Handle Multiple Sequences with GRU/LSTM Dynamic Neural Networks

I am performing system identification using neural networks. My data includes multiple subsets of data (experimental trials) rather than one long continuous dataset. Due to this, I can not simply concatenate the data to train a dynamic neural network since there will be discontinuities. I found this page which gets around this issue when designing a NARX network - https://www.mathworks.com/help/deeplearning/ug/multiple-sequences-with-dynamic-neural-networks.html

However, when I try to use this method with a LSTM or GRU design, i get the following error:

Error using DAGNetwork/predictRNN>iAssertInitialStateIsValidForPredict (line 69)
Incorrect network state. The network expects mini-batches size of 16, but was passed a mini-batch of size 1.

Error in DAGNetwork/predictRNN (line 11)
iAssertInitialStateIsValidForPredict(statefulLayers, dispatcher.MiniBatchSize)

Error in DAGNetwork/predictAndUpdateState (line 123)
[Y, finalState, predictNetwork] = this.predictRNN(X, dispatcher, ...

Error in SeriesNetwork/predictAndUpdateState (line 379)
            [this.UnderlyingDAGNetwork, Y] = this.UnderlyingDAGNetwork.predictAndUpdateState(X, varargin{:});

Error in GRU_setup (line 327)
        [updatedNet,yy] = predictAndUpdateState(updatedNet,v);

Note that the dimension of 16 is referring to the number of experimental trials used to train the network. I concatenated these trials using the catsamples function and got a 1x3600 cell array with each cell have a double array with dimensions 5x16. The number of inputs I have is 5 and the number of samples per test trial is 3600. Additionally, this is a single output system.

Why does the catsample method work for NARX networks but not for LSTM/GRU networks? What is a good way to organize my discontinuous data for training and execution of a LSTM/GRU network?

[ahmed shahin]

did u solve this isuue please?

[Chris P]

No.