How can I normalize data between 0 and 1 ? I want to use logsig...

All is in the question: I want to use logsig as a transfer function for the hidden neurones so I have to normalize data between 0 and 1. The mapminmax function in NN tool box normalize data between -1 and 1 so it does not correspond to what I'm looking for.

Answer 1

[José-Luis]

 bla = 100.*randn(1,10)

 norm_data = (bla - min(bla)) / ( max(bla) - min(bla) )

[Platon]

Thank you for your response. But if I do the normalization manually and train the NN with the normalized data, how can I test the net with new data? normalize them before simulation? if so how can I de-normalize the output calculated from unseen data? It seems complicated (mapminmax do all the stuff automatically)

[José-Luis]

Yes, provided you use the same normalization bounds (the min and max of both datasets). To rescale, please look at the below code.

bla = 100.*randn(1,10)

minVal = min(bla);
maxVal = max(bla);

norm_data = (bla - minVal) / ( maxVal - minVal )
your_original_data = minVal + norm_data.*(maxVal - minVal)

[Aviral Petwal]

No need to denormalize the data. For your Test set also you can normalize the data with the same parameters and feed it to NN. If you trained on Normalised data just normalize your test set using same parameters and feed the data to NN.

Answer 2

[Jurgen]

NDATA = mat2gray(DATA);

[JohnDylon]

Is this exactly a min-max normalizaton?

[Greg Heath]

Why not just try it and find out?

close all, clear all, clc

 [ x1 , t1 ] = simplefit_dataset;  
 DATA1 = [ x1, t1 ];
 DATA2 = [ x1; t1 ];
 whos DATA1 DATA2

 minmax1 = minmax(DATA1)
 minmax2 = minmax(DATA2)

 minmaxMTG1 = minmax( mat2gray(DATA1) )
 minmaxMTG2 = minmax( mat2gray(DATA2) )

Hope this helps.

Greg

Answer 3

[Abhijit Bhattacharjee]

As of MATLAB R2018a, there is an easy one-liner command that can do this for you. It's called NORMALIZE.

Here is an example, where a denotes the vector of data:

a_normalized = normalize(a, 'range');

Answer 4

[Greg Heath]

I like to calculate min, mean, std and max to detect outliers with standardized data (zero mean/unit variance). For normalization and denormalization I just let the training function use defaults

 tansig and linear

however, if the ouput is naturally bounded use

tansig and tansig

or

 tansig and logsig

In short, unless you are plotting you don't have to worry about anything except outliers.

Hope this helps.

Greg

Answer 5

[Angus Steele]

function [ newValue ] = math_scale_values( originalValue, minOriginalRange, maxOriginalRange, minNewRange, maxNewRange )

%   MATH_SCALE_VALUES
%   Converts a value from one range into another
%       (maxNewRange - minNewRange)(originalValue - minOriginalRange)
%    y = ----------------------------------------------------------- + minNewRange      
%               (maxOriginalRange - minOriginalRange)

newValue = minNewRange + (((maxNewRange - minNewRange) * (originalValue - minOriginalRange))/(maxOriginalRange - minOriginalRange));

end