How do I solve the "Referene to non-existent field" error when importing a Tensorflow model?

Hello everyone,

this is my first question on this forum.

I worked for a while in python (Tensorflow) on a simple classification network from my research (I'm a PhD student) and now I'd like to import it in Matlab (I need it in both the frameworks). I did some research and I found the

importKerasNetwork
importKerasLayers

functions in the Deep Learning Toolbox which should solve my problem. Sadly, I was using Keras 2.3.0 in my model (Tensorflow 2.2.0) and I kept getting this warning:

Warning: File '2020_06_19_11_49_10.h5' was saved in Keras version '2.3.0-tf'. Import of Keras versions newer than
'2.2.4' is not supported. The imported model may not exactly match the model saved in the Keras file.

together with the following error:

Reference to non-existent field 'dense_2'.

The first question is: can I do something to still use this model in Matlab?

I decided to go back to Keras 2.2.4 (with Tensorflow 2.1.0) and to re-train my network. Before doing this I decided to try and convert a simple non-trained network (some conv2d, max pooling, dense) but I still got the same error (not the warning, of course):

Reference to non-existent field 'dense_1'.

What am I doing wrong?

Thank in advance for the help,

Diego

UPDATE

This is not only a problem of the dense layer, it also happens when I remove the dense layers from my model, the error simply goes to another kind of layer.

Answer 1

[Sivylla Paraskevopoulou]

You can try to use the importTensorFlowNetwork function, which was introduced in R2021a. importTensorFlowNetwork supports newer TensorFlow versions, imports TensorFlow models in savedModel format, and tries to generate a custom layer when the software cannot convert a TensorFlow layer into an equivalent built-in MATLAB layer.