of these errors:

/usr/bin/ld: /usr/local/lib/libtensorflow-lite.so: undefined reference to `tensorflow::profiler::internal::g_trace_level'
/usr/bin/ld: /usr/local/lib/libtensorflow-lite.so: undefined reference to `tensorflow::profiler::TraceMeRecorder::Record(tensorflow::profiler::TraceMeRecorder::Event&&)'
/usr/bin/ld: /usr/local/lib/libtensorflow-lite.so: undefined reference to `tensorflow::profiler::GetCurrentTimeNanos()'
/usr/bin/ld: /usr/local/lib/libtensorflow-lite.so: undefined reference to `tensorflow::profiler::ScopedMemoryDebugAnnotation::ThreadMemoryDebugAnnotation()'

Could you please offer some suggestions?

Thanks.

[naushir]

I'm afraid we only support TF on our 32-bit releases for this very reason. There has always been a problem building/compiling TFlite in our 64-bit environment.

I don't really have any suggestions as I've not looked into this for a long time, but you might want to revert to a 32-bit build for now if you want to use this module.

[DevelopDaily]

Thanks for the comments. I will keep this issue open for a while because I am trying to convince the Tensorflow people to fix their problems. This is not a libcamera problem.

I fixed it on my local machine. Basically, those "profiler" references, which reside on the tensorflow main branch, should never be referenced inside the TF Lite. Unfortunately, they are referenced, but the CMakeLists.txt of the Lite has never populated any of the code from the main branch. No wonder we see undefined reference.

The tricky part is that I am tempted to use populate_tf_source_vars to bring some code to the Lite. But, that opens a nasty can of worms because the references will escalate to reference a lot of other main branch stuff. That defeats the very motivation of the "Lite" version of Tensorflow, doesn't it?

[naushir]

Is there anything on our side to be done here? If not, I'll close this issue.

[DevelopDaily]

I can close it. The issue was filed with Tensorflow. I am not sure how and when they are going to fix it.