Can a user-defined function be called after weight/bias updates during training?

Hello.

I was wondering if there's a way to specify a function to be executed after weight/bias updates when calling train() or adapt()?

My current solution is to call train multiple times with net.trainParam.epochs=1, but this appears to negatively impact learning.

Thanks

-- Jeff

[Greg Heath]

What is this function supposed todo?

[jeffrey cockburn]

Hi Greg. Well, I would like to track the weights as they update, and I would also like to modify some of them myself along the way. I'm currently just running train with trainParam.epochs=1 to do so, but as we've chatted about, this can shoot the learning alg in the foot.

Thanks again.

Answer 1

[Greg Heath]

In a recent post originated by Jason, I've shown that interrupted training using net.trainParam.epochs = 1 without a validation set can yield the same weights as continued training. However, it may take much longer. How much longer will be different for different problems.

WB(:,i) = getwb(net);

Is sufficient for saving the weights of the ith epoch.

Hope this helps.

Thank you for formally accepting my answer.

Greg

[azie]

hi greg, im using r2008a, i cant find the getwb function. does it only offer using newer version?

Answer 2

[Greg Heath]

The reason the stopping/starting negatively impacts learning is because the learning parameters are reinitialized every time train is called. Even if you obtain the final parameters from

net.trainParam: .showWindow, .showCommandLine, .show, .epochs,
                  .time, .goal, .min_grad, .max_fail, .mu, .mu_dec,
                  .mu_inc, .mu_max

and/or

tr: .mu, .gradient, .val_fail, etc

there is no way to use them to initialize the next call of train unless the function is modified.

Perhaps that will be available in a new version. (I've suggested that to the TBX author.)

Hope this helps.

Thank you for formally accepting my answer.

Greg