using a trained ANN

hi all. I have trained a pattern recognition neural network and have gotten good results (87%). Although, i'm still confused as to how I actually use it in real life. For example, every time i run my network i have to train it and sometimes it takes more than a few tries to get to 87% accuracy. At times the accuracy is as bad as 26%. So my question is, how do i make sure my network remembers what it has learned? I want to save my networks memory when i get 87% accuracy. How do i do that? Second, i was wondering if i could use this network to find the class of an unknown image which i select at runtime. I've used indexing method to separate the training, validation and test data so that the network tests only the images i want it to. Thanks in advance. Have a nice day :)

[인국 강]

how do you run your ann model? because i have no idea

Answer 1

[Greg Heath]

 % FITNET REUSE EXAMPLE: 
 % Train in workspace
 % Save copy to directory
 % Clear original from workspace
 % Load copy from directory to workspace
 % Use copy on "new" data

% If it exists, delete netg from the directory

delete netg.mat

% Clear the workspace and plot before designing netg

 close all, clear all, clc
 [ x,t ] = simplefit_dataset;
 [ I N ] = size(x)        %[1 94]
 [ O N ] = size(t)        %[1 94]
 MSE00   = mean(var(t',1))% 8.3378

 subplot(2,1,1), hold on
 plot(x,'k'), plot(t,'b')
 subplot(2,1,2), hold on
 plot(x,t,'b')

% NOTE: t has 4 local extrema

 netg            = fitnet(4);
 rng(4151941)
 [ netg tr y e ] = train(netg,x,t);
 % y = netg(x); e = t-y;
 stopcriteria = tr.stop      % Validation stop
 NMSE         = mse(e)/MSE00 % 5.8958e-3
 R2           = 1-NMSE       % 0.9941
 plot(x,y,'r')

' netg is in workspace'

whos netg

'netg is not in directory'

dir netg 
dir netg.mat

'Save copy of netg to directory. Becomes netg.mat'

 save( 'netg')
 dir netg.mat

'Next clear original netg from workspace'

 whos netg
 clear netg
 whos netg

'Then load copy of netg from directory to workspace'

 load netg
 whos netg

'Delete copy of netg from directory'

 dir netg.mat 
 delete netg.mat
 dir netg.mat

'Apply netg copy in workspace to "new" data'

 ylr    = netg(fliplr(x));
 diffy  = minmax(ylr-fliplr(y)) % [ 0 0 ]

Hope this helps.

Thank you for formally accepting my answer

Greg

[farheen asdf]

what is X and t here? and what is netg and why do i have to keep deleting and loading it? And lastly, where exactly do i load my new data in this?

[Greg Heath]

Assignment statements [ x,t] = ... meaning [ input target ] = ... are standard syntax for MATLAB example data. See it's use in most of the help and doc examples. For example

 help fitnet
 doc  fitnet

Also see a list of available example datasets

 help nndatasets
 doc nndatasets

netg is the net "G"reg is using for this example.

I was showing how you can save the workspace trained net to the directory. Then at any future time, if it is not in the workspace, you can load it from the directory.

If I didn't erase it in the workspace before loading it from the directory, there would be no reason to load it.

Training is done in the workspace. That is where you read your data in for training and testing. Where the data comes from before you read it in to the workspace depends on where you or MATLAB have it permanently stored.

[farheen asdf]

 ylr    = netg(fliplr(x));
 diffy  = minmax(ylr-fliplr(y)) % [ 0 0 ]

is this the code for applying the network to new data? here x will be the features of the new image correct? and y is what?

[Greg Heath]

ynew = net(xnew);

I just used xnew = fliplr(x), ynew = ylr for convenience

y is defined above