Add files via upload

Main Classify All - DTW-based KNN.m

@@ -0,0 +1,213 @@
+%% Author: Konstantinos Gyftodimos
+
+%% 1. KNN actual calculation - Weight invariants to normalize unit difference + Compare Trajectories not points 
+
+clear;clc;close all;
+
+addpath('..\motion_classification\KNN_Classification_Datasets')
+addpath('..\motion_classification\classification_code')
+
+%Choose over descriptor types calculated with analytical formulas:
+%load('dimles_FS_200.mat'); %81.18 vs 82.12%(DTW)
+%load('dimles_ISA_200.mat'); %74.67 vs 72.12%(DTW) <<<<<<<<
+
+% load('geom_FS_200.mat'); %57.7 vs 50.14% (DTW) <<<<<<<<
+%load('geom_ISA_200.mat'); %64.6 vs 70.01%(DTW)
+
+%load('timebased_FS_200.mat'); %53.10 vs 48.73%(DTW) <<<<<<<<
+load('timebased_ISA_200.mat'); %50.15 vs 56.395%(DTW)
+
+bool_descriptortype_FS=1;
+bool_descriptortype_ISA=0;
+
+
+%% Initialization
+
+%Normalization weights w.r.t max abs value of each invariant
+w1=1/max(max(abs(X_test(:,1,:))));
+w2=1/max(max(abs(X_test(:,2,:))));
+w3=1/max(max(abs(X_test(:,3,:))));  
+w4=1/max(max(abs(X_test(:,4,:))));
+w5=1/max(max(abs(X_test(:,5,:))));
+w6=1/max(max(abs(X_test(:,6,:)))); 
+w=[w1;w2;w3;w4;w5;w6];
+
+%Multiplication of test and train data with corresponding weights
+for weight_index=1:size(w,1)
+    X_train(:,weight_index,:)=w(weight_index)*X_train(:,weight_index,:);
+    X_test(:,weight_index,:)=w(weight_index)*X_test(:,weight_index,:);
+end
+
+
+% if bool_descriptortype_FS 
+%     X_train(:,[3],:)=[]; 
+%     X_train(:,[5],:)=[];
+%     X_test(:,[3],:)=[];
+%     X_test(:,[5],:)=[];
+%     k=4;
+% elseif bool_descriptortype_ISA
+%     X_train(:,[6],:)=[];
+%     X_test(:,[6],:)=[];
+%     X_train(:,[4],:)=[];
+%     X_test(:,[4],:)=[];
+%     X_train(:,[3],:)=[];
+%     X_test(:,[3],:)=[];
+%     k=3;
+% end
+%% Removal of invariants
+
+%Use this for dimensionless FS invariants and change neighbors to k=4
+%Result = 81.18%
+
+% X_train(:,[3],:)=[]; 
+% X_train(:,[5],:)=[];
+% X_test(:,[3],:)=[];
+% X_test(:,[5],:)=[];
+
+
+%Use this for dimensionless ISA invariants and change neighbors k=3
+%Result =74.64%
+
+% X_train(:,[6],:)=[];
+% X_test(:,[6],:)=[];
+% X_train(:,[4],:)=[];
+% X_test(:,[4],:)=[];
+% X_train(:,[3],:)=[];
+% X_test(:,[3],:)=[];
+
+%Use this for geometric FS invariants and change neighbors to k=5
+%Result =57.7%
+
+% X_train(:,[3],:)=[];
+% X_test(:,[3],:)=[];  
+
+%Use this for geometric ISA invariants and change neighbors to k=1
+%Result = 64.60%
+
+% X_train(:,[6],:)=[];
+% X_test(:,[6],:)=[];
+% X_train(:,[4],:)=[];
+% X_test(:,[4],:)=[];
+% X_train(:,[3],:)=[];
+% X_test(:,[3],:)=[];
+
+%Use this for timebased FS invariants and change neighbors to k=2
+%Result = %53.10
+% X_train(:,[3],:)=[];
+% X_test(:,[3],:)=[];
+% X_train(:,[2],:)=[];
+% X_test(:,[2],:)=[];
+
+%Use this for timebased FS invariants and change neighbors to k=5
+%Result = %50.15
+X_train(:,[6],:)=[];
+X_test(:,[6],:)=[];
+X_train(:,[4],:)=[];
+X_test(:,[4],:)=[];
+X_train(:,[3],:)=[];
+X_test(:,[3],:)=[];
+
+%% Actual Calculation
+
+ed = zeros(size(X_test,1),size(X_train,1),size(X_train,2)); 
+
+%For DTW metric, uncomment next line to turn all NaN values of X_test to 0
+X_test(isnan(X_test))=0; 
+
+%Widthband for DTW function
+widthband = round(0.15*size(X_test,3));
+
+parfor test_point=1:size(X_test,1)
+    disp(['Calculating distances for trial ' num2str(test_point) '/' num2str(size(X_test,1))]);
+    ed_single_test = zeros(size(X_train,1),size(X_train,2)); % this is the [100x6] distance matrix for one test point vs all training points for each invariant
+    for train_point=1:size(X_train,1)    
+        ed_singletest_singletrain = zeros(1,size(X_train,2)); % this is the [1x6] distance matrix for one test point vs one training point for each invariant 
+        for invariant=1:size(X_train,2)
+
+            % Compare two [150x1] time series of invariants!
