Different results in training a CNN with Matlab 2018a and Matlab 2019a

Hi, I am training a CNN of classification.

I have trained it with the same seed in Matlab 2018a and Matlab 2019a, after adapting the initialization algorithm in the convolutional layers and the fully connected layer to 'narrow-nomal' in Matlab 2019a, in order to make it comparable. As well, I have set the validation patience in 12 for both releases.

However, I am still getting far different results. Is there anything else that I should take into account to have the same training results? Are there others updates in Matlab 2019a that I should consider?

Thank you in advance

Answer 1

[Greg Heath]

You are making the task difficult by going backwards.

Start with a single hidden node and add nodes one at a time.

Hope this helps

Greg

Answer 2

[Faraz Naqvi]

I also had this problem. In Matlab 2019 there are additional settings, if you see the properties of layers and compare them with the layer properties from previous version i.e. Matlab 2018 you will observe some extra fields in Matlab 2019.

For me, model model had 'WeightsInitializer' 'narrow-normal' is previous and 'glorot' in newer version. You can look at there aforemention link.

https://www.mathworks.com/help/deeplearning/examples/compare-layer-weight-initializers-example.html