Increasing Validation Accuracy for CNN

Question

tyler seudath 2021-12-18

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1613540-increasing-validation-accuracy-for-cnn

回答： Prince Kumar 2022-4-7

Dear all,

I am trying to increase my validation accuracy since it it too small around 27%. Is it possible you can suggest a method to increase it?

Here is the code. Thanks a mil.

% SPLITTING COMPLEX DATA INTO TRAINING, TESTING AND VALIDATION
%Then extract the Re and Imaginary parts from each dataset
Train_inputX1 =All_data_segments; %5000 segments x 400 samples
Train_outputX1 =Modulation_Schemes ;% 5000 segments x 1
Test_X1 =ceil(0.1 *size(All_data_segments,1)); % 10% is catered for testing
idx1_t= randperm(size(All_data_segments,1),Test_X1);%This  will generate the 10% of the random test numbers from 1 to 5000
% 
% % Now I'm gonna partition the data accordingly...
testX1_input = All_data_segments(idx1_t',:); %Phase test input
testX1_output = Modulation_Schemes(idx1_t); %Phase test output
Train_inputX1(idx1_t',:) = []; % assigning '[]' to matrix entries deletes them.
Train_outputX1(idx1_t',:) =[]; % Deletes the outputs that was used to test input
% 
% %Now we need to repeat to find the validation data
Train1_inputX1 = Train_inputX1; %5000 segments x 400 samples
Train1_outputX1 =Train_outputX1;% 5000 segments x 1
Val_X1 =ceil(0.2 *size(Train_outputX1,1)); % 10% is catered for Validation
idx1_v= randperm(size(Train_outputX1,1),Val_X1);%This  will generate the 10% of the random test numbers from 1 to 5000
% 
 % Now I'm gonna partition the data accordingly...
 ValX1_input =Train1_inputX1(idx1_v',:); %Val input
 ValX1_output =Train_outputX1(idx1_v); % Val output
 Train1_inputX1(idx1_v',:) = []; % assigning '[]' to matrix entries deletes them.
 Train1_outputX1(idx1_v',:) =[]; % Deletes the outputs that was used to validation input
% You have a two dimensional array; call it Data for the moment. The rows correspond to different segments and the columns correspond to different samples, so Data(SegmentNumber, SampleNumber)
%is one piece of information. That information has a real and an imaginary part, so you have real(Data(SegmentNumber, SampleNumber)) and imag(Data(SegmentNumber, SampleNumber)) .
%So you have number of segments times number of samples times two (real + imaginary) pieces of information overall.
% But to store number of segments by number of samples by two pieces of information, 
%you can store that in an array that is (NumberOfSegments by NumberOfSamples by 2) large:
DataParts = zeros(size(Train1_inputX1,1), size(Train1_inputX1,2),1,2); %(4500,400,1,2)
DataParts(:,:,:,1) = real(cell2mat(Train1_inputX1));
DataParts(:,:,:,2) = imag(cell2mat(Train1_inputX1)) ;
XTrain=(reshape(DataParts, [400,1,2,3600])); %Train data
DataParts1 = zeros(size(testX1_input,1), size(testX1_input,2),1, 2);
DataParts1(:,:,:,1) = real(cell2mat(testX1_input));
DataParts1(:,:,:,2) = imag(cell2mat(testX1_input)) ;
Ttrain=(reshape(DataParts1,[400,1,2,500])); %Test data
DataParts2 = zeros(size(ValX1_input,1), size(ValX1_input,2),1, 2);
DataParts2(:,:,:,1) = real(cell2mat(ValX1_input));
DataParts2(:,:,:,2) = imag(cell2mat(ValX1_input));
Vtrain =(reshape(DataParts2,[400,1,2,900])); %450 is the number of segments  %400 is the number of samples 
 Valoutfinal= categorical(ValX1_output); %450 values
 testoutfinal =categorical(testX1_output); %500 values
 Trainoutfinal=categorical(Train1_outputX1);%4050 values
%% NETWORK ARCHITECTURE
layers = [imageInputLayer([400 1 2])  % Creating the image layer
    convolution2dLayer([300 1],3,'Stride',1)
    maxPooling2dLayer(2,'Stride',2,'Padding',[0 0 0 1])
    batchNormalizationLayer
    leakyReluLayer
    fullyConnectedLayer(4)
    softmaxLayer
    classificationLayer];
layers2 = [
    imageInputLayer([400 1 2])
    
    convolution2dLayer([3 1],8,'Padding',[0 0 0 1])
    batchNormalizationLayer
    reluLayer   
    
    maxPooling2dLayer(2,'Stride',2)
    
    convolution2dLayer([3 1],16,'Padding',[0 0 0 1])
    batchNormalizationLayer
    reluLayer   
    
    maxPooling2dLayer(2,'Stride',2)
    
    convolution2dLayer([3 1],32,'Padding',[0 0 0 1])
    batchNormalizationLayer
    reluLayer   
    
    fullyConnectedLayer(4)
    softmaxLayer
    classificationLayer];
% Specify training options.  
opts = trainingOptions('adam', ...
    'MaxEpochs',40, ...
    'Shuffle','every-epoch', ...
    'Plots','training-progress', ...
    'Verbose',false, ...
    'ValidationData',{Vtrain,(Valoutfinal)},...
    'ValidationPatience',Inf);
opts1 = trainingOptions('adam', ...
    'MaxEpochs',100, ...
    'InitialLearnRate',1e-3, ...
     'Plots','training-progress', ...
    'Verbose',false, ...
    'ValidationData',{Vtrain,(Valoutfinal)},...
    'ValidationPatience',Inf);
%% Train network
%analyzeNetwork(layers)
net1 = trainNetwork(XTrain,Trainoutfinal,layers,opts);