What is the training accuracy of this model?

I’m trying to classifiy ECG signals using LSTM and MATLAB, the above plot shows that the training accuracy of the system is 100% but when I apply this code to calculate and get the accuracy I get only 20%

LSTMAccuracy = sum(trainPred == Labels)/numel(Labels)*100

Am I missing something here? Or there something wrong I did in my code?

Here is the configuration and the training code:

layers = [ ...
    sequenceInputLayer(1)
    bilstmLayer(100,'OutputMode','last') 
    fullyConnectedLayer(5)
    softmaxLayer
    classificationLayer
    ]

options = trainingOptions('adam', ...
    'MaxEpochs',1750, ...
    'MiniBatchSize', 150, ...
    'InitialLearnRate', 0.0001, ...
    'ExecutionEnvironment',"auto",...
    'plots','training-progress', ...
    'Verbose',false);

net = trainNetwork(Signals, Labels, layers, options);

trainPred = classify(net, Signals,'SequenceLength',1000);

LSTMAccuracy = sum(trainPred == Labels)/numel(Labels)*100
figure
confusionchart(Labels,trainPred,'ColumnSummary','column-normalized',...
              'RowSummary','row-normalized','Title','Confusion Chart for LSTM');

I really appreciate your help, thanks.

Answer 1

[Chunru]

If Labels is character array instead of string array, then numel(Labels) will give the number of characters instead of the number of signals intended.

Labels =["abc"; "def"; "ghi"];
numel(Labels)

ans = 3

Labels =['abc'; 'def'; 'ghi'];
numel(Labels)

ans = 9

My guess is that you have a character arrays with 5 characters for each label. If this is the case, use the following

size(Labels, 2)

ans = 3

[LinkTo]

Thank you for your response, but unfortunately this didn't help.

Each label is only 1 character long. So, If I understand correctly "size" and "numel" will return the same value.

Labels = ["a"; "b"; "c"];
numel(Labels)

ans = 3

Labels = ['a'; 'b'; 'c'];
numel(Labels)

ans = 3

size(Labels, 1)

ans = 3

[Chunru]

Can you check the results of

whos

[LinkTo]

The command whos return the following:

Name Size Bytes Class Attributes

Labels 10080x1 10616 categorical

SIG 2520x1 6330240 cell

Signals 10080x1 25320960 cell

YLabels 2520x1 3056 categorical

i 1x1 8 double

layers 5x1 2206 nnet.cnn.layer.Layer

options 1x1 6333733 nnet.cnn.TrainingOptionsADAM

[Chunru]

How about trainPred? Show the results of the following variables:

Labels
trainPred

[LinkTo]

Sorry I didn't train the model so the variable "trainPred" didn't show before.

Name Size Bytes Class Attributes

Labels 10080x1 10616 categorical

trainPred 10080x1 10616 categorical

As you can see "Labels" and "trainPred" are identical in size and class.

[Chunru]

Show a portion of data

[Labels trainPred]

[LinkTo]

The output is similar to this (This is just an example, as I didn't have the model trained right now)

[Labels trainPred]

ans = 5×2 string array
    "P"    "R"
    "Q"    "S"
    "R"    "R"
    "S"    "P"
    "T"    "P"

And yes, "Labels" and "trainPred" are different, so the accuracy can't be 100%

[Chunru]

Try the following:

trainPred = classify(net, Signals,'SequenceLength','longest');

[LinkTo]

It worked. Thank you.