net reusing

Hi, I'm creating a set of nets using a code more or less like this:

net = patternnet(10);
% net initialization
% ...

for ii=1:10
   net = train(net,in,tar);
   netSet{ii} = net;
end

What I see is the first training takes relatively a lot of time, while the others usually are far faster. So a question comes to my mind: could be each new training starts from the previous net? If so, is there a function to re-randomize the initial weights (without the need to allocate a new net)?

Answer 1

[the cyclist]

Inside your loop, the input to train() is definitely the value of "net" that just calculated in the previous. I am guessing you want to do something like this instead:

net0 = patternnet(10);
% net0 initialization
% ...
for ii=1:10
   net = train(net0,in,tar);
   netSet{ii} = net;
end

This way, you are starting from the initialized value "net0" each time. I don't know enough about neural nets to know if train() has some randomness in it that will make each value of netSet(ii) different.

Also, note that you do not really need the intermediate value "net" inside your loop. You could just do

for ii=1:10
   netSet{ii} = train(net0,in,tar);
end

[Luca Cavazzana]

yes, but the real code is a bit more complex than the one I wrote here, each net has to be trained until a minimum performance rate is obtained, so the "intermediate net" becomes convenient.

Another side effect that comes to my mind is every time I call |train| a new combination of train, validation and test sets are chosen, so probably during the Nth |train| the function validates using some of the data in the (N-k)th one was used for training. After a lot of iterations probably all the data will be somehow used to train the net, causing overfitting.

(sorry for overusing the word "train"...)

Answer 2

[Greg Heath]

Patternnet is selfinitializing. Therefore

1. Initialize the rand RNG

2. Create an outer loop over nH number of candidate values for H, the number of hidden nodes

3. Create an inner loop of Ntrials random initialization designs.

4. Bottom Line: Each of the nH*Ntrials designs begins with a different set of initial weights.

Hope this helps.

Greg

[Luca Cavazzana]

That's more or less what I did.

I was just wondering if |train| re-randomizes the weights by itself.

[Greg Heath]

No, weight randomization only occurs when the net is created

( e.g., net = newff(...))

Greg

[Greg Heath]

No, weight randomization only occurs when the net is created

( e.g., net = newff(...))

Greg

Answer 3

[Greg Heath]

With the current set of net creation functions ( e.g., fitnet, patternnet, feedforwardnet,...) weight initialization does not occur at net creation.

If weights have not been assigned, they will be automatically initialized by train. Otherwise, train will just use the existing weights.

Therefore, when designing multiple nets in a loop, the function configure must be used to initialize the nets at the top of the loop before train is called.

 net = configure(net, x, t);