Need help with Sequence input Deep Learning toolbox error

I am trying to achieve a script to execute sequence-to-label(response) classification.

I am having a problem with how the function trainNetwork wants the input data packaged.

I am trying to figure out what to do with XTrain or XTrainTest, to get the 'trainNetwork' command to accept the input data.

I am receiving the following error: 'Error using trainNetwork

Invalid training data. Predictors must be a numeric array, a datastore, or a table. For networks with sequence input, predictors can also be a cell array of sequences. '

----------------------------------------------------------------------------------------------------------------------------------------------------------------------

clear all;

close all;

clc

%--------------------------------------------------------------------------

% Generating A matrix

NewDimension = uint16(200);

mu = 5;

std = 2.1;

A = normrnd(mu, std, [NewDimension 784] );

%--------------------------------------------------------------------------

N_sample = 60000;

N_test=10000;

%XTrain = zeros(28,28,1,N_sample);

XTrain = zeros(1,NewDimension,1,N_sample);

XTrainTest = cell(N_sample,1);

YTrain=zeros(N_sample,1);

% Please dowload the MNIST data set from http://yann.lecun.com/exdb/mnist/

% and unzip.

fidimg1=fopen('train-images.idx3-ubyte','rb');

fidimg2=fopen('train-labels.idx1-ubyte','rb');

[img,count]=fread(fidimg1,16); % table head

[imgInd,count1]=fread(fidimg2,8); %table head

for k=1:N_sample

[im,~]=fread(fidimg1,[28,28]);

ind=fread(fidimg2,1);

clear testsave, clear testmult

testsave = reshape(im, 784, 1);

testmult = A*testsave;

XTrain(1,1:NewDimension,1,k)=transpose(testmult);

XTrainTest{k} = testmult';

%XTrainTest{k} = double(transpose(testmult));

%XTrain(:,:,1,k)=im';

YTrain(k)=ind;

end

fclose(fidimg1);

fclose(fidimg2);

YTrain=categorical(YTrain);

%-------------------------------------------------------------------------

XTest = zeros(1,NewDimension,N_test);

%XTest = zeros(1,784,1,N_test);

YTest=zeros(N_test,1);

fidimg1=fopen('t10k-images.idx3-ubyte','rb');

fidimg2=fopen('t10k-labels.idx1-ubyte','rb');

[img,count]=fread(fidimg1,16);

[imgInd,count1]=fread(fidimg2,8);

for k=1:N_test

[im,~]=fread(fidimg1,[28,28]);

ind=fread(fidimg2,1);

clear testsave, clear testmult

testsave = reshape(im, 784, 1);

testmult = A*testsave;

XTest(1,1:NewDimension,k)=transpose(testmult);

%XTest(:,:,1,k)=im';% training set building

YTest(k)=ind;

end

fclose(fidimg1);

fclose(fidimg2);

YTest=categorical(YTest);

numHiddenUnits = 400;

numClasses = 10;

numFeatures = NewDimension;

% Modify XTrain and XTest to be 2-D arrays

XTrain = reshape(XTrain, [NewDimension, N_sample]);

%YTrain = transpose(YTrain);

%XTest = reshape(XTest, [NewDimension, N_test]);

layers = [

featureInputLayer(numFeatures)

fullyConnectedLayer(100) % number of neurons in the first hidden layer

batchNormalizationLayer

reluLayer

fullyConnectedLayer(50) % number of neurons in the second hidden layer

batchNormalizationLayer

reluLayer

fullyConnectedLayer(numClasses)

softmaxLayer

classificationLayer];

maxEpochs = 5;

miniBatchSize = 20;

options = trainingOptions('adam', ...

'ExecutionEnvironment','cpu', ...

'MaxEpochs',maxEpochs, ...

'MiniBatchSize',miniBatchSize, ...

'GradientThreshold',1, ...

'Verbose',false, ...

'Plots','training-progress');

net = trainNetwork(XTrainTest,YTrain,layers,options);

[Eva]

I realized, I should add some more context to my submitted question.

I want to take MNIST digit images, and I want to apply a matrix multiplier to the image, that results in a [200 1] sequence. (the MNIST images are changed into a vector of length 784 digits)

I want to do sequence-to-label classification with these input vectors/sequences. So, this changes the image classification problem to a sequence-to-label classification problem. The image classification problem was easy because imageDatastore would take care of prepping the input images for trainNetwork.

For this sequence-to-label problem, I do not know what datastore routine I could use. I keep on getting these data loading errors from the trainNetwork function. Help!

[Eva]

I think I seemed to discover the fault in this simulation, and it had to do with my layer stack up, it was defined for featureinputLayers and not for sequenceInputLayer:

The following seems to allow it to train:

layers = [

sequenceInputLayer(numFeatures)

convolution1dLayer(filterSize, numFilters, 'Padding', 'causal')

reluLayer

layerNormalizationLayer

convolution1dLayer(filterSize, 2*numFilters, 'Padding', 'causal')

reluLayer

layerNormalizationLayer

globalAveragePooling1dLayer

fullyConnectedLayer(numClasses)

softmaxLayer

classificationLayer

];

[Ben]

In general you should be able to use both featureInputLayer and sequenceInputLayer for this problem, by interpreting the image as either a vector 784 features, or a sequence of 784 features. But if you want to use convolution1dLayer you'll need to use sequenceInputLayer since convolution1dLayer operates over the sequence dimension.

For the datastores for sequence data, this example might be a good starting point using only built-in datastores, and this example includes a custom sequenceDatastore implementation that you can look at to get an idea of how you might write your own datastore for this task.