Why does trained model work in MATLAB 2021 but not 2020?

I have a model (initially converted from Tensorflow) which I trained in MATLAB 2021a.

I am now trying to run in 2020b and am getting the following error message:

Warning: While loading an object of class 'DAGNetwork':
Array indices must be positive integers or logical values. 
> In classify_LT (line 2)
In run (line 2) 
Dot indexing is not supported for variables of this type.

Error in DAGNetwork/get.NumInputLayers (line 173)
            val = this.PrivateNetwork.NumInputLayers;

Error in DAGNetwork/predict (line 140)
if this.NumInputLayers == 1

I believe this may be an issue with the Deep Learning Toolbox for Tensorflow-Keras Models toolbox, but am not sure.

Code can be found here, too large to upload zipped.

Answer 1

[Cris LaPierre]

I get the error message in R2020b, R2021a, R2021b, and R2022a prerelease.

If it is working in R2021a for you, do you have any add-ons or toolboxes installed there that are not installed in R2020b?

If you can't get it working, I would suggest contacting support: https://www.mathworks.com/support/contact_us.html

[Cris LaPierre]

One suggestion I was given was ensuring that all layers in your network have names. They currently do not. Can you try that?

[Cris LaPierre]

I have confirmed that this is a bug in R2020b that was fixed in R2021a. Can you update your R2020b to R2021a? If not, I suggest contacting support.