Multidimensional input to SequenceInputLayer

Hello dear community,

i am trying to train a neural network for a regression task.

I do have different cases of data though and the model should be able to differentiate between the different data sets.

For the training, i therefore thought, that i have Data with 2 Channels (2 input neurons), 3 Batches (3 different operating points to be trained) and 50 T (50 timesteps).

Therefore i would have a multi dimensional input matrix.

The target matrix would have one output, 3 Batches and also 50 timesteps.

However, i keep on getting the error:

Error using trainNetwork (line 191)

The training sequences are of feature dimension 2 50 but the input layer expects sequences of feature dimension 2.

Error in delme2 (line 46)

net = trainNetwork(permute(X, [1 3 2]), permute(Y, [1 3 2]), layers, options);

I have checked the internet, but i don't really find a solution to this.

Thank you in Advance !

Here is the Code

% Manuell definierte Dimensionen

B = 3; % Anzahl der Arbeitspunkte (Samples)

T = 50; % Anzahl der Zeitschritte

C = 2; % Anzahl der Eingangsmerkmale

O = 1; % Anzahl der Ausgabemerkmale

% Beispielhafte Daten für verschiedene Arbeitspunkte (APs)

X = zeros(C, B, T);

Y = zeros(O, B, T);

% Arbeitspunkt 1

X(1, 1, :) = 2000;

X(2, 1, :) = 70;

Y(1, 1, :) = rand(1, 1, T) * 100 + 50;

% Arbeitspunkt 2

X(1, 2, :) = 4000;

X(2, 2, :) = 50;

Y(1, 2, :) = rand(1, 1, T) * 100 + 40;

% Arbeitspunkt 3

X(1, 3, :) = 3000;

X(2, 3, :) = 60;

Y(1, 3, :) = rand(1, 1, T) * 100 + 45;

% Netzwerkparameter

numHiddenUnits = 50; % Anzahl der LSTM-Zellen

% Netzwerkarchitektur definieren

layers = [ ...

sequenceInputLayer(C) % C Eingangsmerkmale

lstmLayer(numHiddenUnits, 'OutputMode', 'sequence') % LSTM Layer

fullyConnectedLayer(O) % O Ausgabemerkmale

regressionLayer]; % Regressions-Layer für kontinuierliche Ausgaben

% Trainingsoptionen festlegen

options = trainingOptions('adam', ...

'MaxEpochs', 100, ...

'MiniBatchSize', 1, ...

'GradientThreshold', 1, ...

'InitialLearnRate', 0.01, ...

'Verbose', false, ...

'Plots', 'training-progress');

% Netz trainieren

net = trainNetwork(permute(X, [1 3 2]), permute(Y, [1 3 2]), layers, options);

% Permutieren, um X und Y in die erwartete Form C x T x B und O x T x B zu bringen

% Beispielhafte Inferenz für einen neuen Arbeitspunkt

XTest = zeros(C, 1, T); % Testdaten in der Form C x 1 x T

XTest(1, 1, :) = 3500;

XTest(2, 1, :) = 65;

% Kurve vorhersagen

YPred = predict(net, permute(XTest, [1 3 2])); % Permutieren, um XTest in die erwartete Form C x T x 1 zu bringen

% Vorhergesagte Temperaturkurve anzeigen

plot(squeeze(YPred)); % Squeeze, um die 2D-Kurve für das Plotting zu erhalten

xlabel('Zeitschritte');

ylabel('Vorhergesagte Temperatur');

title('Vorhergesagte Temperaturkurve für neuen Arbeitspunkt');

Answer 1

[Sahas]

Hi @tobias,

As per my understanding, you are training a neural network for a regression task and encountering an error when using the “trainNetwork” function.

I was able to reproduce the error at my end and identified the issue. There was a mismatch between the expected input dimensions of the network and the actual dimensions of the data being fed into it.

I made the following changes in the code to ensure smooth data processing and handle dimension mismatching:

• The X, XTest and Y matrices are arranged in format (features, timesteps, samples) which is the expected input format for the LSTM layer.
• The “permute” function is removed from the “trainNetwork” function arguments as it was unnecessary.
• The data was reshaped to fit the expected input format for “trainNetwork” function.

% Manually defined dimensions
B = 3;  % Number of operating points (samples)
T = 50; % Number of time steps
C = 2;  % Number of input features
O = 1;  % Number of output features

%%%%%Changes here to match the expected input format%%%%%
% Example data for different operating points (OPs)
X = zeros(C, T, B); 
Y = zeros(O, T, B); 

% Operating point 1
X(1, :, 1) = 2000;  
X(2, :, 1) = 70;   
Y(1, :, 1) = rand(1, T) * 100 + 50; 

% Operating point 2
X(1, :, 2) = 4000;  
X(2, :, 2) = 50;   
Y(1, :, 2) = rand(1, T) * 100 + 40;  

% Operating point 3
X(1, :, 3) = 3000;  
X(2, :, 3) = 60;    
Y(1, :, 3) = rand(1, T) * 100 + 45;  

% Network parameters
numHiddenUnits = 50; % Number of LSTM cells

% Define network architecture
layers = [ ...
    sequenceInputLayer(C)             % C input features
    lstmLayer(numHiddenUnits, 'OutputMode', 'sequence') % LSTM layer
    fullyConnectedLayer(O)            % O output features
    regressionLayer];                 % Regression layer for continuous outputs

% Set training options
options = trainingOptions('adam', ...
    'MaxEpochs', 100, ... 
    'MiniBatchSize', 1, ...
    'GradientThreshold', 1, ...
    'InitialLearnRate', 0.01, ...
    'Verbose', false, ...
    'Plots', 'training-progress');

% Reshape data to fit the expected input format for trainNetwork
% Each sequence is of size B, with cell of size C x T
XCell = squeeze(mat2cell(X, C, T, ones(1, B)));
YCell = squeeze(mat2cell(Y, O, T, ones(1, B)));

% Train network
net = trainNetwork(XCell, YCell, layers, options);

%%%%%Changes here to match the expected output format%%%%%
% Example inference for a new operating point
XTest = zeros(C, T, 1); % Test data in the form C x T x 1
XTest(1, :, 1) = 3500;  
XTest(2, :, 1) = 65;    

% Predict curve
YPred = predict(net, squeeze(XTest));

% Display predicted temperature curve
plot(squeeze(YPred)); % Squeeze to obtain the 2D curve for plotting
xlabel('Time Steps');
ylabel('Predicted Temperature');
title('Predicted Temperature Curve for New Operating Point');

Refer to the following MathWorks documentation on “trainNetwork” for more information:

https://www.mathworks.com/help/deeplearning/ref/trainnetwork.html

Hope this is beneficial!