create target for neural network

3 views (last 30 days)
Hi all,
I had extracted feature vector of an image and saved it in a excel document. feature vector is 42x42 dimension. I don't know how to create target for this input so i can train the neural network. I need to do emotion classification. 5 classes. So, how do i create target vector and train the network?

Accepted Answer

Chandra Kurniawan
Chandra Kurniawan on 13 Feb 2012
Hi, Remo
Feature vector that extracted from an image ideally m x 1 matrix.
You said that it's 42 x 42. Is it means 42 feature of 42 images??
If so, then you need to create 5 x 42 matrix of target.
First create matrix of zero 5 x 42.
for feature#1, this feature belong to which emotion? Let say 3rd class. Then you need to mark at first column 3rd row as 1.
t = [0 0 0 0 0 ... until 42
0 0 0 0 0 ...
1 0 0 0 0 ...
0 0 0 0 0 ...
0 0 0 0 0 ...
And so on.
I have an example below. In this example I have 5x5 feature vector. and 5x5 matrix of target.
p = [1 0 1 0 1
0 1 1 0 1
1 0 0 1 1
0 1 0 1 1
1 0 0 1 1];
t = [0 0 1 0 0
0 0 0 0 1
1 0 0 0 0
0 1 0 0 0
0 0 0 1 0];
PR = [0 1;0 1;0 1;0 1;0 1];
net = newff(PR,[5 25 25 5],{'logsig','logsig','logsig','logsig'},'traingda');
net.trainParam.epochs = 1500;
net.trainParam.goal = 0;
net = train(net,p,t);
Then try to simulate the first feature.
sim(net,p(:,1))
And the result :
ans =
0.0032
0.0003
0.9955
0.0000
0.0029
It means this inputs belong to 3rd class.
I hope this will helps.
  3 Comments
gfgf99
gfgf99 on 31 Aug 2015
Edited: gfgf99 on 31 Aug 2015
Chandra Kurniawan Can you please explain above written code for more convince .
PR = [0 1;0 1;0 1;0 1;0 1]; net = newff(PR,[5 25 25 5],{'logsig','logsig','logsig','logsig'},'traingda'); net.trainParam.epochs = 1500; *net.trainParam.goal = 0; net = train(net,p,t);* .
Thank you.
Best regard
debasmita bhoumik
debasmita bhoumik on 9 Apr 2016
I have made two feature matrices ( the purpose was to classify weather infection exist or not in a leaf) of size 80x16 and 40x16 respectively. The first one was for training images(80 images with 16 features) and other is test images(40 images with 16 features). I had also classified them accurately using SVM classifier. Now i want to classify those same images with the help of NEURAL NETWORK. Please suggest what can I do

Sign in to comment.

More Answers (1)

Greg Heath
Greg Heath on 14 Feb 2012
remo asked on 12 Feb 2012 at 12:29
>I had extracted feature vector of an image and saved it in a excel document. feature vector is 42x42 >dimension. I don't know how to create target for this input so i can train the neural network. I need to >do emotion classification. 5 classes
The input features must be column vectors. If you extract a feature matrix from an original matrix, you can convert it to a feature vector using the colon operator. For example, if you have a feature matrix fm with size(fm) = [ r, c ], then fv = fm(:) will be the corresponding feature vector with size(fv) = [ r*c, 1 ].
If you have N feature vectors from N images, the N columns form an input matrix, p, with size(p) = [ I N ] and I = r*c.
If the N images come from k classes, the corresponding target matrix, t, contains N columns of the k-dimensional unit matrix eye(k). The row containing the "1" is the class index for the corresponding input vector. size(t) = [ O N ] with O = k.
The output matrix is readily formed using the function IND2VEC. For example
>> t = full(ind2vec([ 1 3 5 2 4 ]))
t =
1 0 0 0 0
0 0 0 1 0
0 1 0 0 0
0 0 0 0 1
0 0 1 0 0
Given a feature vector, the resulting elements in the output vector can be interpreted as approximations to the input-conditional class posterior probabilities. The input vector is then assigned to the class corresponding to the maximum output.
A single hidden layer with H nodes is sufficient. The resulting node topology is I-H-O. This will result in Neq = N*O nonlinear equations to be "solved" for Nw = (I+1)*H+(H+1)*O unknown weights. For training to convergence, it is required that Neq >= Nw. However, the condition Neq >> Nw is preferable to mitigate real-world noise and measurement error. This is equivalent to requiring
H <= Hub is required but H << Hub is desired where the upper bound Hub is given by
Hub = ( Neq - O ) / ( I + O + 1 ).
A practical value for H can be obtained by trial and error keeping the above conditions in mind.
Hope this helps.
Greg
  1 Comment
remo
remo on 14 Feb 2012
Hi Greg,
thank you for ur time in answering my question. I would try ur method in creating my target vector. Thank you sir.

Sign in to comment.

Categories

Find more on Image Data Workflows in Help Center and File Exchange

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!