CNN+BILSTM Architecture

Hello

Could someone please let me know if my implmentation of CNN+BILSTM network is correct? Am not getting good performance:

I am trying to classify 12-Lead ECG signals

inputSize = [1250 12 1];
numHiddenUnits = 10;
layers = [ ...
    sequenceInputLayer(inputSize,'Name','input','normalization','none')
    sequenceFoldingLayer('Name','fold')
    convolution2dLayer([21 1],16,'Name','conv1','Padding','same')
    maxPooling2dLayer([7 1],'Stride',7,'Name','maxpool1','Padding','same')
    convolution2dLayer([17 1],32,'Name','conv2','Padding','same')
    maxPooling2dLayer([6 1],'Stride',6,'Name','maxpool2','Padding','same')
    convolution2dLayer([13 1],64,'Name','conv3','Padding','same')
    maxPooling2dLayer([7 1],'Stride',7,'Name','maxpool3','Padding','same')

    sequenceUnfoldingLayer('Name','unfold')
    flattenLayer('Name','flatten')
    
    bilstmLayer(numHiddenUnits,'OutputMode','last','Name','bilstm1')
    
    fullyConnectedLayer(numClasses, 'Name','fc')
    softmaxLayer('Name','softmax')
    classificationLayer('Name','classification')];

Thank you

[Asvin Kumar]

If you are referring to any paper or material, providing that as a reference would help anyone from the community validate or make suggestions if that's what you're looking for.

Answer 1

[Mohanad Alkhodari]

anyone?