Thread Subject: NN Object Size - NN Toolbox V6.0 (R2008a)

Hello,
Recently, while working on project that requires a relatively large training set, MATLAB kept on running out of memory. On further analysis, I noticed that the size of neural network object created with 'newff' command was abnormally large and it appears that it is due to the fact that NN object contains the input data set as one of its subobject structures.
--
  predictive_model 1x1 69797369 network
  training_input_data_set 50x170910 68364000 double
  
  >> predictive_model.inputs{1,1}
  
  ans =
  
       exampleInput: [50x170910 double]
--
Is there an option disable this feature of an NN object? This feature instead of helping to ease the burden of MATLAB memory utilization is becoming a contributing factor.

Also, am I missing soemthing or 'plotroc' or 'roc' functions can not actually calculate area under the ROC curve?

Thanks in advance for all the help.
Farooq

On Jan 6, 2:37=A0pm, "Farooq Azam" <fa...@ieee.org> wrote:
> Hello,
> Recently, while working on project that requires a relatively large train=
ing set, MATLAB kept on running out of memory. On further analysis, I notic=
ed that the size ofneuralnetwork object created with 'newff' command was ab=
normally large and it appears that it is due to the fact that NN object con=
tains the input data set as one of its subobject structures.
> --
> =A0 predictive_model =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 1x1 =
=A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 69797369 =A0network =A0 =A0 =A0 =A0 =A0 =A0
> =A0 training_input_data_set =A0 =A0 =A0 =A0 =A0 50x170910 =A0 =A0 =A0 =A0=
 =A0 =A068364000 =A0double
>
> =A0 >> predictive_model.inputs{1,1}
>
> =A0 ans =3D
>
> =A0 =A0 =A0 =A0exampleInput: [50x170910 double]
> --
> Is there an option disable this feature of an NN object? This feature ins=
tead of helping to ease the burden of MATLAB memory utilization is becoming=
 a contributing factor.
>
> Also, am I missing soemthing or 'plotroc' or 'roc' functions can not actu=
ally calculate area under the ROC curve?
>
> Thanks in advance for all the help.
> Farooq

Not familiar with R2008 (e.g., PLOTROC,ROC,...).

Perhaps the memory problems are with TRAINLM.
For a fix;

help train

and see alternative strategies of increasing
memory with LM training or sithching to CG

Hope this helps.

Greg

Dr. Greg,

Thank you for your response. I did code scaled conjugate gradient method myself and now problem is resolved.

Farooq

Greg Heath <heath@alumni.brown.edu> wrote in message <ca2bd280-417b-45dc-98ed-bca9e0bcf279@z28g2000prd.googlegroups.com>...
> On Jan 6, 2:37=A0pm, "Farooq Azam" <fa...@ieee.org> wrote:
> > Hello,
> > Recently, while working on project that requires a relatively large train=
> ing set, MATLAB kept on running out of memory. On further analysis, I notic=
> ed that the size ofneuralnetwork object created with 'newff' command was ab=
> normally large and it appears that it is due to the fact that NN object con=
> tains the input data set as one of its subobject structures.
> > --
> > =A0 predictive_model =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 1x1 =
> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 69797369 =A0network =A0 =A0 =A0 =A0 =A0 =A0
> > =A0 training_input_data_set =A0 =A0 =A0 =A0 =A0 50x170910 =A0 =A0 =A0 =A0=
> =A0 =A068364000 =A0double
> >
> > =A0 >> predictive_model.inputs{1,1}
> >
> > =A0 ans =3D
> >
> > =A0 =A0 =A0 =A0exampleInput: [50x170910 double]
> > --
> > Is there an option disable this feature of an NN object? This feature ins=
> tead of helping to ease the burden of MATLAB memory utilization is becoming=
> a contributing factor.
> >
> > Also, am I missing soemthing or 'plotroc' or 'roc' functions can not actu=
> ally calculate area under the ROC curve?
> >
> > Thanks in advance for all the help.
> > Farooq
>
> Not familiar with R2008 (e.g., PLOTROC,ROC,...).
>
> Perhaps the memory problems are with TRAINLM.
> For a fix;
>
> help train
>
> and see alternative strategies of increasing
> memory with LM training or sithching to CG
>
> Hope this helps.
>
> Greg