Difference between fitrnet and fitnet

When I first learn to use neural networks, I usually used nnstart toolbox, which uses fitnet. However, when I started actually to build my own neural networks with more layers and optimal hyperparameters, I found that R2021a introduced fitrnet that looks more powerful. I was wondering what is the difference between these two neural network functions.

Answer 1

[David Willingham]

Hi Alex,

In R2021a, SMLT introduced the fitting functions fitcnet and fitrnet, which train shallow classification and regression neural network models. If you'd like to customize the network, there is a newer framework for building shallow and deep neural networks. Here are 2 examples of how to do this:

Create Simple Deep Learning Network for Classification

Train Convolutional Neural Network for Regression

David

[Alex Liu]

Thank you so much, David!

[Alessandro Niccolai]

Dear David,

While working on Neural Network, I've compared the results of the two functions "fitrnet" and "fitnet". The results I've obtained shows that the "fitnet" function works much better both in terms of computational time and of performance.

Even usign the hyperparameter optimization of "fitrnet", the results are much less than the "fitnet". Here you can find the results I've obtained:

Neural Network - fitrnet

Elapsed time is 105.777746 seconds.

R2 = 0.99957

MSE = 0.028211%

Neural Network - fitnet

Elapsed time is 23.113578 seconds.

R2 = 0.99997

MSE = 0.00082949%

In both the cases, I've computed the performance on a different dataset.

Have you experienced something similar?

Attached you can find the input I've used and below the code that performs the comparison:

clc
clear 
close all

load('trainingTest');

%% Train fitrnet

disp('Neural Network - fitrnet')

disp('-- Traninig --')
tic
Mdl = fitrnet(...
    x, t, ...
    'LayerSizes', [ 86  23 294], ...
    'Activations', 'tanh', ...
    'Lambda', 0.022295, ...
    'IterationLimit', 2000, ...
    'Standardize', true);
toc

disp('-- Running --')
yTest = predict(Mdl,xTest);

% R2

R2 = regress(yTest,tTest);
disp([' R2 = ',num2str(R2)])

% MSE

MSE = mean((tTest-yTest).^2)/mean(yTest)^2*100;
disp([' MSE = ',num2str(MSE),'%'])


%% Train fitnet

disp('Neural Network - fitnet')

hiddenLayers = [10 10];
nLayers = length(hiddenLayers)+1;
net = fitnet(hiddenLayers);

net.trainFcn = 'trainbr';
net.trainParam.epochs = 1000;

net.trainParam.showWindow = false;
net.trainParam.showCommandLine = false;
net.trainParam.show = 10;
net.trainParam.max_fail = 3;
net.trainParam.goal = 1e-9;


net.divideFcn = 'dividerand';
net.divideParam.testRatio = 0.15;
net.divideParam.valRatio = 0.15;

tic
[net,tr] = train(net,x',t');
toc

yTest = net(xTest')';

% R2

R2 = regress(yTest,tTest);
disp([' R2 = ',num2str(R2)])

% MSE

MSE = mean((tTest-yTest).^2)/mean(yTest)^2*100;
disp([' MSE = ',num2str(MSE),'%'])

[David Willingham]

Hi Allesandro,

The two functions have different solver implementations under the hood. Also they have different number of layers and layers sizes.

My recommendation is that you can always try both and see which method works best for the given application.