Error in the regularization of Neural Networks and understanding 'perform'

4 次查看（过去 30 天）
Harleen Singh Makhija
编辑： Harleen Singh Makhija 2019-1-5
Hello all,
In MATLAB 2017b..
I have a neural network for fitting pupose.
[I N] = size(x); % I = 3, N = 500
[O N] = size (t); % O = 1
Following is my code only for regularization purpose..
clc;
close all; rng(0);
x = input_vel_mass_pos_steps';
t = torque_steps';
trainFcn = 'trainrp';
% Create a Fitting Network
hiddenLayerSize = 9;
net = fitnet(hiddenLayerSize,trainFcn);
%regularization Lambda
regularization_term = [0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0.11 0.12 0.13 0.14 0.15 0.16 0.17 0.18 0.19 0.2 0.21 0.22 0.23 0.24 0.25 0.26 0.27 0.28 0.29 0.3 0.31 0.32 0.33 0.34 0.35 0.36 0.37 0.38 0.39 0.4 0.41 0.42 0.43 0.44 0.45 0.46 0.47 0.48 0.49 0.5 0.51 0.52 0.53 0.54 0.55 0.56 0.57 0.58 0.59 0.6 0.61 0.62 0.63 0.64 0.65 0.66 0.67 0.68 0.69 0.7 0.71 0.72 0.73 0.74 0.75 0.76 0.77 0.78 0.79 0.8 0.81 0.82 0.83 0.84 0.85 0.86 0.87 0.88 0.89 0.9 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98 0.99 1];
mm = size(regularization_term,2);
x_train = x(:,sstrain);t_train = t(:,sstrain); %sstrain - indices got from a different program
x_test = x(:,sstest);t_test = t(:,sstest); %%sstest - indices got from a different program
trainPerformance = zeros(10,mm); %zeros(j,mm)
testPerformance = zeros(10,mm);
for j = 1:10
for i = 1:mm
net = fitnet(hiddenLayerSize,trainFcn);
net.divideFcn = '';
net.trainParam.epochs = 300;
%nett.performFcn='msereg';
%nett.performParam.ratio = regularization_term(i);
net.performParam.regularization = regularization_term(i);
[net trr] = train(net,x_train,t_train);
y_train = nett(x_train);
trainPerformance(j,i) = sqrt(perform(net,t_train,y_train));
%trainPerformance(j,i) = norm((y_train-t_train)/sqrt(length(y_train)));
y_test = net(x_test);
testPerformance(j,i) = sqrt(perform(net,t_test,y_test));
%testPerformance(j,i) = norm((y_test-t_test)/sqrt(length(y_test)));
end
end
plot(regularization_term, mean(trainPerformance),regularization_term,mean(testPerformance))
hold on;
legend('trainperformance-RMSE','testperformacne-RMSE')
xlabel('Regularization Ratio')
ylabel('RMSE')
My questions are as following..
  1. What does the 'perform' function does? As per the theory it takes into consideration
net.performFcn and
net.performParam
net.performFcn = mse by default and By using net.performFcn.regularization = regularization_term(i) why then at *last* (in the command window to check) when I do testPerformance = sqrt(perform(nett,t_test,y_test)) and testPerformance = norm((y_test-t_test)/sqrt(length(y_test))) the results are not equal? Ain't they suppose to be equal?
2. when my input code is net.performFcn.regularization = regularization_term(i) and testperformance = sqrt(perform(nett,t_test,y_test)) I get very weird graph as follows -
How can it be a straight line? More importantly how come test error was less than training error? By this graph, I won't be able to get the optimised lambda value from this. This should have been a curve. The testing should have gone from down to up as a curve and by penalty we should have seen the training error going up..
3. I used in mine(later) from the previous code(which I saw) as net.performFcn = 'msereg' and net.performParam.ratio = regularization(i) used to work for regularization but its not the case in new versions as I checked later (after using 'msereg') net.performFcn is still 'mse'. Question is what is then now net.performParam.ratio? and what's the difference between net.performParam.ratio and net.performParam.regularization? Because though net.performFcn is not 'msereg' my code above ran for net.performParam.ratio = regularization(i). Why did it ran then for net.performParam.ratio in the first place itself? I then used it anyhow to see my results..
4. Below are my cases for plot of RMSE vs regularization(i), all of which are abnormal. Any help understanding where am I going wrong is appreciated and concept of my regularization will be clear..
  • using net.performParam.regularization = regularization_term(i) and testPerformance = norm(y_test-t_test)
  • using net.performParam.ratio = regularization_term(i) and testPerformance = sqrt(perform(nett,t_test,y_test))
  • using net.performParam.ratio = regularization_term(i) and testPerformance = norm(y_test-t_test)

请先登录，再回答此问题。

回答（0 个）

类别

在 Help Center 和 File Exchange 中查找有关 Define Shallow Neural Network Architectures 的更多信息

标签

Community Treasure Hunt

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

Start Hunting!

Translated by