Unable to reshape the data to 3D on Matlab for LSTM training?

I was trying to reshape my data array from 2D to 3D to be able to be loaded to LSTM for training, i encountered the dimensionality problem in the beginning when I could not reshape my data into ```[samples, time steps, features]``` (in my case, ```[2097 1000 1]```, since the MATLAB always automatically turns it into ```[[2097 1000]```.

The reshaping method I used is

sequence_length = 1000;
num_features = 1; %  1D time series

% Check the number of samples
num_samples = floor(length(final_filtered_signal) / sequence_length);
X = reshape(final_filtered_signal(1:num_samples * sequence_length), [num_samples, sequence_length, num_features]);

My initial layer structure is as following:

% Define the CNN-LSTM model
layers = [
    sequenceInputLayer(num_features, 'MinLength', sequence_length)
    
    convolution1dLayer(3, 64, 'Padding', 'same')
    batchNormalizationLayer
    reluLayer
    maxPooling1dLayer(2, 'Stride', 2)
    
    convolution1dLayer(3, 128, 'Padding', 'same')
    batchNormalizationLayer
    reluLayer
    maxPooling1dLayer(2, 'Stride', 2)
    
    convolution1dLayer(3, 256, 'Padding', 'same')
    batchNormalizationLayer
    reluLayer
    
    lstmLayer(100, 'OutputMode', 'last')
    
    fullyConnectedLayer(64)
    reluLayer
    dropoutLayer(0.5)
    fullyConnectedLayer(2) % 2 output neurons for binary classification
    softmaxLayer
    classificationLayer
    ];

With this layer set-up I run into error:

Error using trainNetwork
The training sequences are of feature dimension 1677 but the input layer expects sequences of feature dimension 1.

Error in noise_kernel_deconvolve (line 260)
net = trainNetwork(X_train, Y_train, layers, options);

Since I keep run into dimensionality problem, I changed the layer to :

% Define the CNN-LSTM model
layers = [
    sequenceInputLayer([sequence_length, num_features])
    
    convolution1dLayer(3, 64, 'Padding', 'same')
    batchNormalizationLayer
    reluLayer
    maxPooling1dLayer(2, 'Stride', 2)
    
    convolution1dLayer(3, 128, 'Padding', 'same')
    batchNormalizationLayer
    reluLayer
    maxPooling1dLayer(2, 'Stride', 2)
    
    convolution1dLayer(3, 256, 'Padding', 'same')
    batchNormalizationLayer
    reluLayer
    
    flattenLayer  % Add this layer to flatten the output before LSTM
    
    lstmLayer(100, 'OutputMode', 'last')
    
    fullyConnectedLayer(64)
    reluLayer
    dropoutLayer(0.5)
    fullyConnectedLayer(2) % 2 output neurons for binary classification
    softmaxLayer
    classificationLayer
    ];

basically changing the ```sequenceInputLayer(num_features, 'MinLength', sequence_length)``` to accept the 2D matrix format of my input data. After this I encountered new error:

Error using trainNetwork
The training sequences are of feature dimension 1677  1000 but the input layer expects sequences of feature dimension 1000     1.

Error in noise_kernel_deconvolve (line 230)
net = trainNetwork(X_train, Y_train, layers, options);

I am now a bit confused how does LSTM require the input data and the layer to be structured... since it seems doesn't matter if I have 2D input X or 3D input ```X```, I would always encounter the dimension problem when I start running the

net = trainNetwork(X_train, Y_train, layers, options);

Answer 1

[Ayush Aniket]

Hi Jingyi,

The input format required by the LSTM network in MATLAB for dataset of sequences is a Nx1 cell array where each element is a c-by-s matrix, where c is the number of features of the sequence and s is the sequence length. Refer to the following document link to read about various input formats: https://www.mathworks.com/help/deeplearning/ref/trainnetwork.html#mw_36a68d96-8505-4b8d-b338-44e1efa9cc5e

In your first approach, the correct format for X_Train would be a cell array with samples x 1 (i.e. 2097 x 1) dimension, wherein each entry should be in the fromat of features x sequence_length (i.e. 1 x 1000).

Refer to the following MATLAB answer: https://www.mathworks.com/matlabcentral/answers/2127811-lstm?s_tid=srchtitle