preparation data for Convolutional Neural Network in regression approach

Question

Abolfazl Nejatian 2017-11-20

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/368043-preparation-data-for-convolutional-neural-network-in-regression-approach

评论： John Albert 2018-1-2

dear my friends Thanks for your guidance.

I've another question. after that, I was disappointed with LSTM Regression Layer in Matlab, i tried to use a Normal CNN for my Time series prediction. in my codes initially, i changed the dimensional of data from [1*1000] to [3*997] (with consideration time delays).

because it should be similar to an image input then changed this 2-dimensional matrix to 4-dimensional matrices, to be a proper Input data for CNN layer. and after that trained a CNN network and then finally check the prediction results with the real targets, the strange point is prediction scales is much less than the real targets, I don't have any idea Why this happened?

Many Thanks Abolfazl Nejatian

Here is my code

    function RegressionCNN()
    clc; clear; close all;
    %%load data
    X = xlsread('forexTest.xlsx');
    NumDelay = 4;
    Delays = setDelays(NumDelay);
    [Inputs, Targets] = CreateTimeSeriesData(X', Delays);
    Inputs  = Inputs';
    Targets = Targets';
    nData   = size(Inputs,1);
    PERM = 1:nData; % Permutation to Shuffle Data
    pTrain       = .8;
    nTrainData   = round(pTrain*nData);
    TrainInd     = PERM(1:nTrainData);
    TrainInputs  = Inputs(TrainInd,:);
    TrainTargets = Targets(TrainInd,:);
    pTest       = 1-pTrain;
    nTestData   = nData-nTrainData;
    TestInd     = PERM(nTrainData+1:end);
    TestInputs  = Inputs(TestInd,:);
    TestTargets = Targets(TestInd,:);
    %%Creat LSTM Layers 
    layers = [ ...
        imageInputLayer([size(Inputs,2) 1],'Name','rinput');
        convolution2dLayer(1,25);
        reluLayer();
        fullyConnectedLayer(1);
        regressionLayer('Name','routput')];
    %%Train LSTM network
    maxEpochs = 200;
    shuffle   = 'never';
    miniBatchSize = 10;
    options = trainingOptions('sgdm', ...
        'MaxEpochs',maxEpochs, ...
        'Shuffle', shuffle,...
        'MiniBatchSize',miniBatchSize,...
        'Verbose',1,...
        'Plots','training-progress',...
        'SequenceLength','longest',...
        'ExecutionEnvironment','cpu');
    Y   = TrainTargets;
    X   = TrainInputs';
    X   = build4DData(X);
    net = trainNetwork(X,Y,layers,options);
    %%Test Network
    XTest       = TestInputs';
    TestTargets = TestTargets;
    XTest       = build4DData(XTest);
    YPred = predict(net,XTest, ...
        'MiniBatchSize',miniBatchSize,...
         'ExecutionEnvironment','cpu');
     %%plot the results
    close all
    format long
    YPred  = double(YPred);
    whos('TestTargets')
    whos('YPred')
    Errors = TestTargets - YPred;
    MSE    = mean(Errors.^2);
    RMSE   = sqrt(MSE);
    ErrorMean = mean(Errors);
    ErrorStd  = std(Errors);
    subplot(2,2,[1 2]);
    plot(TestTargets);
    hold on;
    plot(YPred);
    legend('Targets','Outputs');
    ylabel('Targets and Outputs');
    grid on;
    subplot(2,2,3);
    plot(Errors);
    title(['MSE = ' num2str(MSE) ', RMSE = ' num2str(RMSE)]);
    ylabel('Errors');
    grid on;
    subplot(2,2,4);
    histfit(Errors, 50);
    title(['Error Mean = ' num2str(ErrorMean) ', Error StD = ' num2str(ErrorStd)]);
    function XNew = build4DData(x)
    XNew = zeros(size(x,1),1,1,size(x,2));
    XNew(:,1,1,:) = x(:,:);