目錄
importslpdb函數
——將頭文件、RR文件、註釋文件構成結構體：OutputData = struct('filename', fileName, 'time', anTimeGeneratedCell, ...'rr', rrCollection, 'annotation', anClassGeneratedCell, 'age', age, ...'gender', gender, 'weight', weight);
function OutputData = importslpdb(fileName)
%Import and synchronize a slpdb recording
% Syntax:
% OutputData = importslpdb(fileName)
%
% Input:
% *) fileName - slpdb file name to be imported. Example: 'slp01a'.
% file must be located in 'slpdb' folder,
% three file formats needed: .hea, .rr, and .an
%
% Output:
% *) OutputData - struct contains synchronized RR and annotation
fileName = strcat('slpdb/', cell2mat(fileName));
SEC_PER_EPOCH = 30; % amount of seconds in one epoch (value for slpdb is 30)
OutputData = [];
fprintf('\n%s DATA IMPORT...\n', fileName);
%% IMPORT HEADER DATA 導入頭文件數據
fprintf('Importing header file...\n');
fid = fopen(strcat(fileName, '.hea'), 'r');
headerFile = textscan(fid, '%s%s%s%s%s%s%s%*[^\n]');
fclose(fid);
heaSamplingFreq = strsplit(char(headerFile{3}(1)), '/');
heaSamplingFreq = str2double(cell2mat(heaSamplingFreq(1)));
heaTotalSamples = str2double(cell2mat(headerFile{4}(1)));
heaRecLengthInSec = ceil(heaTotalSamples/heaSamplingFreq);
heaTotalEpoch = ceil(heaRecLengthInSec/SEC_PER_EPOCH);
heaIdx = size(headerFile{1}, 1); % get last line index of file
if cell2mat(headerFile{1}(end-1)) == '#'
% decease index by 1 for 'slp37.hea',
% because the last line is not age, gender, and weight information
heaIdx = heaIdx - 1;
end
age = headerFile{2}(heaIdx);
gender = headerFile{3}(heaIdx);
weight = headerFile{4}(heaIdx);
% output of "IMPORT HEADER DATA" section:
% *) heaTotalEpoch - total epoch according to header data
% *) age - age of the subject
% *) gender - gender of the subject
% *) weight - weight of the subject
% END OF IMPORT HEADER DATA
%% IMPORT ANNOTATION DATA 導入註釋數據
fprintf('Importing annotation file...\n');
fid = fopen(strcat(fileName, '.an'), 'r');
anFile = textscan(fid, '%s%s%s%s%s%s%s%*[^\n]');
fclose(fid);
% remove header of annotation data (first line)
% remove this string: 'Elapsed time Sample # Type Sub Chan Num Aux'
for i=1:size(anFile, 2)
anFile{i}(1) = [];
end
% change first epoch's 'start time' into 0:00.000
anTemp = cell2mat(anFile{1}(1));
anTemp(end-2:end) = 48;
anFile{1}(1) = cellstr(anTemp);
anTime = anFile{1};
anClass = anFile{7};
% output of "IMPORT ANNOTATION DATA" section:
% *) anTime - time from annotation file (cell array)
% *) anClass - annotation (cell array)
% END OF IMPORT ANNOTATION DATA
%% IMPORT RR DATA 導入RR數據
fprintf('Importing RR file...\n');
fid = fopen(strcat(fileName, '.rr'), 'r');
rrFile = textscan(fid, '%s%s%s%s%s%*[^\n]');
fclose(fid);
rrConvertedTime = rrFile{1};
for i=1:size(rrConvertedTime, 1);
rrStartTimeChar = cell2mat(rrConvertedTime(i)); % convert cell into char
rrStartTimeChar(end-2:end) = 48; % xx:xx:xx.aaa -> change 'aaa' part to '000'
% split start time by ":" into matrix 求時長
rrStartTimeMat = strsplit(char(rrConvertedTime(i)), ':')';
% get seconds from the last element
rrSecond = str2double(cell2mat(rrStartTimeMat(end)));
% epoch grouping
rrWhichGroup = floor(rrSecond/SEC_PER_EPOCH)*SEC_PER_EPOCH;
% set epoch grouping
if rrWhichGroup == 0
rrStartTimeChar(end-5) = 48;
elseif rrWhichGroup == 30
rrStartTimeChar(end-5) = 51;
end
rrStartTimeChar(end-4) = 48;
rrConvertedTime(i) = mat2cell(rrStartTimeChar, 1);
end
% change RR value from 'array of cell' into 'array of double' 改變RR值屬性
rrNum = zeros(size(rrFile{3}, 1), 1);
for i=1:size(rrFile{3}, 1)
rrNum(i) = str2double(cell2mat(rrFile{3}(i)));
end
rrTime = rrFile{1};
% output of "IMPORT RR DATA" section:
% *) rrConvertedTime - rounded RR start time according to the epoch
% example: 1:34:31.328 -> 1:34:30.000
% 1:50:13.616 -> 1:50:00.000
% *) rrNum - RR value (array of double)
% *) rrTime - start time of un-rounded RR
% END OF IMPORT RR DATA
%% VALIDITY CHECK 數據有效性檢驗（A.註釋、B.RR數據）
fprintf('Data Validity Check:\n');
% A. Annotation File Check
% *) generate annotation time acocrding to total epoch from header file
% *) result: anTimeGeneratedMat -> Matrix size: number of epoch X 3 (h,m,s)
anTimeGeneratedMat = zeros(heaTotalEpoch, 3);
for i=2:heaTotalEpoch
anTimeGeneratedMat(i, 3) = anTimeGeneratedMat(i-1, 3) + SEC_PER_EPOCH;
anTimeGeneratedMat(i, 2) = anTimeGeneratedMat(i-1, 2);
anTimeGeneratedMat(i, 1) = anTimeGeneratedMat(i-1, 1);
if anTimeGeneratedMat(i, 3) >= 60
anTimeGeneratedMat(i, 3) = 0;
anTimeGeneratedMat(i, 2) = anTimeGeneratedMat(i-1, 2) + 1;
if anTimeGeneratedMat(i, 2) >= 60
anTimeGeneratedMat(i, 2) = 0;
anTimeGeneratedMat(i, 1) = anTimeGeneratedMat(i-1, 1) + 1;
end
end
end
% convert anTimeGeneratedMat into anTimeGeneratedCell for easier comparison
anTimeGeneratedCell = cell(size(anTimeGeneratedMat, 1), 1);
for i = 1:size(anTimeGeneratedMat, 1)
if anTimeGeneratedMat(i, 1) == 0 % when the 'hour' is 0
anTimeGeneratedCell(i) = ...
cellstr(strcat(sprintf('%d',anTimeGeneratedMat(i, 2)), ...
sprintf(':%02d.000',anTimeGeneratedMat(i, 3))));
else
temp = strcat(sprintf('%d', anTimeGeneratedMat(i, 1)), ...
sprintf(':%02d',anTimeGeneratedMat(i, 2)));
anTimeGeneratedCell(i) = ...
cellstr(strcat(temp, sprintf(':%02d.000',anTimeGeneratedMat(i, 3))));
end
end
% *) ANNOTATION FILE CHECK 1 (Total epoch of each data):
fprintf(' CHECK 1: ');
if heaTotalEpoch == size(anTime, 1) && ...
size(unique(rrConvertedTime), 1) == size(anTimeGeneratedMat, 1) && ...
heaTotalEpoch == size(unique(rrConvertedTime), 1)
fprintf('[SUCCESS] heaTotalEpoch (%d) == size(anTime, 1) (%d) == ', ...
