Trying reduce overfitting of training plot

1 view (last 30 days)
Previoulsy tried running network with two sets of data however was not succesful. Achieved progres with running one per dataset however want to know how I can reduce any overfitting even though the validation accuracy seems good.
clc; clear all; close all;
%Import/Upload data
load Projectdata.mat
% change to label vector
CS1 = categories(categorical(INS_output));
Z2 = [];
for i = 1 : length(INS_output)
Z2(i,1) = find(INS_output(i)==CS1);
end
Yo2 = INS_output;
INS_output = Z2;
%transposing insulin data
InsulinReadings_T = InsulinReadings';
rand('seed', 0)
ind = randperm(size(InsulinReadings_T, 1));
InsulinReadings_T = InsulinReadings_T(ind, :);
INS_output = INS_output(ind);
InsulinReadings_train = InsulinReadings_T;
train_InsulinReadings = InsulinReadings_train(1:84,:);
train_INS_output = INS_output(1:84);
InsulinReadingsTrain=(reshape(train_InsulinReadings',[1758,1,1,84]));
val_InsulinReadings = InsulinReadings_train(85:102,:);
val_INS_output = INS_output(85:102);
InsulinReadingsVal=(reshape(val_InsulinReadings', [1758,1,1,18]));
test_InsulinReadings = InsulinReadings_train(103:120,:);
test_INS_output = INS_output(103:120);
InsulinReadingsTest=(reshape(test_InsulinReadings', [1758,1,1,18]));
%% NETWORK ARCHITECTURE
layers = [imageInputLayer([1758 1 1]) % Creating the image layer
convolution2dLayer([102 1],3,'Stride',1)
batchNormalizationLayer
reluLayer
maxPooling2dLayer(2,'Stride',2,'Padding',[0 0 0 1])
dropoutLayer
fullyConnectedLayer(1)
regressionLayer];
% Specify training options.
opts = trainingOptions('sgdm', ...
'MaxEpochs',500, ...
'Shuffle','every-epoch', ...
'Plots','training-progress', ...
'Verbose',false, ...
'ValidationData',{InsulinReadingsVal,val_INS_output,},...
'LearnRateDropFactor',0.2,...
'LearnRateDropPeriod',5,...
'ExecutionEnvironment', 'cpu', ...
'ValidationPatience',Inf);
%% Train network
%net = trainNetwork(XTrain,Trainoutfinal,layers,opts);
yc1 = train_INS_output(:);
net2 = trainNetwork(InsulinReadingsTrain,yc1,layers,opts);
%% Compare against testing Data
INS_outputpredicted = predict(net2, InsulinReadingsTest)
predictionError = test_INS_output - INS_outputpredicted;
squares = predictionError.^2;
rmse = sqrt(mean(squares))
figure
scatter(INS_outputpredicted, test_INS_output,'+')
title ('True value vs Predicted Value')
xlabel ("Predicted Value")
ylabel ("True Value")
hold on
plot([-3 3], [-7 7], 'b--')
  3 Comments
Ive J
Ive J on 5 Mar 2022
Your train and validation sets should be non-overlapping, and if this is the case here (as should be), your model is protected against overfitting simply because train/validation sets never met each other.

Sign in to comment.

Accepted Answer

yanqi liu
yanqi liu on 7 Mar 2022
may be set dropoutLayer(value) to reduce more parameters during training
  1 Comment
Nathaniel Porter
Nathaniel Porter on 7 Mar 2022
Is a 0.5 dropout layer value good or should increase and determine if it helps

Sign in to comment.

More Answers (0)

Products


Release

R2021b

Community Treasure Hunt

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

Start Hunting!

Translated by