how to apply sliding window in deep learning LSTM

Hi All,

i am currenty using MATLAB Deep Learning Toolbox to build a LSTM-Network to do timeseries regression. Since my data has 12 features and 1 response. The input size of the sequenceInputLayer is set to 12. However, i want to apply a sliding window to my data, which means the dimension of the input of the sequenceInputLayer should be a matrix. (if sliding window is 3,then the input should be [12*3] or [3*12] Matrix?) Is that possible and how to realize it? And also, how does Matlab know that the input data is time series?

Layers = [ ...
    sequenceInputLayer(numFeatures)
    gruLayer(numHiddenUnits,'OutputMode','sequence')
    dropoutLayer(0.5)
    fullyConnectedLayer(64)
    dropoutLayer(0.5)
    fullyConnectedLayer(numResponses)
    regressionLayer];

Answer 1

[prabhat kumar sharma]

Hi Tony,

Yes, it's possible to apply a sliding window to your data and use a matrix as input for the sequenceInputLayer.

You can do the following updates.

• If your original data is in a 2D matrix of size (numSamples, numFeatures), reshape it into a 3D tensor of size (numSamples - windowSize + 1, windowSize, numFeatures). This creates the sliding window sequences.
• Change the inputSize property of the sequenceInputLayer to [windowSize, numFeatures]. This tells the layer to expect matrix inputs with the specified dimensions.

layers = [ ...
    sequenceInputLayer([windowSize, numFeatures]) 
    ...  % Rest of your code
];

MATLAB doesn't inherently know that your data is time series. It treats it as a 3D tensor with dimensions corresponding to samples, timesteps, and features.

The LSTM and GRU layers are designed to handle sequential data, so they can learn temporal patterns within your time series.

I hope it helps!