'size(unique(rrConvertedTime), 1) (%d) == ', ...
'size(anTimeGeneratedMat, 1) (%d)\n', heaTotalEpoch, size(anTime, 1), ...
size(unique(rrConvertedTime), 1), size(anTimeGeneratedMat, 1));
else
fprintf('[WARNING] heaTotalEpoch (%d) != size(anTime, 1) (%d) != ', ...
'size(unique(rrConvertedTime), 1) (%d) != ', ...
'size(anTimeGeneratedMat, 1) (%d)\n', heaTotalEpoch, size(anTime, 1), ...
size(unique(rrConvertedTime), 1), size(anTimeGeneratedMat, 1));
end
% *) ANNOTATION FILE CHECK 2 (Check equality of anTimeGeneratedCell and anTime):
fprintf(' CHECK 2: ');
if size(anTime, 1) == heaTotalEpoch
for i=1:heaTotalEpoch
if ~strcmp(anTimeGeneratedCell{i}, anTime{i})
fprintf('[FAILED ] anTimeGeneratedCell is NOT EQUAL to anTime\n');
return
end
end
fprintf('[SUCCESS] anTimeGeneratedCell is EQUAL to anTime\n');
else
fprintf('[WARNING] size(anTime, 1) (%d) != heaTotalEpoch (%d), ', ...
'anTimeGeneratedCell will be used\n', size(anTime, 1), heaTotalEpoch);
end
% *) ANNOTATION FILE CHECK 3 (Check annotation value must be '1', '2', '3','4', 'W', 'R', or {'MT', 'M' -> these two will be removed later}):
fprintf(' CHECK 3: ');
distinctClass = char(unique(anClass));
for i=1:size(distinctClass, 1)
if distinctClass(i) ~= '1' && distinctClass(i) ~= '2' ...
&& distinctClass(i) ~= '3' && distinctClass(i) ~= '4' ...
&& distinctClass(i) ~= 'W' && distinctClass(i) ~= 'R' ...
&& distinctClass(i) ~= 'M'
fprintf('[WARNING ] Annotation values is NOT OK\n');
return
end
end
fprintf('[SUCCESS] Annotation values is OK\n');
% B. RR File Check
% *) RR FILE CHECK 1 (Check equality of size(unique(rrConvertedTime), 1)
% and heaTotalEpoch):
fprintf(' CHECK 4: ');
if size(unique(rrConvertedTime), 1) ~= heaTotalEpoch
fprintf('[WARNING] size(unique(rrConvertedTime), 1) (%d) != ', ...
'heaTotalEpoch (%d)\n', size(unique(rrConvertedTime), 1), heaTotalEpoch);
else
fprintf('[SUCCESS] size(unique(rrConvertedTime), 1) (%d) == ', ...
'heaTotalEpoch (%d)\n', size(unique(rrConvertedTime), 1), heaTotalEpoch);
end
% END OF VALIDITY CHECK
%% SYNCHRONIZE RR AND ANNOTATION DATA 同步RR和註釋數據
epochCounter = 1;
rrCounter = 1;
% rrCollection = each row contains RRs of associated epoch
rrCollection = cell(heaTotalEpoch, 1);
% rrTimeCollection = each row contains RR time of associated epoch
rrTimeCollection = cell(heaTotalEpoch, 1);
for i=1:size(rrConvertedTime, 1) % looping for each rrConvertedTime in that file
if strcmp(rrConvertedTime(i), anTimeGeneratedCell(epochCounter))
% when i-th RR time is equal to annotation time of current epoch
rrCollection{epochCounter}(rrCounter) = rrNum(i);
rrTimeCollection{epochCounter}(rrCounter) = rrTime(i);
rrCounter=rrCounter+1;
elseif ~strcmp(rrConvertedTime(i), anTimeGeneratedCell(epochCounter)) ...
&& ~strcmp(rrConvertedTime(i), anTimeGeneratedCell(epochCounter+1))
% when i-th RR time is not equal to annotation time of current epoch
% and i-th RR time is not equal to annotation time of the next epoch
while ~strcmp(rrConvertedTime(i), anTimeGeneratedCell(epochCounter+1))
epochCounter = epochCounter + 1;
end
rrCounter=1;
epochCounter=epochCounter+1;
rrCollection{epochCounter}(rrCounter) = rrNum(i);
rrTimeCollection{epochCounter}(rrCounter) = rrTime(i);
rrCounter=rrCounter+1;
elseif ~strcmp(rrConvertedTime(i), anTimeGeneratedCell(epochCounter)) ...
&& strcmp(rrConvertedTime(i), anTimeGeneratedCell(epochCounter+1))
% when i-th RR time is not equal to annotation time of current epoch
% and i-th RR time is equal to annotation time of the next epoch
rrCounter=1;
epochCounter=epochCounter+1;
rrCollection{epochCounter}(rrCounter) = rrNum(i);
rrTimeCollection{epochCounter}(rrCounter) = rrTime(i);
rrCounter=rrCounter+1;
end
end
% END OF SYNCHRONIZE RR AND ANNOTATION DATA
%% SYNCHRONIZED DATA VALIDITY CHECK 同步數據有效性檢驗
% generate new annotation matrix, fill the time without annotation 生成新的註釋矩陣,在沒有註釋的情況下填充時間
anClassGeneratedCell = cell(heaTotalEpoch, 1);
je = 1;
for i=1:heaTotalEpoch
if strcmp(anTimeGeneratedCell(i), anTime(je))
% if the time is the same, copy the annotation
anClassGeneratedCell(i) = anClass(je);
if je < size(anClass, 1)
je = je + 1;
end
else
% if the time is different, fill with 'none'
anClassGeneratedCell(i) = {'none'};
end
end
fprintf('Removing invalid epoch:\n');
isExists = 0;
for i=heaTotalEpoch:-1:1
flag = 0;
if sum(rrCollection{i}) < 28 || sum(rrCollection{i}) > 32
% set flag to remove incomplete RR data of that epoch by:
% check the sum of RR interval from each epoch,
% can't be below 28 or higher than 32
% (according to slp04 data, min sum is 29 and max is 30)
flag = 1;
fprintf(' Epoch %d (time: %s) of %s data is removed because ', ...
