Deploying neural network in R2014a using genfunction instead of sim and .mat file means the application has to be recompiled and deployed each time retraining is required

Previously in older releases of Matlab (e.g R2011a) and the Neural network tool box it was possible to train a network and save it in a .mat file. You could then compile the code that runs the neural network and the code would use the .mat file and the sim function to run the network. This meant that if you needed to retrain the network, you could simply rerun the training and replace the .mat file without having to redeploy and reregister the whole solution.

We have a deployed a neural network solution running from Excel on many machines and we don’t want to have to re-register .dll files every time we need to recalibrate our models. We recalibrate every few months and previously this just meant replacing .mat files.

Can you please advise how to do the above with the new version of Matlab (R2014a) and the neural network toolbox given that it now no longer supports the sim function and instead uses the genfunction?

Or is this no longer possible, meaning recalibration requires full redeployment? If so, why would Mathworks do this and make things so much less flexible?

Thanks

Answer 1

[David Ringo]

Turns out the SIM function is still available in R2014a. The issue was actually unrelated to the SIM function but rather to a bug in the Matlab Compiler as below:

http://www.mathworks.com/support/bugreports/1059029

32-bit compiled components generated with mcc and deploytool from MATLAB R2014a fail with 'undefined function' during runtime Description

On Windows, when using 32-bit MATLAB R2014a to deploy MATLAB code using mcc or deploytool, the deployed component fails during runtime with the 'undefined function' error.

This occurs because 32-bit MATLAB is most often installed into the 'C:\Program Files (x86)' directory, and the parentheses in the path name cause some functions to be missing in the generated component.