Parallel computing in Neural Networks is not using all the Workers in Matlab 2015b

Hi everyone!

I am using a Time Delay Neural Network that takes a long time to be trained. So, I have used parallel processing and the training time had a great decrease.

The curious thing is that when using Matlab 2015a the parallel processing works all the workers:

pool = parpool;
tdnn = train(tdnn,Xs,Ts,'useParallel','yes','showResources','yes');

The result is:

Starting parallel pool (parpool) using the 'local' profile ... connected to 4 workers.

Computing Resources:
Parallel Workers:
  Worker 1 on MBP-de-Hugo, MATLAB on MACI64
  Worker 2 on MBP-de-Hugo, MATLAB on MACI64
  Worker 3 on MBP-de-Hugo, MATLAB on MACI64
  Worker 4 on MBP-de-Hugo, MATLAB on MACI64

However, when using the same code in Matlab 2015b, just one worker is used:

Starting parallel pool (parpool) using the 'local' profile ... connected to 4 workers.

Computing Resources:
Parallel Workers:
  Worker 1 on MBP-de-Hugo, MEX on MACI64
  Worker 2 on MBP-de-Hugo, Unused
  Worker 3 on MBP-de-Hugo, Unused
  Worker 4 on MBP-de-Hugo, Unused

How to use the others Workers in Matlab 2015b?

Thanks in advance!

Answer 1

[Tushar Athawale]

This is a bug in the "train" function of the Neural Network toolbox in the way the time information of the timedelaynet is analyzed prior to sending to the workers, making it appear internally as though the network training cannot be parallelized.

This behavior is a side-effect of the following bug:

https://www.mathworks.com/support/bugreports/1310148

To work around this issue, please follow these instructions in the above link to apply the patch.

[orlem lima dos santos]

hello I'm having the same problem but for matlab 2015a and the bugfix above is only for matlab 2015b

any suggestion?

[Hugo Mendonça]

Sorry for the completely delayed answer.

You just must get a latter matlab version or download a corrective patch on the link provided above.