'incomplete RR data\n', i, anTimeGeneratedCell{i}, fileName);
elseif strcmp(anClassGeneratedCell(i), {'none'})
% set flag to remove no annotation epoch
flag = 1;
fprintf(' Epoch %d (time: %s) of %s data is removed because ', ...
'no annotation\n', i, anTimeGeneratedCell{i}, fileName);
elseif strcmp(anClassGeneratedCell(i), {'MT'}) || ...
strcmp(anClassGeneratedCell(i), {'M'})
% set flag to remove 'MT' or 'M' annotation epoch
flag = 1;
fprintf(' Epoch %d (time: %s) of %s data is removed because ', ...
'the annotation is %s\n', i, anTimeGeneratedCell{i}, fileName, ...
anClassGeneratedCell{i});
end
% when the flag is 1, remove the data
if flag == 1
anTimeGeneratedCell{i} = [];
rrCollection{i} = [];
anClassGeneratedCell{i} = [];
isExists = 1;
end
end
% print message if no invalid epoch
if ~isExists
fprintf('No invalid epoch\n');
end
% delete empty row
anTimeGeneratedCell = ...
anTimeGeneratedCell(~cellfun(@isempty, anTimeGeneratedCell));
rrCollection = rrCollection(~cellfun(@isempty, rrCollection));
anClassGeneratedCell = ...
anClassGeneratedCell(~cellfun(@isempty, anClassGeneratedCell));
%END OF SYNCHRONIZED DATA VALIDITY CHECK
%% PREPARE THE OUTPUT 輸出信息
OutputData = struct('filename', fileName, 'time', anTimeGeneratedCell, ...
'rr', rrCollection, 'annotation', anClassGeneratedCell, 'age', age, ...
'gender', gender, 'weight', weight);
% END OF PREPARE THE OUTPUT
end
loadmatobject函數
——導入第index例樣本的字段及對應數據
function Data = loadmatobject(fileName, index)
%Load .mat object by index
% Syntax:
% Data = loadmatobject(dir, index)
%
% Input:
% *) fileName - file name
% *) index - index of .mat's variable to be returned
%
% Output:
% *) Data - index-th variable returned
Data = load(fileName);
fieldName = fieldnames(Data); %返回結構體的字段名稱
Data = Data.(fieldName{index}); %返回數據
end
extractfeatures函數
——沒有輸出變量，輸出非歸一化、歸一化的特徵文件（'xlsx', 'mat'）
function extractfeatures(SlpdbData, destination, outputFormat)
%Extract HRV Features
% Syntax:
% extractfeatures(SlpdbData, destination, outputFormat)
%
% Input:
% *) SlpdbData - struct generated from importslpdb() function % 從importslpdb()函數生成的結構
% *) destination - directory of the result
% *) outputFormat - output format: 'xlsx', 'mat', 'all'
%
% Output:
% No output variables, but there are two files output:
% hrv_features_unorm - unnormalized features
% hrv_features_norm - normalized features
% target - matrix total samples X 6 (1 - 6 classes target)
nSamples = size(SlpdbData, 1);
nClasses = length(unique([SlpdbData.annotation])); % unique：找出數據矩陣中所有不重複數，確定分類類數
hrv = zeros(nSamples, 25);
target = zeros(nSamples, nClasses);
target(:, [1 5]) = NaN;
for i=1:nSamples
rr_diff = diff(SlpdbData(i).rr); % diff函數式用於求導數和差分
hrv(i, 1) = HRVFeature.AVNN(SlpdbData(i).rr);
hrv(i, 2) = HRVFeature.SDNN(SlpdbData(i).rr);
hrv(i, 3) = HRVFeature.RMSSD(rr_diff);
hrv(i, 4) = HRVFeature.SDSD(rr_diff);
hrv(i, 5) = HRVFeature.NNx(50, rr_diff);
hrv(i, 6) = HRVFeature.PNNx(hrv(i, 5), size(SlpdbData(i).rr, 2));
hrv(i, 7) = HRVFeature.HRV_TRIANGULAR_IDX(SlpdbData(i).rr);
hrv(i, 8) = HRVFeature.SD1(hrv(i, 4));
hrv(i, 9) = HRVFeature.SD2(hrv(i, 2), hrv(i, 4));
hrv(i, 10) = HRVFeature.SD1_SD2_RATIO(hrv(i, 8), hrv(i, 9));
hrv(i, 11) = HRVFeature.S(hrv(i, 8), hrv(i, 9));
[TP,pLF,pHF,LFHFratio,VLF,LF,HF,f,Y,NFFT] = ...
HRVFeature.fft_val_fun(SlpdbData(i).rr,2);
hrv(i, 12) = TP;
hrv(i, 13) = pLF;
hrv(i, 14) = pHF;
hrv(i, 15) = LFHFratio;
hrv(i, 16) = VLF;
hrv(i, 17) = LF;
hrv(i, 18) = HF;
% set class annotation
switch SlpdbData(i).annotation
case '1'
target(i,6) = 1; % 標記類別
target(i,4) = 1;
target(i,3) = 1;
target(i,2) = 1;
case '2'
target(i,6) = 2;
target(i,4) = 1;
target(i,3) = 1;
target(i,2) = 1;
case '3'
target(i,6) = 3;
target(i,4) = 2;
target(i,3) = 1;
target(i,2) = 1;
case '4'
target(i,6) = 4;
target(i,4) = 2;
target(i,3) = 1;
target(i,2) = 1;
case 'R'
target(i,6) = 5;
target(i,4) = 3;
target(i,3) = 2;
target(i,2) = 1;
case 'W'
target(i,6) = 6;
target(i,4) = 4;
target(i,3) = 3;
target(i,2) = 2;
otherwise
fprintf('Invalid Annotation');
return
end
end
hrv( :, ~any(hrv,1) ) = [];
% create a new dir if not exists
dirList = dir;
isDirExists = 0;
for i=1:length(dir)
if dirList(i).isdir && strcmp(dirList(i).name, destination)
isDirExists = 1;
end
end
if ~isDirExists
mkdir(destination);
end
% save the data into destination
hrv_features_unorm = hrv;
hrv_features_norm = normalizedata(hrv, -1, 1);
if strcmp(outputFormat, 'xlsx') || strcmp(outputFormat, 'all') % strcmp是用於做字符串比較，保存成相應的特徵文件
xlswrite(strcat(destination, 'hrv_features_unorm.xlsx'), ...
hrv_features_unorm);
xlswrite(strcat(destination, 'hrv_features_norm.xlsx'), ...
hrv_features_norm);
xlswrite(strcat(destination, 'target.xlsx'), target);
end
if strcmp(outputFormat, 'mat') || strcmp(outputFormat, 'all')
save(strcat(destination, 'hrv_features_unorm.mat'), 'hrv_features_unorm');
save(strcat(destination, 'hrv_features_norm.mat'), 'hrv_features_norm');
save(strcat(destination, 'target.mat'), 'target');
end
end
getindexrange函數
——通過index得到輸入向量的範圍
function range = getindexrange(nSamplesEachData, index)
%Get range of inputted vector by index. For example [2 3 4 5] is the
%inputted nSamplesEachData and index is 2. Then, the output is [3 4 5].