+            time_series1 = squeeze(X_test(test_point,invariant,:)); % test data, compress [1 x 1 x 150] into [150x1]
+            time_series2 = squeeze(X_train(train_point,invariant,:));  % train data, compress [1 x 1 x 150] into [150x1]
+            ed_singletest_singletrain(invariant)=dtw(time_series1,time_series2,'euclidean'); 
+        end
+        ed_single_test(train_point,:) = ed_singletest_singletrain;
+    end
+    ed(test_point,:,:) = ed_single_test;
+end
+
+ed_new = sum(ed,3); % sum of distances over 3rd dimension 
+[ed_final,ind] = sort(ed_new,2); %put all elements in ed in increasing order in each row
+
+%% Find the nearest k for each data point of the testing data
+
+k=5; 
+k_nn=ind(:,1:k);
+t_labels = zeros(size(k_nn));
+
+%transform k_nn to t_labels, where instead of the number of column as each
+%value, the class that corresponds to this column is inserted.
+
+for i=1:size(k_nn,1)
+    for j=1:size(k_nn,2)
+        if k_nn(i,j)>=1 && k_nn(i,j)<=10
+            t_labels(i,j)=1;
+        elseif k_nn(i,j)>=11 && k_nn(i,j)<=20
+            t_labels(i,j)=2;
+        elseif k_nn(i,j)>=21 && k_nn(i,j)<=30
+            t_labels(i,j)=3;
+        elseif k_nn(i,j)>=31 && k_nn(i,j)<=40
+            t_labels(i,j)=4;
+        elseif k_nn(i,j)>=41 && k_nn(i,j)<=50
+            t_labels(i,j)=5;
+        elseif k_nn(i,j)>=51 && k_nn(i,j)<=60
+            t_labels(i,j)=6;
+        elseif k_nn(i,j)>=61 && k_nn(i,j)<=70
+            t_labels(i,j)=7;
+        elseif k_nn(i,j)>=71 && k_nn(i,j)<=80
+            t_labels(i,j)=8;
+        elseif k_nn(i,j)>=81 && k_nn(i,j)<=90
+            t_labels(i,j)=9;
+        elseif k_nn(i,j)>=91 && k_nn(i,j)<=100
+            t_labels(i,j)=10;
+
+        end
+    end
+end
+
+% Pick the class that appears the most inside t_labels, depending on k
+
+t_labels_final=mode(t_labels,2); 
+
+%it chooses the lowest value when same class appears, do it without mode.
+
+%calculate the classification accuracy
+actual_labels=zeros(size(X_test,1),1);
+
+actual_labels(1:109,1)=1;
+actual_labels(110:207,1)=2;
+actual_labels(208:296,1)=3;
+actual_labels(297:399,1)=4;
+actual_labels(400:486,1)=5;
+actual_labels(487:596,1)=6;
+actual_labels(597:690,1)=7;
+actual_labels(691:783,1)=8;
+actual_labels(784:883,1)=9;
+actual_labels(884:983,1)=10;
+
+%Accuracy
+accuracy=length(find(actual_labels==t_labels_final))/size(X_test,1);
+disp('Overall Classification Accuracy:')
+disp(accuracy);
+
+%Confusion Matrix
+C=confusionmat(actual_labels,t_labels_final);
+
+for i=1:size(C,1)
+C(i,:) = C(i,:)./sum(C(i,:));
+end
+
+disp('Confustion Matrix:');
+disp(C);
+
+
+
+