trainNetwork not running because categorical row vector not found even though it's there

I'm running a deep learning script over a signal dataset using a sequence-to-label architecture I put together in the Deep Learning Network designer. The input has two channels and I'm attempting to output into 41 separate classes using 1D-CNN/LSTM. I defined the train and test data as such:

idx = splitlabels(labels,[0.8,0.1,0.1])
trainidx = idx{1}
validx = idx{2}
testidx = idx{3}

traindata = sigdata(trainidx)
trainlabels = labels(trainidx)
valdata = sigdata(validx)
vallabels = labels(validx)
testdata = sigdata(testidx)
testlabels = labels(testidx)

The training options are defined as:

opts = trainingOptions("adam", ...
    "Plots","training-progress", ...
    "ValidationData",{valdata vallabels}, ...
    "InitialLearnRate",0.0001, ...
    "MaxEpochs",15, ...
    "Shuffle","every-epoch", ...
    "ValidationFrequency",20, ...
    "MiniBatchSize",32)

and the network architecture looks like:

12×1 Layer array with layers:

     1   'sequence'      Sequence Input          Sequence input with 2 dimensions
     2   'conv1d'        1-D Convolution         8 3 convolutions with stride 1 and padding 'same'
     3   'batchnorm'     Batch Normalization     Batch normalization
     4   'relu'          ReLU                    ReLU
     5   'maxpool1d'     1-D Max Pooling         Max pooling with pool size 2, stride 1, and padding 'same'
     6   'conv1d_1'      1-D Convolution         8 3 convolutions with stride 1 and padding 'same'
     7   'batchnorm_1'   Batch Normalization     Batch normalization
     8   'relu_1'        ReLU                    ReLU
     9   'lstm'          LSTM                    LSTM with 128 hidden units
    10   'fc'            Fully Connected         41 fully connected layer
    11   'softmax'       Softmax                 softmax
    12   'classoutput'   Classification Output   crossentropyex
    

I'd prefer not to use global pooling layers, because that would lead to a very low classification accuracy. Using all of these inputs, I attempt to train the network:

    damagenet = trainNetwork(traindata,trainlabels,layers_1,opts)

However, doing this leads to the following compilation error:

Error using trainNetwork

Invalid training data for sequence-to-sequence classification network. For multiple observations, the responses

must be a cell array of categorical row vectors. For a single observation, the responses can be also a single

categorical row vector.

This doesn't make any sense to me, as I'm not trying to run a sequence-to-sequence; I'm trying sequence-to-label. Hence the size-41 fully-connected layer. Furthermore, I am using a cell array of categorical row vectors. In Workspace, my labels variable is listed as a 4141x1 categorical. And again, I don't get this error when I try it with a global pooling layer after the last relu layer. I just get a classification with really low accuracy.

What should I do to my network, training data, or options for the code to run without error? Any help or advice is appreciated.

Thanks!

[Matt J]

It would be easiest if you attached a .mat file containing traindata,trainlabels,layers_1,opts so we could run the code ourselves. Please put them all in one .mat file. Do not post 4 separate .mat files.

[Said El-Hawwat]

Would it be possible to send it to you personally via email or some other means? The file size is too big to attach here, even if I compress it

[Matt J]

We don't need the entire data set. Just a subset that reproduces the error.

[Said El-Hawwat]

I'm sorry but it would be nice if I could just email it to you personally. Even when I take a subset it still tells me that the filesize is too big to attach

Answer 1

[Cris LaPierre]

Perhaps this example will help?

https://www.mathworks.com/help/deeplearning/ug/classify-sequence-data-using-lstm-networks.html

I'm not able to duplicate the error, but using the data from the example, the following ran without error.

idx = splitlabels(labels,[0.8,0.1,0.1])
trainidx = idx{1}
validx = idx{2}
testidx = idx{3}
traindata = data(trainidx)
trainlabels = labels(trainidx)
valdata = data(validx)
vallabels = labels(validx)
testdata = data(testidx)
testlabels = labels(testidx)

% layers generated by Deep Network Designer
layers_1 = [
    sequenceInputLayer(numChannels,"Name","input")
    convolution1dLayer(8,3,"Name","conv1d","Padding","same")
    batchNormalizationLayer("Name","batchnorm")
    reluLayer("Name","relu")
    maxPooling1dLayer(2,"Name","maxpool1d","Padding","same")
    convolution1dLayer(8,3,"Name","conv1d_1","Padding","same")
    batchNormalizationLayer("Name","batchnorm_1")
    reluLayer("Name","relu_1")
    lstmLayer(128,"Name","lstm","OutputMode","last")
    fullyConnectedLayer(numClasses,"Name","fc")
    softmaxLayer("Name","softmax")
    classificationLayer("Name","classification")];

opts = trainingOptions("adam", ...
    "Plots","training-progress", ...
    "ValidationData",{valdata vallabels}, ...
    "InitialLearnRate",0.0001, ...
    "MaxEpochs",15, ...
    "Shuffle","every-epoch", ...
    "ValidationFrequency",20, ...
    "MiniBatchSize",32);

damagenet = trainNetwork(traindata,trainlabels,layers_1,opts)

[Said El-Hawwat]

Thanks for the help; this allowed me to get around the categorical dataset issue.

Unfortunately my training accuracy is still too low though.