%Explanation:
%The sum of [2 3 4 5] is 14 (there are 14 items).
%If index = 1, so the output is [1 2] -> total elements are 2
%If index = 2, so the output is [3 4 5] -> total elements are 3
%If index = 3, so the output is [6 7 8 9] -> total elements are 4
%If index = 4, so the output is [10 11 12 13 14] -> total elements are 5
% Syntax:
% range = getindexrange(nSamplesEachData, index)
%
% Input:
% *) nSamplesEachData - total number of data in each index % 每個index中的數據總數
% *) index - index to be retrieved
%
% Output:
% *) range - a vector contains ordered number of associated index % 包含相關index的有序數量向量
if sum(index > length(nSamplesEachData)) >= 1
disp('Index limit exceeded');
return
end
range = [];
for i=1:length(index)
if index(i) == 1
startNum = 1;
endNum = nSamplesEachData(index(i));
else
startNum = sum(nSamplesEachData(1:index(i)-1)) + 1;
endNum = startNum + nSamplesEachData(index(i)) - 1;
end
range = [range startNum:endNum];
end
end
PSOforELM函數
——PSO的初始化、結合ELM的PSO迭代
function [result, startTime, endTime] = PSOforELM(nFeatures, trainingData, ...
testingData, PSOSettings)
%Running PSO with ELM for feature selection and number of hidden nodes % 使用ELM運行PSO來選擇特徵和隱藏節點的數量
%optimization
% Syntax:
% [result, startTime, endTime] = PSOforELM(nFeatures, trainingData, ...
% testingData, PSOSettings)
%
% Input:
% *) nFeatures - total number of features to be selected
% *) trainingData - training data (Matrix size: total samples X nFeatures)
% *) testingData - testing data (Matrix size: total samples X nFeatures)
% *) PSOSettings - struct contains PSO parameters, examples:
% PSOSettings.MAX_ITERATION = MAX_ITERATION;
% PSOSettings.nParticles = 20;
% PSOSettings.W = 0.6;
% PSOSettings.c1 = 1.2;
% PSOSettings.c2 = 1.2;
% PSOSettings.Wa = 0.95;
% PSOSettings.Wf = 0.05;
%
% Output:
% *) result - struct contains records of PSO ELM result % 此結構包含PSO ELM結果的記錄
% *) startTime - time when the experiment starts
% *) endTime - time when the experiment ends
startTime = clock;
%% PSO PARAMETER PREPARATION PSO各參數定義及初始化
% max total bits for hidden nodes
nHiddenBits = length(dectobin(size(trainingData, 1))); % dectobin十-二進制轉換
populationPosition = rand(PSOSettings.nParticles, nFeatures+nHiddenBits) > 0.5; % 隨機矩陣中的每一個數與0.5比較，若值小於0.5，populationPosition矩陣中相對應的值返回1，否則返回0
% 不符合條件的重新更新
for i=1:PSOSettings.nParticles
while bintodec(populationPosition(i, nFeatures+1:end)) < nFeatures || ...
bintodec(populationPosition(i, nFeatures+1:end)) > size(trainingData, 1) || ...
sum(populationPosition(i, 1:nFeatures)) == 0
populationPosition(i, :) = rand(1, nFeatures+nHiddenBits) > 0.5;
end
end
populationVelocity = int64(zeros(PSOSettings.nParticles, 1)); % 初始化爲0（十進制）
% 定義pBest結構體的字段：position、fitness、trainingAccuracy、testingAccuracy
pBest(PSOSettings.nParticles).position = [];
pBest(PSOSettings.nParticles).fitness = [];
pBest(PSOSettings.nParticles).trainingAccuracy = [];
pBest(PSOSettings.nParticles).testingAccuracy = [];
% 各字段初始化
for i=1:PSOSettings.nParticles
pBest(i).position = false(1, nFeatures+nHiddenBits);
% max fitness value
pBest(i).fitness = repmat(-1000000, PSOSettings.nParticles, 1);
pBest(i).trainingAccuracy = 0;
pBest(i).testingAccuracy = 0;
end
% 定義gBest結構體的字段：position、fitness、trainingAccuracy、testingAccuracy、fromIteration、fromParticle
gBest.position = false(1, nFeatures+nHiddenBits);
gBest.fitness = -1000000; % max fitness value all particle all iteration
gBest.trainingAccuracy = [];
gBest.testingAccuracy = [];
gBest.fromIteration = [];
gBest.fromParticle = [];
% 定義result結構體的字段
result(PSOSettings.MAX_ITERATION+1).iteration = [];
result(PSOSettings.MAX_ITERATION+1).populationPosition = [];
result(PSOSettings.MAX_ITERATION+1).pBest = [];
result(PSOSettings.MAX_ITERATION+1).time = [];
result(PSOSettings.MAX_ITERATION+1).trainingAccuracy = [];
result(PSOSettings.MAX_ITERATION+1).testingAccuracy = [];
result(PSOSettings.MAX_ITERATION+1).model = [];
result(PSOSettings.MAX_ITERATION+1).gBest = [];
% END OF PSO PARAMETER PREPARATION
%% INITIALIZATION STEP 初始化步驟
%fitness function evaluation 適應度函數的評價
[modelArr, trainAccArr, testAccArr, timeArr, populationFitness, pBest] = ...
evaluatefitness(PSOSettings, nFeatures, trainingData, testingData, ...
populationPosition, pBest); % evaluatefitness函數具體見本節最後
gBest = gbestupdate(nFeatures, trainAccArr, testAccArr, populationFitness, ...
populationPosition, gBest, 0); % gbestupdate函數具體見本節最後
% save initial data 保存原始數據
result(1).iteration = 0;
result(1).populationPosition = populationPosition;
result(1).pBest = pBest;
result(1).time = timeArr;
result(1).trainingAccuracy = trainAccArr;
result(1).testingAccuracy = testAccArr;
%result(1).model = modelArr;
result(1).gBest = gBest;
% END OF INITIALIZATION STEP
%% PSO ITERATION PSO的迭代
for iteration=1:PSOSettings.MAX_ITERATION
% Update Velocity
r1 = rand();
r2 = rand();
for i=1:PSOSettings.nParticles
% calculate velocity value 計算速度值
positionDec = int64(bintodec(populationPosition(i, :)));
populationVelocity(i, 1) = PSOSettings.W * populationVelocity(i, 1) + ...
PSOSettings.c1 * r1 * (bintodec(pBest(i).position) - positionDec) ...
+ PSOSettings.c2 * r2 * (bintodec(gBest.position) - positionDec);
% update particle position
newPosDec = abs(int64(positionDec + populationVelocity(i, 1)));
newPosBin = dectobin(newPosDec);
% if the total bits is lower than nFeatures + nHiddenBits,add zeros in front
if size(newPosBin, 2) < (nFeatures + nHiddenBits)
newPosBin = ...
