How to fix "Are you using old TFLite binary with newer model?"

Question

I have a very small custom model built with tensorflow 2.0 that I can convert to a model.tflite and run it on a Coral, but when I load this model into the jevois and run the jevois tensorflow saliency demo I get the following error, from the console.

<!--StartFragment-->INF TensorFlow::operator(): Loaded model /jevois/share/tensorflow/fma224/model.tflite

ERR TensorFlow::Report: Op builtin_code out or range: 74. Are you using old TFLite binary with newer model?

ERR TensorFlow::Report: Op builtin_code out or range: 82. Are you using old TFLite binary with newer model?

ERR TensorFlow::Report: Registration failed.

FTL TensorFlow::operator(): Failed to construct interpreter

Where does the TFLite binary live so that I can replace it with a new one?<!--EndFragment-->

what version of tensorflow does the jevois use by default?

Answer 1

Darn, we are kind of stuck at TF 1.9.9, please see the exact version in

https://github.com/jevois/jevoisbase/blob/master/Contrib/reinstall.sh

we get this exact version (the long number is the commit ID on GitHub):

get_github tensorflow tensorflow 25c197e02393bd44f50079945409009dd4d434f8

And then we compile only the C++ source files we need (discarding anything that would try to open a window and such), and those get integrated into libjevjoisbase, please see here, section about TensorFlow:

https://github.com/jevois/jevoisbase/blob/master/CMakeLists.txt

We have tried to update to a newer version but their build system has changed and we have not yet been able to track down what exactly gets compiled and how through their bazel-based build. If you can somehow figure out an updated list of sources for the one that starts at line 378 of the above CMakeLists.txt, then we can try that update to a newer tensorflow again.

Otherwise, maybe you can make your model work under TF 1.9.9?