How to Train CNN on Custom dataset in matrix form

Hi everyone, i hope you are doing well

i have the following dataset myFile.txt includes 102x5,in which first 4 coloums are the Number of Observation and the last column are the Discrete labels/Classes for the dataset. I want to train simple CNN on this dataset

How i can randomly split the data into train/validation/ test and reshape the data to train simple CNN

when i run numel(trainlabel) it shows 102 Classes but i have only 5 Classes

I have got this error when i run the below code

Caused by:

Layer 2: Input size mismatch. Size of input to this layer is different from the expected input size.

Inputs to this layer:

from layer 1 (size 102(S) × 4(S) × 1(C) × 1(B))

%myFile.txt includes 102x5,in which first 4 coloums are the
% Number of Observation and the last 5 column are the Discrete labels/Classes for
% the dataset
dataset=readmatrix('myFile.txt');
traindata=dataset(:,1:4);
trainlabel=categorical(dataset(:,5));

% The following show 102 Classes but i have only 5 Classes
numClasses = numel(trainlabel)

%How to Split Data Randomly into Train,Validation and Test

% The below code Split the data 80/20 in Train and Test but not randomly
PD = 0.80 ;  % percentage 80%
Ptrain = traindata(1:round(PD*length(traindata)),:) ; Ttrain = trainlabel(1:round(PD*length(traindata))) ;
Ptest = traindata(round(PD*length(traindata)):end,:) ;Ttest = trainlabel(round(PD*length(traindata)):end) ;


% I have defined the following Network for training,

layers = [ ...
    imageInputLayer([102 4 1])
    convolution2dLayer(5,20)
    reluLayer
    maxPooling2dLayer(2,'Stride',2)
    fullyConnectedLayer(5)
    softmaxLayer
    classificationLayer];

options = trainingOptions('sgdm', ...
    'MaxEpochs',20,...
    'InitialLearnRate',1e-4, ...
    'Verbose',false, ...
    'Plots','training-progress');
net = trainNetwork(traindata,trainlabel,layers,options);

Answer 1

[yanqi liu]

yes，sir，may be reshape your data，such as

sz = size(dataset);
dataset = dataset(randperm(sz(1)),:);
traindata=dataset(:,1:4);
trainlabel=categorical(dataset(:,5));

classes = unique(trainlabel)
numClasses = numel(unique(trainlabel))

PD = 0.80 ;  
Ptrain = []; Ttrain = [];
Ptest = []; Ttest = [];
for i = 1 : length(classes)
    indi = find(trainlabel==classes(i));
    indi = indi(randperm(length(indi)));
    indj = round(length(indi)*PD);
    Ptrain = [Ptrain; traindata(indi(1:indj),:)]; Ttrain = [Ttrain; trainlabel(indi(1:indj),:)]; 
    Ptest = [Ptest; traindata(indi(1+indj:end),:)]; Ttest = [Ttest; trainlabel(indi(1+indj:end),:)]; 
end

Ptrain=(reshape(Ptrain', [4,1,1,size(Ptrain,1)]));
Ptest=(reshape(Ptest', [4,1,1,size(Ptest,1)]));

layers = [imageInputLayer([4 1 1])
    convolution2dLayer([3 1],3,'Stride',1)
    batchNormalizationLayer
    reluLayer
    maxPooling2dLayer(2,'Stride',2,'Padding',[0 0 0 1])
    dropoutLayer
    fullyConnectedLayer(numClasses)
    softmaxLayer
    classificationLayer];

options = trainingOptions('adam', ...
    'MaxEpochs',3000, ...
    'Shuffle','every-epoch', ...
    'Plots','training-progress', ...
    'Verbose',false, ...
    'ValidationData',{Ptest,Ttest},...
    'ExecutionEnvironment', 'cpu', ...
    'ValidationPatience',Inf);

net = trainNetwork(Ptrain,Ttrain,layers,options);

[Med Future]

@yanqi liu Can you please explain why there is loop for Ptrain and Ptest? and why you reshape in this way?

reshape(Ptrain', [4,1,1,size(Ptrain,1)]

[Med Future]

@yanqi liu i think Training is not correct.

How can i you convolution1dLayer if i want to change network

[yanqi liu]

yes，sir

because the input data is 1*4 and output data is 1*1 for classify

so we use [4,1,1,batches] to reshape it

then，just make convolution2dLayer one dimension to 1 to get convolution1dLayer use

for more information，may be read my book：http://product.dangdang.com/28472603.html

[Med Future]

@yanqi the input data has 4 columns thats why it is 1*4?

[Med Future]

@yanqi can you also share with code with convolution1dLayer?

[yanqi liu]

yes，sir，input has 4 dimension，so it process as 1*4

there is no convolution1dLayer，so we use convolution2dLayer and set one dimension as 1 to process

[Med Future]

@yanqi liu did this neural Network perform well?

[yanqi liu]

yes，sir，or use lstm，may be same result

[Med Future]

@yanqi liu Can you please share 1D CNN network for this?

[yanqi liu]

yes，sir，may be make model structure like this

layers = [ ...
    sequenceInputLayer(4)
    lstmLayer(100,'OutputMode','sequence')
    dropoutLayer(0.3)
    lstmLayer(50,'OutputMode','sequence')
    dropoutLayer(0.2)
    fullyConnectedLayer(numClasses)
    softmaxLayer
    classificationLayer];