[zeros(1, (nFeatures + nHiddenBits) - size(newPosBin, 2)) ...
newPosBin];
end
% if the number of hidden node is more than the number of samples
if bintodec(newPosBin(1, nFeatures+1:end)) > size(trainingData, 1) ...
|| size(newPosBin(1, nFeatures+1:end), 2) > nHiddenBits
newPosBin = ...
[newPosBin(1, 1:nFeatures) dectobin(size(trainingData, 1))];
end
% if the number of selected features is 0 如果選擇的特徵數量爲0
while sum(newPosBin(1, 1:nFeatures)) == 0
newPosBin(1, 1:nFeatures) = rand(1, nFeatures) > 0.5;
end
% set the new value of position
populationPosition(i, :) = newPosBin;
end
% fitness function evaluation
[modelArr, trainAccArr, testAccArr, timeArr, populationFitness, pBest] = ...
evaluatefitness(PSOSettings, nFeatures, trainingData, testingData, ...
populationPosition, pBest);
gBest = gbestupdate(nFeatures, trainAccArr, testAccArr, ...
populationFitness, populationPosition, gBest, iteration+1);
% save data
result(iteration+1).iteration = iteration;
result(iteration+1).populationPosition = populationPosition;
result(iteration+1).pBest = pBest;
result(iteration+1).time = timeArr;
result(iteration+1).trainingAccuracy = trainAccArr;
result(iteration+1).testingAccuracy = testAccArr;
%result(iteration+1).model = modelArr;
result(iteration+1).gBest = gBest;
end
% END OF PSO ITERATION
endTime = clock;
end
PSOforSVM函數
function [result, startTime, endTime] = PSOforSVM(nFeatures, trainingData, ...
testingData, PSOSettings)
%Running PSO with SVM for feature selection
% Syntax:
% [result, startTime, endTime] = PSOforSVM(nFeatures, trainingData, ...
% testingData, PSOSettings)
%
% Input:
% *) nFeatures - total number of features to be selected
% *) trainingData - training data (Matrix size: total samples X nFeatures)
% *) testingData - testing data (Matrix size: total samples X nFeatures)
% *) PSOSettings - struct contains PSO parameters, examples:
% PSOSettings.MAX_ITERATION = MAX_ITERATION;
% PSOSettings.nParticles = 20;
% PSOSettings.W = 0.6;
% PSOSettings.c1 = 1.2;
% PSOSettings.c2 = 1.2;
% PSOSettings.Wa = 0.95;
% PSOSettings.Wf = 0.05;
%
% Output:
% *) result - struct contains records of PSO SVM result
% *) startTime - time when the experiment starts
% *) endTime - time when the experiment ends
startTime = clock;
%% PSO PARAMETER PREPARATION
populationPosition = rand(PSOSettings.nParticles, nFeatures) > 0.5;
for i=1:PSOSettings.nParticles
while sum(populationPosition(i, 1:nFeatures)) == 0
populationPosition(i, :) = rand(1, nFeatures) > 0.5;
end
end
populationVelocity = int64(zeros(PSOSettings.nParticles, 1)); % in decimal value
% pBest
pBest(PSOSettings.nParticles).position = [];
pBest(PSOSettings.nParticles).fitness = [];
pBest(PSOSettings.nParticles).trainingAccuracy = [];
pBest(PSOSettings.nParticles).testingAccuracy = [];
for i=1:PSOSettings.nParticles
pBest(i).position = false(1, nFeatures);
% max fitness value
pBest(i).fitness = repmat(-1000000, PSOSettings.nParticles, 1);
pBest(i).trainingAccuracy = 0;
pBest(i).testingAccuracy = 0;
end
% gBest
gBest.position = false(1, nFeatures);
gBest.fitness = -1000000; % max fitness value all particle all iteration
gBest.trainingAccuracy = [];
gBest.testingAccuracy = [];
gBest.fromIteration = [];
gBest.fromParticle = [];
% initialize struct data
result(PSOSettings.MAX_ITERATION+1).iteration = [];
result(PSOSettings.MAX_ITERATION+1).populationPosition = [];
result(PSOSettings.MAX_ITERATION+1).pBest = [];
result(PSOSettings.MAX_ITERATION+1).time = [];
result(PSOSettings.MAX_ITERATION+1).trainingAccuracy = [];
result(PSOSettings.MAX_ITERATION+1).testingAccuracy = [];
result(PSOSettings.MAX_ITERATION+1).gBest = [];
% END OF PSO PARAMETER PREPARATION
%% INITIALIZATION STEP
%fitness function evaluation
[trainAccArr, testAccArr, timeArr, populationFitness, pBest] = ...
evaluatefitness(PSOSettings, nFeatures, trainingData, testingData, ...
populationPosition, pBest);
gBest = gbestupdate(nFeatures, trainAccArr, testAccArr, populationFitness, ...
populationPosition, gBest, 0);
% save initial data
result(1).iteration = 0;
result(1).populationPosition = populationPosition;
result(1).pBest = pBest;
result(1).time = timeArr;
result(1).trainingAccuracy = trainAccArr;
result(1).testingAccuracy = testAccArr;
result(1).gBest = gBest;
% END OF INITIALIZATION STEP
%% PSO ITERATION
for iteration=1:PSOSettings.MAX_ITERATION
% Update Velocity
r1 = rand();
r2 = rand();
for i=1:PSOSettings.nParticles
% calculate velocity value
positionDec = int64(bintodec(populationPosition(i, :)));
populationVelocity(i, 1) = PSOSettings.W * populationVelocity(i, 1) + ...
PSOSettings.c1 * r1 * (bintodec(pBest(i).position) - positionDec) ...
+ PSOSettings.c2 * r2 * (bintodec(gBest.position) - positionDec);
% update particle position
newPosDec = abs(int64(positionDec + populationVelocity(i, 1)));
newPosBin = dectobin(newPosDec);
% if the total bits is lower than nFeatures, add zeros in front
if size(newPosBin, 2) < nFeatures
newPosBin = [zeros(1, (nFeatures)-size(newPosBin, 2)) newPosBin];
end
% if the total bits is higher than nFeatures, get first nFeatures
if size(newPosBin, 2) > nFeatures
newPosBin = newPosBin(1, 1:nFeatures);
end
% if the number of selected features is 0
while sum(newPosBin(1, 1:nFeatures)) == 0
newPosBin(1, 1:nFeatures) = rand(1, nFeatures) > 0.5;
end
% set the new value of position
populationPosition(i, :) = newPosBin;
end
% fitness function evaluation
[trainAccArr, testAccArr, timeArr, populationFitness, pBest] = ...
evaluatefitness(PSOSettings, nFeatures, trainingData, testingData, ...
populationPosition, pBest);
gBest = gbestupdate(nFeatures, trainAccArr, testAccArr, ...
populationFitness, populationPosition, gBest, iteration+1);
% save data
result(iteration+1).iteration = iteration;
result(iteration+1).populationPosition = populationPosition;
result(iteration+1).pBest = pBest;
result(iteration+1).time = timeArr;
result(iteration+1).trainingAccuracy = trainAccArr;
result(iteration+1).testingAccuracy = testAccArr;
result(iteration+1).gBest = gBest;
end
% END OF PSO ITERATION
endTime = clock;
end
extractresults函數
——沒有輸出變量，有Excel輸出文件（每次實驗結果、所有實驗的最好結果）、保存gBest隨迭代變化圖片
function extractresults(resultRootFolder, nFeatures, classNum, nExperiments, ...
nIterations)
%Extract raw results of experiment using PSOELM or PSOSVM method
% Syntax:
% extractresults(resultRootFolder, nFeatures, classNum, nExperiments, ...
