Dividerand - neural network training

I'm using a neural network model and for some simulations I've used the function 'dividerand' instead of 'divideind'. Without using the command [net,tr]=train(net,.....) where I could check how the data were randomly separated in training, validation and testing datasets, is it possible to verify how the random separation was executed (analysing the indices)?

Answer 1

[Greg Heath]

If you do not include tr as a training output, the only way to obtain the dividerand indices is to call dividerand before calling newff/newfit/fitnet/patternnet/feedforwardnet and nullify it's use within those net creation functions.

However, if all you are concerned about is duplicating a run, just specify the same random number seed before calling the creation function.

Hope this helps.

Greg

Answer 2

[Sean de Wolski]

I am not sure I understand. You can analyze the indces just by looking at the outputs from dividerannd. Can you please clarify your question a little further and/or provide a short example?

Answer 3

[Greg Heath]

The structure tr in the double output [net tr ] = train(net,x,t); will contain the train/val/test indices.

Greg

Answer 4

[JSousa Sousa]

Thanks for your prompt answers Sean de Wolski and Greg Heath.

In some simulations I've just used the command net= train(net,...), without specifying an output tr where I could have a structure with the train/val/test indices. In that way, is that any possibility to verify how the data were effectively split in different substets?

I know that to repeat a simulation using the same scenario it must be well known: - the network weights/biases initialization (using for example the command revert to obtain the initial parameters) and - the data separation criteria (difficult to me to recognize without the tr structure).

Thanks for your time and attention

João.

Answer 5

[JSousa Sousa]

I have a saved neural network model with good results that I would like to analyse in more detail. So the point is well understood, I'm interested to duplicate this training process (another similar run). However without a tr structure and also without the random number seed saved for this case I think I won't get exactly the same data division.

Do you know how to recover this seed?

Thanks again

JSousa.

[Greg Heath]

No.

You should always specify the random number seed/state before calling a function that uses random numbers.

The only thing you can do now is to use a loop to create a lot of designs and try to find one that has a similar performance to the original.

Greg