Error using predictAndUpdateState (LSTM NN)

Hello guys,

I am trying to conduct a regression analysis using a LSTM neural network.

I am using 8 variables as input, and obtaining 1 output.

My knowledge in Deep Learning Toolbox is limited, therefore I have used Neural Network Fitting App to create the network.

Once exported, I am trying to predict into the future using the function predictAndUpdateState. However, I keep getting the same error message:

% Xnew is a cell array with the 8 inputs I want to use to predict.
>> for i = 2:numTimeStepsTest
    v = Xnew(:,i);
    [net1,score] = predictAndUpdateState(net1,v);
    scores(:,i) = score; 
end 
Undefined function 'predictAndUpdateState' for input arguments of type 'network'.

As I understand, a LSTM network is a recurrent neural network, therefore I don't know where the mistake could be.

As I said, my knowledge is very limited, so I would appreciate any help on this matter.

Many thanks,

Natalia

[NN]

following

Answer 1

[Vineet Joshi]

Hi!

Background:

In the following code I have used the command line equivalent of 'Neural Network Fitting App' to create a simple network.

trainFcn = 'trainlm';
hiddenLayerSize = 10;
net = fitnet(hiddenLayerSize,trainFcn);
class(net)

ans = 'network'

As you can see the 'fitnet' returns a network of type 'network'.

From the error shared by you, it looks like your case is same as well since input argument is of type 'network'.

Understanding the Error:

The documentation of predictAndUpdateState states that the input network can be of two types only. It can either be a SeriesNetwork object or a DAGNetwork object.

Possible Workaround:

The most strightforward workaround is to create a SeriesNetwork object or a DAGNetwork object. Attaching a few links to help you with this.

Helpful Links:

• predictAndUpdateState - Network Class
• DAGNetwork
• SeriesNetwork

Answer 2

[Pjeter Berisa]

I think the root problem is the same as discussed in this question:

How are the functions "train" and "trainNetwork" different underneath? - MATLAB Answers - MATLAB Central (mathworks.com)