% nIterations)
%
% Input:
% *) resultRootFolder - root directory of the results
% *) nFeatures - total features
% *) classNum - number of class in vector -> [2 3 4 6]
% *) nExperiments - total experiments of PSO
% *) nIterations - total iterations of each PSO
%
% Output:
% No output variables, there are excel output files:
% [method]_[filename]_extracted_result - results of each experiment
% [method]_result - best of all experiments
%
% Parameter Example:
% resultRootFolder = 'PSOELM_raw_result';
% nFeatures = 18;
% classNum = [2 3 4 6];
% nExperiments = 25;
% nIterations = 100;
fileNames = {'slp01a' 'slp01b' 'slp02a' 'slp02b' 'slp03' 'slp04' ...
'slp14' 'slp16' 'slp32' 'slp37' 'slp41' 'slp45' 'slp48' ...
'slp59' 'slp60' 'slp61' 'slp66' 'slp67x'};
method = strsplit(resultRootFolder, '_');
method = method{1}; % PSOSVM | PSOELM
nClassClassifiers = length(classNum); % 4
headerEachExp = [];
switch method
case 'PSOELM'
headerEachExp = {'Experiment', 'gBestFitness', 'TrainAcc', ...
'TestAcc', 'ProcessTime(Sec)', 'HiddenNodes', 'SelectedFeatures'};
case 'PSOSVM'
headerEachExp = {'Experiment', 'gBestFitness', 'TrainAcc', ...
'TestAcc', 'ProcessTime(Sec)', 'SelectedFeatures'};
end
headerBestExp = headerEachExp;
headerBestExp{1} = 'RecordingName';
% write header for main excel result (only 1 excel)
for i=1:length(classNum)
xlswrite(sprintf('%s/%s_result.xlsx', resultRootFolder, method), ...
headerBestExp, sprintf('%d classes', classNum(i)));
end
for iFile=1:length(fileNames) % loop for each file
% eachFileFolder example: PSOELM_raw_result/PSOELM_slp01a_raw_result
eachFileFolder = sprintf('%s/%s_%s_raw_result', resultRootFolder, ...
method, fileNames{iFile});
for iClass=1:nClassClassifiers % loop for each class number
matFileName = sprintf('%s/%s_%s_%dclasses_raw_result.mat', ...
eachFileFolder, method, fileNames{iFile}, classNum(iClass));
ExperimentResult = loadmatobject(matFileName, 1);
temp = zeros(nExperiments, length(headerEachExp)-1);
nBits = ...
length(ExperimentResult(4). ...
iterationResult(end).gBest.position) - nFeatures;
gBestParticles = false(nExperiments, nFeatures+nBits);
tempCell = cell(nExperiments, 1);
% get the last gBest result of each experiment
for iExp=1:nExperiments
lastResult = ExperimentResult(iExp).iterationResult(end);
temp(iExp, 1) = iExp;
temp(iExp, 2) = lastResult.gBest.fitness;
temp(iExp, 3) = lastResult.gBest.trainingAccuracy;
temp(iExp, 4) = lastResult.gBest.testingAccuracy;
temp(iExp, 5) = ...
etime(ExperimentResult(iExp).endTime, ...
ExperimentResult(iExp).startTime);
if strcmp(method, 'PSOELM')
temp(iExp, 6) = ...
bintodec(lastResult.gBest.position(nFeatures+1:end));
end
tempCell(iExp, 1) = ...
{bintostringorder(lastResult.gBest.position(1, 1:nFeatures))};
gBestParticles(iExp, :) = lastResult.gBest.position;
end
eachFileExcelPath = sprintf('%s/%s_%s_extracted_result.xlsx', ...
eachFileFolder, method, fileNames{iFile});
xlswrite(eachFileExcelPath, headerEachExp, ...
sprintf('%d classes', classNum(iClass)), 'A1');
xlswrite(eachFileExcelPath, temp, ...
sprintf('%d classes', classNum(iClass)), 'A2');
xlswrite(eachFileExcelPath, tempCell, ...
sprintf('%d classes', classNum(iClass)), ...
sprintf('%s2', getexcelcolumncode(length(headerEachExp))));
% get the best experiment of each classification
bestExpIdx = find(temp(:, 2) == max(temp(:, 2)));
if length(bestExpIdx) > 1
% if have the same gBest fitness value, get the max of testAcc
bestExpIdx = ...
bestExpIdx(temp(bestExpIdx, 4) == max(temp(bestExpIdx, 4)));
if length(bestExpIdx) > 1
% if have the same testAcc, get the max of trainAcc
bestExpIdx = bestExpIdx(temp(bestExpIdx, 3) == ...
max(temp(bestExpIdx, 3)));
if length(bestExpIdx) > 1
% if have the same trainAcc,
% get the min of selected features
bestExpIdx = bestExpIdx( ...
sum(gBestParticles(bestExpIdx, 1:nFeatures), 2) == ...
min(sum(gBestParticles(bestExpIdx, 1:nFeatures), 2)));
if length(bestExpIdx) > 1
% if have the same selected feature,
% check the method used
switch method
case 'PSOELM'
bestExpIdx = ...
bestExpIdx(temp(bestExpIdx, 6) == ...
min(temp(bestExpIdx, 6)));
if length(bestExpIdx) > 1
% if have the same selected feature,
% get the first
bestExpIdx = bestExpIdx(1);
end
case 'PSOSVM'
bestExpIdx = bestExpIdx(1);
end
end
end
end
end
% mark the best index
xlswrite(eachFileExcelPath, {'BEST EXPERIMENT'}, ...
sprintf('%d classes', classNum(iClass)), ...
sprintf('%s%d', ...
getexcelcolumncode(length(headerEachExp)+1), bestExpIdx+1));
%{
% gather gBest fitness of the best experiment
gBest = zeros(nIterations, 1);
for iItr=1:nIterations
gBest(iItr) = ...
ExperimentResult(bestExpIdx). ...
iterationResult(iItr+1).gBest.fitness;
end
% save graphics
f = figure;
plot(1:nIterations, gBest);
ylabel('gBest Fitness'); xlabel('Iteration');
title(sprintf('[%s] Best Experiment of %s (%d classes)', ...
method, fileNames{iFile}, classNum(iClass)));
saveas(f, ...
sprintf('%s/[%s] Best Experiment of %s (%d classes).png', ...
eachFileFolder, method, fileNames{iFile}, classNum(iClass)));
close all;
%}
% save result to main excel
switch method
case 'PSOELM'
xlswrite( ...
sprintf('%s/%s_result.xlsx', resultRootFolder, ...
method), ...
