Differences between Deep Learning Toolboxes versions 18b and 19a

I have successfully set up a GoogLeNet network using the Deep Learning Toolbox (version 19a) on a standalone computer. I copied the LayerGraph variable to my networked computer which runs version 18b. When I load the variable I get the following warnings and the LayerGraph variable is empty:

Warning: While loading an object of class 'nnet.cnn.layer.ImageInputLayer':

Reference to non-existent field 'IsAverageImageMeanPerChannel'.

Warning: While loading an object of class 'nnet.cnn.layer.FullyConnectedLayer':

Index exceeds the number of array elements (1).

Warning: While loading an object of class 'nnet.cnn.LayerGraph':

Dot indexing is not supported for variables of this type.

I have read the release notes for the Deep Learning Toolbox version 19a, and cannot see why it should not be backward compatible (although I admit that I did not fully understand all the comments).

Am I trying to do something that is impossible, or have I overlooked something?

Answer 1

[Shubh Sahu]

'IsAverageImageMeanPerChannel' is introduced in R2019a and for 'final layers' dot notation is not supported in 2018b. For more information refer these links

• https://www.mathworks.com/help/releases/R2018b/deeplearning/ug/pretrained-convolutional-neural-networks.html

• https://www.mathworks.com/help/releases/R2018b/deeplearning/examples/train-deep-learning-network-to-classify-new-images.html

[Benjamin Zeff]

I have run into the same problem as Nick recently, trying to share a DAGNetwork created in MATLAB 2019a with someone using 2018b. I understand the basic idea here that there were some classes/fields that have changed. Is there any way, however, to write out the trained network to make it backward compatible?

I have been able to share trained networks of this type between the two systems in the past. In those cases, the networks were created by the default segnetLayers(...) command and then trained. There was no problem opening them and using them in R2018b. My newer network uses the replaceLayers function (R2019a) to create a non-normalized input and a weighted classification layer output. This network cannot be opened in R2018b, and I get the errors above. But the final architecture of the new and old networks is basically the same! Where is the 'IsAverageImageMeanPerChannel' and 'final layers' dot notation coming in? Can I circumvent that?

Thank you for any help you can offer.