Multiple output in TDNN (Time Delay Neural Network)

Hello,

My name is Oresti and I am currently working with time delay neural networks ("timedelaynet").

I have three inputs signals and three outputs signals, and the problem appears when the the data is prepared (for taking into account the delay established).

[Is_tr,Ii_tr,Ai_tr,Os_tr] = preparets(net,cell_inputs_train,cell_outputs_train);

The idea is to define the particular NN by establishing the different parameters, but actually this is in principle not allowed for the outputs:

Considering 'net' the struct for the TDNN, defined as:

net = timedelaynet(DELAYS,LAYERS);

it is absolutely feasible to define the number of inputs

net.numInputs = Ninputs;

but as far as I know you cannot modify the number of outputs

net.numOutputs (READ ONLY)

because you have only read permissions.

Of course, I know that the problem can be split so defining three independent TDNN for each output (my current solution) but I would like to know why I cannot define it directly.

If someone has an idea I will really appreciate the answer.

Thanks in advance,

Oresti.

Answer 1

[Greg Heath]

% I have three inputs signals and three outputs signals, and the problem % appears when the the data is prepared (for taking into account the % delay established). % % % [Xs,Xi,Ai,Ts,EWs,shift] = preparets(net,Xnf,Tnf,Tf,EW) % % [Is_tr,Ii_tr,Ai_tr,Os_tr] = preparets(net,cell_inputs_train, ... % cell_outputs_train);

1. IF YOU PUT TRAINING DATA INTO PREPARETS, YOU WILL HAVE TO PREVENT CONFIGURE OR TRAIN FROM USING THE DEFAULT DIVIDERAND TO OBTAIN VAL & TEST DATA.

% The idea is to define the particular NN by establishing the different % parameters, but actually this is in principle not allowed for the outputs: % % Considering 'net' the struct for the TDNN, defined as: % % net = timedelaynet(DELAYS,LAYERS);

2. NO. THE NO. OF HIDDEN LAYER NODES, NOT THE NUMBER OF LAYERS.

% it is absolutely feasible to define the number of inputs % % net.numInputs = Ninputs; % % but as far as I know you cannot modify the number of outputs % % net.numOutputs (READ ONLY) % % because you have only read permissions.

3.NO. NUMBER OF INPUTS AND OUTPUTS RECOGNIZED BY THE NET IS DETERMINED BY ONFIGURE OR TRAIN

% Of course, I know that the problem can be split so defining three % independent TDNN for each output (my current solution) but I would % like to know why I cannot define it directly. % % If someone has an idea I will really appreciate the answer.

SEE 3 ABOVE. ALSO, MULTIDIMENSIONAL DOCUMENTATION EXAMPLES.

HOPE THIS HELPS.

GREG

Answer 2

[Khalid]

I know this is a really old post, but I arrived here via google search and didn't find any of the answers worked for me. The answer which I found worked is here.

Basically for multi-dimensional inputs, or to produce multi-step forecast outputs all at once (equivalent to a multi-dimensional output) you need to create your inputs and outputs using

con2seq()

- Khalid.

Answer 3

[Greg Heath]

In the call to preparenets the input and output dimensions are automatically determined.

Hope this helps.

Greg

[Oresti]

Hi Greg,

Conversely I experience this problem:

??? Error using ==> preparets at 164

Number of output signals does not match number of network's non-feedback outputs.

Error in ==> tdnn_test at 114

[Is_tr,Ii_tr,Ai_tr,Os_tr] = preparets(net,cell_inputs_train,cell_outputs_train);

Error in ==> TDNN_ESTIMATION_EXAMPLE_POS_ACC at 151

TDNN_struct = tdnn_test(I_Tr,O_Tr,I_Ts,O_Ts,...

Basically what I am passing as inputs and outputs are three signals (X,Y,Z) for the input and three signals (X,Y,Z) for the output.

Any idea?

Thanks in advance,

Oresti.