How to define FusedBatchNormV3 function in Custom Layers with Custom DenseNet?

[Lin__Vaan]

Hi all,

Firstly I make my custom model, DenseNet of 13 layers, and I try to convert my model to IR files with the following error:

[ ERROR ]  List of operations that cannot be converted to Inference Engine IR:
[ ERROR ]      FusedBatchNormV3 (3)
[ ERROR ]          batch_normalization/FusedBatchNormV3
[ ERROR ]          batch_normalization_1/FusedBatchNormV3
[ ERROR ]          batch_normalization_2/FusedBatchNormV3
[ ERROR ]  Part of the nodes was not converted to IR. Stopped.
 For more information please refer to Model Optimizer FAQ (https://docs.openvinotoolkit.org/latest/_docs_MO_DG_prepare_model_Model_Optimizer_FAQ.html), question #24.

Next I follow this url: https://github.com/david-drew/OpenVINO-Custom-Layers/blob/master/2019.r2.0/ReadMe.Windows.2019.r2.md

And it works! My model now can be converted to .bin and .xml file.

Now I need Inference Engine Custom Layer Implementation for the Intel® CPU, the url give me a sample that define cosh function in ext_cosh.cpp, but I don't know how to define FusedBatchNormV3 function in .cpp, could anyone have some solutions about this?

I use Anaconda and create two environment: TF 1.15 and TF 2.0, my custom model was constructed and training on TF 2.0, and I converted it on TF 1.15.

Here is my model and .h5 file.