Is it possible to use a three-dimensional matrix to train LSTM network in Matlab R2020b?

Hello everyone,

Previously I trained the neural networks with the trainNetwork(XTrain0',YTrain',layers0,options), with XTrain0 as the input variables, expressed as a (7670 x 7) array, and the YTrain as the output variable that I would like to predict in the future, expressed as a (7670 x 1) array. The result looks good..

However, I currently have difficulty on training the LSTM network with 3-dimensional input data. I want to determine the details fo the original input data variable by breaking them into 139 segments. Originally I tried to use (1058460 x 7) array and (1058460 x 1) array as inputs, but the result of prediction was highly inaccurate.

and I tried to test whether I can do it by creating 3-dimensional variables: the input ends up looking like (7670 x 7 x 139) array and the input is (7670 x 1 x 139) array, but I have difficulty on rewriting the network (layers0). Here is my original modified version of layers0:

layers0 = [ ...

sequenceInputLayer([7,139,1])

lstmLayer(100,'OutputMode','sequence')

fullyConnectedLayer(1)

dropoutLayer(0.4)

regressionLayer];

I tried to look up every page of MATLAB but cannot find the solution. Please help me improve my layers0 so that I can input my 3-dimensional variables. Thank you.

[Aditya Patil]

Can you elaborate on what issue/error you are facing? sequenceInputLayer accepts 2D/3D inputs, so this should be fine.

[Pai-Feng Teng]

If you input sequence layer in the format of 3-dimensional, they will tell you that it must be scalar.

I look up the internet and I can't find any code with sequenceInputLayer that accepts 3D inputs

[Aditya Patil]

Here is an example for 3D input, https://www.mathworks.com/help/deeplearning/ref/nnet.cnn.layer.sequenceinputlayer.html#mw_d1a76588-f98f-4d19-9130-d4411b4a5ee9