How to connect a 1d convolution layer after a LSTM layer with output mode "last"?

I am working on classify sEMG data, and I want to build a LSTM-CNN network. I want to first employ a LSTM layer and take the last output as the input of a 1D convolution layer.

layers = [ ...
    sequenceInputLayer(sEMG_channels)  
    lstmLayer(numHiddenUnits,'OutputMode','sequence')
    lstmLayer(numHiddenUnits,'OutputMode','last')
    convolution1dLayer(3,25,'Padding','same')
    batchNormalizationLayer
    reluLayer
    dropoutLayer(0.4)
    convolution1dLayer(3,10,'Padding','same')
    batchNormalizationLayer
    reluLayer
    fullyConnectedLayer(23)
    softmaxLayer
    classificationLayer];

However, when I run it. It returns that the convolution1dLayer has 0 temporal dimension and 0 spatial dimension.

my input data XTrain is a 1000*1 cell and in each cell is a 48*30000 double matrix which means that my sEMG data has 48 channels and 30000 time points.

What is the output of lstmLayer(numHiddenUnits,'OutputMode','last') like?

how can i connect LSTM with CNN?

[Manikanta Aditya]

layers = [ ...
    sequenceInputLayer(48) % Assuming sEMG_channels = 48
    lstmLayer(numHiddenUnits, 'OutputMode', 'last')
    reshape([1, 1, numHiddenUnits]) % Reshape LSTM output to have a spatial dimension
    convolution1dLayer(3, 25, 'Padding', 'same')
    batchNormalizationLayer
    reluLayer
    dropoutLayer(0.4)
    convolution1dLayer(3, 10, 'Padding', 'same')
    batchNormalizationLayer
    reluLayer
    fullyConnectedLayer(23)
    softmaxLayer
    classificationLayer];

Check this.

When you set the OutputMode of an LSTM layer to 'last', the output of that layer will be a 2D matrix of size numObservations x numHiddenUnits. Here, numObservations is the number of sequences in your input data, and numHiddenUnits is the number of hidden units in the LSTM layer.

In your case, since your input data XTrain is a 1000x1 cell array, where each cell contains a 48x30000 matrix (48 channels and 30000 time points), the output of the lstmLayer(numHiddenUnits,'OutputMode','last') will be a 2D matrix of size 1000 x numHiddenUnits.

To connect this LSTM output to a 1D convolutional layer, you need to reshape the output to have a valid spatial dimension. You can do this by adding a permute layer or a reshape layer after the LSTM layer.

[Yufan]

Thank you for answer my question!

But adding

reshape([1, 1, numHiddenUnits]) 

Error using reshape
Not enough input arguments.

does not work! There is not enough input arguments.

how can i convey the output of LSTM layer into the reshape layer?

I want to conduct a network like this.

[Manikanta Aditya]

layers = [ ...
    sequenceInputLayer(48) % Assuming sEMG_channels = 48
    lstmLayer(numHiddenUnits, 'OutputMode', 'last')
    reshapeLayer([1 1 numHiddenUnits],'Name','reshape') % Reshape LSTM output to have a spatial dimension
    convolution1dLayer(3, 25, 'Padding', 'same')
    batchNormalizationLayer
    reluLayer
    dropoutLayer(0.4)
    convolution1dLayer(3, 10, 'Padding', 'same')
    batchNormalizationLayer
    reluLayer
    fullyConnectedLayer(23)
    softmaxLayer
    classificationLayer];