Use of catsamples() to concatenate sequences for a neural network

I have developed a MIMO NARX network where I have multiple sequences (trials). Originally, I concatenated the data with one trial after the other to create "continuous" input and target time series. Because the data is not really continuous, I get a discontinuation between each concatenation. After some digging, I found these 2 links that talks about problem ( https://www.mathworks.com/help/nnet/ug/understanding-neural-network-toolbox-data-structures.html ) and the solution ( https://www.mathworks.com/help/nnet/ug/multiple-sequences-with-dynamic-neural-networks.html ). However I am having troubles implementing the catsamples() command based on how my data is setup.

For the inputs, I have a cell array of 3x30 where each column is a sequence of different length of [2X1] (2 inputs) and each row will be a different input series. I have different experimental conditions saved in 1 cell array but each will be tested separately as a single input series.

EMG =3×30 cell array
  Columns 1 through 5
    {1×134 cell}    {1×137 cell}    {1×133 cell}    {1×134 cell}    {1×136 cell}
    {1×134 cell}    {1×137 cell}    {1×133 cell}    {1×134 cell}    {1×136 cell}
    {1×134 cell}    {1×137 cell}    {1×133 cell}    {1×134 cell}    {1×136 cell}
  Columns 6 through 10 ...

EMG{1, 1}{1, 1} = 1×134 cell array
  Columns 1 through 5
    [2×1 double]    [2×1 double]    [2×1 double]    [2×1 double]    [2×1 double]
  Columns 6 through 10 ...

The output is just a 1x30 where each column is a sequence/trial and each cell has 2 target outputs.

Angle_Moment=1×30 cell array
  Columns 1 through 5
    {1×134 cell}    {1×137 cell}    {1×133 cell}    {1×134 cell}    {1×136 cell}
 Columns 6 through 10 ...

    Angle_Moment{1, 1}{1, 1} = 1×134 cell array
      Columns 1 through 5
        [2×1 double]    [2×1 double]    [2×1 double]    [2×1 double]    [2×1 double]
      Columns 6 through 10 ...

All the examples I found of catsamples() have each sequence in a separate variables. How can I implement

x_mul = catsamples(x1,x2,x3,'pad');

but with all the sequences in one single cell array with 2 inputs/outputs? Should I set up my data differently? How?

Also, I am planning on dividing my data using

net.divideFcn ='divideblock';

with 70-20-10%. Does that means it will take the first 21 values of the {1x30}?

Please respond if you have insight in either of the 2 questions (catsamples or divideFcn).

Thank you,

Answer 1

[ErikaZ]

catsamples( EMG_InputSeries{i,:},'pad'); % For EMG
catsamples( Angle_Moment_TargetSeries{:},'pad'); % For Ankle_Moment

This will create a 1x400s cell array with 2x30 cell inside each.

For default 'divideblock', it takes 70 percent of 400s time points for training. If needed to take 70% of the Q = 30, the mode must be changed to sample

net.divideMode = 'sample';

This will take 21 points from every of the 400s time points for training.

[Torsten K]

Hello everybody,

i know the thread is old but i'm looking for a solution to a similar problem. I have 150 time series of different lengths (max 600 timesteps), 2 inputs and 1 target and I want to train a NARX network. I converted the time series into a cell array X for the inputs and a cell array T for the targets:

X = 1×600 cell array

Columns 1 through 2 [...] Columns 599 through 600

{2×150 double} {2×150 double} [...] {2×150 double} {2×150 double}

T = 1×600 cell array

Columns 1 through 2 [...] Columns 599 through 600

{1×150 double} {1×150 double} [...] {1×150 double} {1×150 double}

Now I would like to use 110 "complete" time series for training, 20 for validation and 20 for testing. The time series should not be torn apart, as each time series reflects a single dynamic process, so it makes no sense for me to train with the first 70% of the timeseries, validate with the next 15% and test with the last 15%.

Does anyone have any suggestion on how to implement this approach in Matlab code?

Best regards

Torsten