[fileNames(iFile) temp(bestExpIdx, 2) ...
temp(bestExpIdx, 3) temp(bestExpIdx, 4) ...
temp(bestExpIdx, 5) temp(bestExpIdx, 6) ...
tempCell(bestExpIdx)], ...
sprintf('%d classes', classNum(iClass)), ...
sprintf('A%d', iFile+1));
case 'PSOSVM'
xlswrite( ...
sprintf('%s/%s_result.xlsx', resultRootFolder, ...
method), ...
[fileNames(iFile) temp(bestExpIdx, 2) ...
temp(bestExpIdx, 3) temp(bestExpIdx, 4) ...
temp(bestExpIdx, 5) tempCell(bestExpIdx)], ...
sprintf('%d classes', classNum(iClass)), ...
sprintf('A%d', iFile+1));
end
end
end
end
evaluatefitness、gbestupdate函數
function [modelArr, trainAccArr, testAccArr, timeArr, populationFitness, ...
pBest] = evaluatefitness(PSOSettings, nFeatures, trainingData, ...
testingData, populationPosition, pBest)
modelArr(PSOSettings.nParticles).inputWeight = []; % 定義modelArr結構體字段inputWeight、outputWeight
modelArr(PSOSettings.nParticles).outputWeight = [];
trainAccArr = zeros(PSOSettings.nParticles, 1);
testAccArr = zeros(PSOSettings.nParticles, 1);
timeArr = zeros(PSOSettings.nParticles, 1);
populationFitness = zeros(PSOSettings.nParticles, 1);
for i=1:PSOSettings.nParticles
tic;
% TRAINING
maskedTrainingFeature = featuremasking(trainingData, ... % featuremasking函數具體見最後
populationPosition(i, 1:nFeatures)); % remove unselected features
% prepare the target data
% (example: transformation from 4 into [0 0 0 1 0 0])
trainingTarget = full(ind2vec(trainingData(:,end)'))';
[Model, trainAcc] = trainELM(maskedTrainingFeature, trainingTarget, ... % trainELM函數具體見最後
bintodec(populationPosition(i, nFeatures+1:end)));
% TESTING
maskedTestingFeature = featuremasking(testingData, ... % featuremasking函數具體見最後
populationPosition(i, 1:nFeatures)); % remove unselected features
% prepare the target data
% (example: transformation from 4 into [0 0 0 1 0 0])
testingTarget = full(ind2vec(testingData(:,end)'))';
testAcc = testELM(maskedTestingFeature, testingTarget, Model); % testELM函數具體見最後
populationFitness(i, 1) = fitness(PSOSettings.Wa, PSOSettings.Wf, ...
testAcc, populationPosition(i, 1:nFeatures)); % function fitnessValue = fitness(Wa, Wf, acc, featureMask)
% fitnessValue = Wa * acc + Wf * (1 - (sum(featureMask)/length(featureMask)));
% end
% pBest update
ischanged = 0;
% 滿足以下任意一條件，更新變量
if populationFitness(i, 1) > pBest(i).fitness
ischanged = 1;
elseif populationFitness(i, 1) == pBest(i).fitness
if pBest(i).testingAccuracy < testAcc
ischanged = 1;
elseif pBest(i).trainingAccuracy < trainAcc
ischanged = 1;
elseif sum(pBest(i).position(1, 1:nFeatures)) > ...
sum(populationPosition(i, 1:nFeatures))
ischanged = 1;
elseif bintodec(pBest(i).position(1, nFeatures+1:end)) > ...
bintodec(populationPosition(i, nFeatures+1:end))
ischanged = 1;
end
end
if ischanged
pBest(i).fitness = populationFitness(i, 1);
pBest(i).position = populationPosition(i, :);
pBest(i).trainingAccuracy = trainAcc;
pBest(i).testingAccuracy = testAcc;
end
% end of pBest update
modelArr(i) = Model;
timeArr(i) = toc;
trainAccArr(i) = trainAcc;
testAccArr(i) = testAcc;
end
end
function gBest = gbestupdate(nFeatures, trainAccArr, testAccArr, ...
populationFitness, populationPosition, gBest, iteration)
if max(populationFitness) >= gBest.fitness
found = find(populationFitness == max(populationFitness));
if length(found) > 1
% if have the same gBest fitness value, get the max of testAcc 如果具有相同的gBest適應度值，則獲取testAcc的最大值
found = found(testAccArr(found) == max(testAccArr(found)));
if length(found) > 1
% if have the same testAcc, get the max of trainAcc 如果有相同的testAcc，得到trainAcc的最大值
found = found(trainAccArr(found) == max(trainAccArr(found)));
if length(found) > 1
% if have the same trainAcc, get the min of selected features 如果具有相同的trainAcc，則獲取所選特徵的最小值
found = ...
found(sum(populationPosition(found, 1:nFeatures), 2) ...
== min(sum(populationPosition(found, 1:nFeatures), 2)));
if length(found) > 1
% if have the same selected feature,get the min of hidden node
hn = zeros(length(found), 1);
for i=1:length(found)
hn(i, 1) = bintodec(populationPosition(found(i), ...
nFeatures+1:end));
end
found = found(hn == min(hn));
if length(found) > 1
found = found(1);
end
end
end
end
end
gBest.fitness = populationFitness(found);
gBest.position = populationPosition(found, :);
gBest.trainingAccuracy = trainAccArr(found);
gBest.testingAccuracy = testAccArr(found);
gBest.fromIteration = iteration;
gBest.fromParticle = found;
end
end
function [trainAccArr, testAccArr, timeArr, populationFitness, pBest] = ...
evaluatefitness(PSOSettings, nFeatures, trainingData, testingData, ...
populationPosition, pBest)
trainAccArr = zeros(PSOSettings.nParticles, 1);
testAccArr = zeros(PSOSettings.nParticles, 1);
timeArr = zeros(PSOSettings.nParticles, 1);
populationFitness = zeros(PSOSettings.nParticles, 1);
for i=1:PSOSettings.nParticles
tic;
% TRAINING
maskedTrainingFeature = featuremasking(trainingData, ...
populationPosition(i, 1:nFeatures)); % remove unselected features
Model = trainSVM(maskedTrainingFeature, trainingData(:,end), 'RBF');
trainAcc = testSVM(maskedTrainingFeature, trainingData(:,end), Model);
% TESTING
maskedTestingFeature = featuremasking(testingData, ...
populationPosition(i, 1:nFeatures)); % remove unselected features
testAcc = testSVM(maskedTestingFeature, testingData(:,end), Model);
populationFitness(i, 1) = fitness(PSOSettings.Wa, PSOSettings.Wf, ...
