Hi,
On this subject I finally don't use the Labview implementation of Tensorflow.
I use instead the C Api from tensorflow.dll (runtime library compiled by TF team).
https://www.tensorflow.org/install/lang_c
The advantage is that you're independent from the NI development and you can use the last stable TF version with GPU acceleration (CPU and GPU versions of the dll are available).
You also can get rid of the frozen model with all the previously mentioned bugs and use the native model.
To do that, I've created my own dll with MSVC. It instantiates a C++ object to maintain a pointer TFsession in Labview.
I also wrapped the dll call behind a Labview API and reserved the required output memory from Labview which was not obvious.
This way, I can compute inferences very quickly using the GPU.
For instance, classification on 256x256 images gets an answer after 4ms.
Thus, it can be integrated in real-time application.
HTH
Thierry
