Going from trainNetwork to trainnet

I've attached 3 files(see post below for latest version of these files):

• trainNetworkEXAMPLE - my original trainNetwork implementation
• trainnetEXAMPLE - the trainnet implementation
• example.csv - data file with predictors and targets

The codes for the two examples are identical, the difference is only in the formatting of the input matrices.

trainNetworkEXAMPLE works as expected.

trainnetEXAMPLE works but convergence of the solver is different and solution is poor.

Both codes end with:

Training stopped: Met validation criterion

What am I getting wrong?

[Matt J]

Seemingly nothing. Why do you think something is wrong?

[psousa]

NMSE for training set and test set are considerably worse for version using trainnet.

Answer 1

[Sourabh]

Hi @psousa

I too encountered the similar issue when using “trainnet” and “trainNetwork” method.

The workaround that worked in my case was to:

1. Use @mse as the loss function instead of "mse" in “trainnet”.

  [net,info] = trainnet(XTrain,TTrain,layers,@mse,options); 

2. Set 'GradientThreshold' to ‘Inf’ in ‘trainingOptions’ of both the programs.

  options = trainingOptions('adam',  
  ... 
      'GradientThreshold',Inf, 
  ... 
  ); 

Kindly refer to the below image:

[psousa]

Matlab support got back to me within a couple weeks and that was the solution they offered.

Thanks for looking into it.