testAcc, populationPosition(i, 1:nFeatures));
% pBest update
ischanged = 0;
if populationFitness(i, 1) > pBest(i).fitness
ischanged = 1;
elseif populationFitness(i, 1) == pBest(i).fitness
if pBest(i).testingAccuracy < testAcc
ischanged = 1;
elseif pBest(i).trainingAccuracy < trainAcc
ischanged = 1;
elseif sum(pBest(i).position(1, 1:nFeatures)) > ...
sum(populationPosition(i, 1:nFeatures))
ischanged = 1;
end
end
if ischanged
pBest(i).fitness = populationFitness(i, 1);
pBest(i).position = populationPosition(i, :);
pBest(i).trainingAccuracy = trainAcc;
pBest(i).testingAccuracy = testAcc;
end
% end of pBest update
timeArr(i) = toc;
trainAccArr(i) = trainAcc;
testAccArr(i) = testAcc;
end
end
function gBest = gbestupdate(nFeatures, trainAccArr, testAccArr, ...
populationFitness, populationPosition, gBest, iteration)
if max(populationFitness) >= gBest.fitness
found = find(populationFitness == max(populationFitness));
if length(found) > 1
% if have the same gBest fitness value, get the max of testAcc
found = found(testAccArr(found) == max(testAccArr(found)));
if length(found) > 1
% if have the same testAcc, get the max of trainAcc
found = found(trainAccArr(found) == max(trainAccArr(found)));
if length(found) > 1
% if have the same trainAcc, get the min of selected features
found = ...
found(sum(populationPosition(found, 1:nFeatures), 2) ...
== min(sum(populationPosition(found, 1:nFeatures), 2)));
if length(found) > 1
% if have the same selected feature, get the first
found = found(1);
end
end
end
end
gBest.fitness = populationFitness(found);
gBest.position = populationPosition(found, :);
gBest.trainingAccuracy = trainAccArr(found);
gBest.testingAccuracy = testAccArr(found);
gBest.fromIteration = iteration;
gBest.fromParticle = found;
end
end
featuremasking、trainELM、testELM函數
function maskedFeature = featuremasking(feature, mask)
%Retrieve masked features
% Syntax:
% maskedFeature = featuremasking(feature, mask)
%
% Input:
% *) feature - feature collection
% (Matrix size: total samples X total features)
% *) mask - logical matrix, 1 means selected, 0 is not selected
% (Matrix size: 1 X total features)
% Output:
% *) maskedFeature - matrix with selected features only
% (Matrix size: total samples X total selected features
maskedFeature = zeros(size(feature, 1), sum(mask));
j = 1;
for i=1:sum(mask)
if mask(1, i) == 1
maskedFeature(:, j) = feature(:, i);
j = j + 1;
end
end
end
function [ELMModel, trainAcc] = trainELM(feature, target, nHiddenNode)
%Train Extreme Learning Machine (ELM) model
% Syntax:
% [ELMModel, trainAcc] = trainELM(feature, target, nHiddenNode)
%
% Input:
% *) feature - feature collection
% (Matrix size: total samples X total features)
% *) target - target of each sample
% (Matrix size: total samples X total classes)
% Example: class 4 -> target is [0 0 0 1]
% *) nHiddenNode - total hidden nodes of ELM
%
% Output:
% *) ELMModel.inputWeight - input weight of ELM
% (Matrix size: nHiddenNode (+1 for bias) X total features)
% *) ELMModel.outputWeight - output weight of ELM
% (Matrix size: total classes X nHiddenNode)
% *) trainAcc - training accuracy
if size(feature, 2) == 0
fprintf('Someting went wrong, no feature selected.');
return
end
% STEP 1: RANDOMLY ASSIGN INPUT WEIGHT AND BIAS 隨機分配輸入權重和偏差
minWeight = -1;
maxWeight = 1;
inputWeight = (maxWeight-minWeight) .* ...
rand(nHiddenNode, size(feature, 2)+1) + minWeight;
% STEP 2: CALCULATE THE HIDDEN LAYER OUTPUT MATRIX H 計算隱含層輸出矩陣H
% linear combination of hidden output 線性組合的隱藏輸出
hiddenOutput = (inputWeight(:, 1:end-1) * feature')+ ...
repmat(inputWeight(:, end), 1, size(feature, 1));
% apply activation function on hidden output 對隱藏輸出應用激活函數
hiddenOutput = sigmoid(hiddenOutput); % function y = sigmoid(x)
% y = 1./(1 + exp(-1.*x));
% end
% STEP 3: CALCULATE THE OUTPUT WEIGHT B 計算輸出權值B
% estimate output weight
outputWeight = target' * pinv(hiddenOutput); % pinv函數-求矩陣的僞逆矩陣
% STEP 4: APPLY MODEL TO TRAINING DATA 將模型應用訓練數據
% linear combination of predicted output 預測輸出的線性組合
predictedOutput = outputWeight * hiddenOutput;
% apply activation function on predicted output 對預測輸出應用激活函數
predictedOutput = sigmoid(predictedOutput);
maxPred = max(predictedOutput);
predictedClass = zeros(size(predictedOutput, 2), 1);
for i=1:size(predictedOutput, 2)
class = find(predictedOutput(:, i) == maxPred(i));
predictedClass(i) = class(1, 1);
end
trainAcc = sum(predictedClass == vec2ind(target')')/ ... % vec2ind：向量到索引，向量中每一列中的所有元素有且只能有一個爲1，向量的列數即爲索引矩陣的列數
size(predictedOutput, 2) * 100;
ELMModel.inputWeight = inputWeight;
ELMModel.outputWeight = outputWeight;
end
function testAcc = testELM(feature, target, ELMModel)
%Test Extreme Learning Machine (ELM) model
% Syntax:
% testAcc = testELM(feature, target, ELMModel)
%
% Input:
% *) feature - feature collection
% (Matrix size: total samples X total features)
% *) target - target of each sample
% (Matrix size: total samples X total classes)
% Example: class 4 -> target is [0 0 0 1]
% *) ELMModel - ELMModel generated from trainELM() function
%
% Output:
% *) testAcc - testing accuracy
% linear combination of hidden output
hiddenOutput = (ELMModel.inputWeight(:, 1:end-1) * feature')+ ...
repmat(ELMModel.inputWeight(:, end), 1, size(feature, 1));
% apply activation function on hidden output
hiddenOutput = sigmoid(hiddenOutput);
% linear combination of predicted output
predictedOutput = ELMModel.outputWeight * hiddenOutput;
% apply activation function on predicted output
predictedOutput = sigmoid(predictedOutput);
maxPred = max(predictedOutput);
predictedClass = zeros(size(predictedOutput, 2), 1);
for i=1:size(predictedOutput, 2)
class = find(predictedOutput(:, i) == maxPred(i));
predictedClass(i) = class(1, 1);
end
testAcc = sum(predictedClass == vec2ind(target')')/ ...
size(predictedOutput, 2) * 100;
end
