Trying to get 80% and greater accuracy from network. Can someone help in editing my code to reach to 80% or close too?

2 views (last 30 days)
Nathaniel Porter
Nathaniel Porter on 16 Dec 2021
Commented: yanqi liu on 17 Dec 2021
clc; clear all; close all;
load generated_data.mat
% 2289*180
% 6 classes
X1_T = X1';
rand('seed', 0)
ind = randperm(size(X1_T, 1));
X1_T = X1_T(ind, :);
Y1 = categorical(Y1(ind));
% Split Data
X1_train = X1_T;
train_X1 = X1_train(1:120,:);
train_Y1 = Y1(1:120);
% Data Batch
XTrain=(reshape(train_X1', [2289,120]));
val_X1 = X1_train(121:150,:);
val_Y1 = Y1(121:150);
XVal=(reshape(val_X1', [2289,30]));
test_X1 = X1_train(151:180,:);
test_Y1 = Y1(151:180);
XTest=(reshape(test_X1', [2289,30]));
numFeatures = size(X1_T,2);
% number of hidden units represent the size of the data
numHiddenUnits = 180;
%number of classes represent different patients normal,LIS,type2....
numClasses = length(categories(categorical(Y1)));
layers = [ ...
sequenceInputLayer(numFeatures)
dropoutLayer(0.1)
bilstmLayer(numHiddenUnits,'OutputMode','sequence')
fullyConnectedLayer(numClasses)
instanceNormalizationLayer
softmaxLayer
classificationLayer];
options = trainingOptions('sgdm', ...
'MaxEpochs',150, ...
'GradientThreshold',1, ...
'Verbose',false, ...
'ValidationData',{XVal, val_Y1},...
'LearnRateDropFactor',0.2,...
'LearnRateDropPeriod',5,...
'Plots','training-progress');
% Train
net = trainNetwork(XTrain,train_Y1,layers,options);
% Test
miniBatchSize = 27;
YPred = classify(net,XTest, ...
'MiniBatchSize',miniBatchSize,...
'ExecutionEnvironment', 'cpu');
acc = mean(YPred(:) == categorical(test_Y1(:)))
figure
t = confusionchart(categorical(test_Y1(:)),YPred(:));
  2 Comments
Nathaniel Porter
Nathaniel Porter on 16 Dec 2021
My validation accuracy at highest(without the InstanceNormalizationLayer) is 67.77% but I am currently trying to improve it to roughly 80%. Just asking if my code be manipulated currently to achive this number ?

Sign in to comment.

Accepted Answer

yanqi liu
yanqi liu on 16 Dec 2021
Edited: yanqi liu on 16 Dec 2021
yes,sir,may be use
clc; clear all; close all;
load generated_data.mat
% 2289*180
% 6 classes
rand('seed', 0)
X1_T = X1';
YC1 = categorical(Y1);
CS = categories(YC1);
train_index = []; val_index = []; test_index = [];
for i = 1 : length(CS)
indi = find(YC1==CS{i});
% Shuffling data
indi = indi(randperm(length(indi)));
% 2/3---train, 1/6---val, 1/6---test
index1 = round(length(indi)*2/3);
index2 = round(length(indi)*(2/3+1/6));
train_index = [train_index indi(1:index1)];
val_index = [val_index indi(1+index1:index2)];
test_index = [test_index indi(1+index2:end)];
end
ind = [train_index val_index test_index];
X1_T = X1_T(ind, :);
Y1 = categorical(Y1(ind));
% Split Data
X1_train = X1_T;
train_X1 = X1_train(1:120,:);
train_Y1 = Y1(1:120);
% Data Batch
XTrain=(reshape(train_X1', [2289,120]));
val_X1 = X1_train(121:150,:);
val_Y1 = Y1(121:150);
XVal=(reshape(val_X1', [2289,30]));
test_X1 = X1_train(151:180,:);
test_Y1 = Y1(151:180);
XTest=(reshape(test_X1', [2289,30]));
numFeatures = size(X1_T,2);
% number of hidden units represent the size of the data
numHiddenUnits = 500;
%number of classes represent different patients normal,LIS,type2....
numClasses = length(categories(categorical(Y1)));
layers = [ ...
sequenceInputLayer(numFeatures)
%dropoutLayer(0.5)
instanceNormalizationLayer
bilstmLayer(numHiddenUnits,'OutputMode','sequence')
%dropoutLayer(0.5)
instanceNormalizationLayer
bilstmLayer(round(numHiddenUnits/2),'OutputMode','sequence')
fullyConnectedLayer(numClasses)
instanceNormalizationLayer
softmaxLayer
classificationLayer];
options = trainingOptions('sgdm', ...
'MaxEpochs',100, ...
'GradientThreshold',1, ...
'Verbose',false, ...
'ValidationData',{XVal, val_Y1},...
'LearnRateDropFactor',0.2,...
'LearnRateDropPeriod',5,...
'Plots','training-progress');
% Train
net = trainNetwork(XTrain,train_Y1,layers,options);
% Test
miniBatchSize = 27;
YPred = classify(net,XTest, ...
'MiniBatchSize',miniBatchSize,...
'ExecutionEnvironment', 'cpu');
acc = mean(YPred(:) == categorical(test_Y1(:)))
figure
t = confusionchart(categorical(test_Y1(:)),YPred(:));
acc =
0.9667
>>
  2 Comments
yanqi liu
yanqi liu on 17 Dec 2021
yes,sir,we know it is 6 classes,so just for every class,we choose 2/3、1/6、1/6 as train、val、test data,through this method,we can get the data split index

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