How to understand the input and output arguments of preparets function for network training

For training NARX network, we need to prepare input and target time series data using preparets function. However, the preparets function has many input and output arguments which are confused for me to understand their functioning. I have read the examples about preparets, yet it is still difficult to understand. What is the best practice to understand the preparets and correctly use it under different network training scenarios ?

help preparets

 preparets - Prepare input and target time series data for network simulation or training
    This MATLAB function takes these arguments: net — Neural network Xnf —
    Non-feedback inputs Tnf — Non-feedback targets Tf — Feedback targets EW
    — Error weights (optional)and returns these arguments: Xs — Shifted
    inputs Xi — Initial input delay states Ai — Initial layer delay states
    Ts — Shifted targets EWs — Shifted error weights shift — The number of
    timesteps truncated from the front of X and T in order to properly fill
    Xi and Ai.

    Syntax
      [Xs,Xi,Ai,Ts,EWs,shift] = preparets(net,Xnf,Tnf,Tf,EW)

    Input Arguments
      net - Input network
        network
      Xnf - Non-feedback inputs
        cell array
      Tnf - Non-feedback targets
        cell array
      Tf - Feedback targets
        cell array
      EW - Error weights
        cell array

    Output Arguments
      Xs - Shifted inputs
        cell array
      Xi - Initial input delay states
        cell array
      Ai - Initial layer delay states
        cell array
      Ts - Shifted targets
        cell array
      EWs - Shifted error weights
        cell array
      shift - Timesteps
        scalar

    Examples
      openExample('nnet/PrepareDataForOpenAndClosedLoopNetworksExample')

    See also adddelay, closeloop, narnet, narxnet, openloop, removedelay,
      timedelaynet

    Introduced in Deep Learning Toolbox in R2010b
    Documentation for preparets
       doc preparets

EDIT: removed copyright code.

Answer 1

[Soumya]

Hi,

'preparets' is used to prepare and align time series data for dynamic neural networks such as NARX, NAR, and time-delay networks prior to training or simulation. Below is a detailed explanation:

• The function automatically applies the input and layer delays defined in the network ('net.inputDelays', 'net.layerDelays') to ensure the data is consistent with the network architecture.
• It partitions the data into:
• Initial delay conditions → 'Xi' (input states) and 'Ai' (layer states)
• Aligned training/simulation data → 'Xs' (inputs) and 'Ts' (targets)
• The function removes the initial timesteps required to fill delay buffers. The number of removed samples is returned as 'shift'.

Input arguments interpretation:

• 'net' → neural network object
• 'Xnf' → external (non-feedback) inputs
• 'Tnf' → non-feedback targets
• 'Tf' → feedback targets (used in NARX configurations)
• 'EW' → optional error weights

Output arguments interpretation:

• 'Xs' → shifted and aligned inputs
• 'Xi' → initial input delay states
• 'Ai' → initial layer delay states
• 'Ts' → shifted and aligned targets
• 'EWs' → shifted error weights
• 'shift' → number of truncated initial timesteps
• For a typical open-loop NARX network (training phase):['Xs', 'Xi', 'Ai', 'Ts'] = 'preparets'('net', 'X', {}, 'T')
• For a closed-loop configuration (simulation/prediction phase):['Xs', 'Xi', 'Ai', 'Ts'] = 'preparets'('net', 'X', 'T', 'T')
• Users should not manually shift or align time series data, as 'preparets' is designed to handle all delay and alignment operations internally.
• All time series data must be provided as cell arrays, where each cell represents one timestep.
• The function is typically called just before 'train' or 'sim', ensuring compatibility between the data and the network.

Please refer to the following documentation for more details:

• https://www.mathworks.com/help/deeplearning/ref/preparets.html
• https://www.mathworks.com/help/deeplearning/ref/narxnet.html

I hope this help!