help needed with setup for an ltsm for sequence classification

Hello,

I'm trying to create a classifier for sequence data using the lstm network. I have time-series data with 25 timepoints where 8 features are computed.

Following the examples on the matlab help (japanese vowels, human activity data) results in the classifier not learning anything (flat accuracy at chance level, and no loss values at all in the training progress window). I'm new at setting up DNN's so i'm not sure at what level something is going wrong: input data format, layer spec, training options, etc. What kind of network and training options would work well for this kind of data?

My data:

65 samples of 8 features x 25 timepoints, stored in a cell array X={65,1}(8x25)
label is the 4 classes the data belong to: Y=65x1 categorical

Here are the layers and options:

inputSize = 8; % the 8 features
numHiddenUnits = 100;
numClasses = 4; % 4 classes

layers = [ ...
    sequenceInputLayer(inputSize)
    bilstmLayer(numHiddenUnits,,'OutputMode','last') 
    fullyConnectedLayer(numClasses)
    softmaxLayer
    classificationLayer]

options = trainingOptions('adam', ...
    'MaxEpochs',100, ...
    'InitialLearnRate', 0.01, ...
    'SequenceLength', 25, ...
    'GradientThreshold', 1, ...
    'plots','training-progress', ...
    'Verbose',false);

net = trainNetwork(X,Y,layers,options);

Any help is appreciated!