Error using trainNetwork (line 184) The training sequences are of feature dimension 1 792 but the input layer expects sequences of feature dimension 1.

% Parameters
n_samples = 1000;
signal_frequency = 10;
time = linspace(0, 1, n_samples)';
I_signal = cos(2 * pi * signal_frequency * time);
Q_signal = sin(2 * pi * signal_frequency * time);

% Create pairs of I and Q signals
X = I_signal(1:end-1); % I signal as input
y = Q_signal(2:end);   % Corresponding Q signal as target output

% Define sequence length (number of time steps in each sequence)
sequence_length = 10;  % You can adjust this as needed

% Prepare sequences of input and target data
X_sequences = buffer(X, sequence_length, sequence_length-1, 'nodelay').';
y_sequences = buffer(y, sequence_length, sequence_length-1, 'nodelay').';

% Split the dataset into training and testing sets
rng(0); % Set random seed for reproducibility
split_ratio = 0.8;
split_idx = round(split_ratio * size(X_sequences, 1));
X_train = X_sequences(1:split_idx, :);
y_train = y_sequences(1:split_idx, :);
X_test = X_sequences(split_idx+1:end, :);
y_test = y_sequences(split_idx+1:end, :);

% Reshape data for LSTM input (numObservations, numTimeSteps, numFeatures)
X_train = reshape(X_train, [1, size(X_train, 1), sequence_length]);
y_train = reshape(y_train, [1, size(y_train, 1), sequence_length]);
X_test = reshape(X_test, [1, size(X_test, 1), sequence_length]);
y_test = reshape(y_test, [1, size(y_test, 1), sequence_length]);

% Create an LSTM-based neural network
layers = [
    sequenceInputLayer(1) % Sequence input layer for single scalar input
    lstmLayer(64, 'OutputMode', 'sequence')
    lstmLayer(32, 'OutputMode', 'sequence')
    fullyConnectedLayer(1)
    regressionLayer % Regression layer
];

options = trainingOptions('adam', ...
    'MaxEpochs', 50, ...
    'MiniBatchSize', 32, ...
    'ValidationData', {X_test, y_test}, ...
    'Plots', 'training-progress', ...
    'InitialLearnRate', 0.01);

% Train the LSTM network
net = trainNetwork(X_train, y_train, layers, options);

% Use the trained model to estimate Q signals from new I signals
new_I_signals = cos(2 * pi * signal_frequency * time); % Replace with your new I signals
new_X_sequences = buffer(new_I_signals, sequence_length, sequence_length-1, 'nodelay').';
new_X_sequences = reshape(new_X_sequences, [1, size(new_X_sequences, 1), sequence_length]);

estimated_Q_signals = predict(net, new_X_sequences);

% Plot the original Q signals and estimated Q signals
figure;
plot(time(2:end), Q_signal(2:end), '--', 'DisplayName', 'Original Q Signal');
hold on;
plot(time(2:end), estimated_Q_signals(2:end), 'DisplayName', 'Estimated Q Signal');
xlabel('Time');
ylabel('Amplitude');
legend('Location', 'best');
title('Q Signal Estimation with LSTM');
grid on;

I am using this code for regression problem. But I am getting Error again and again. Any help or what should I do?

Error using trainNetwork (line 184)

The training sequences are of feature dimension 1 792 but the input layer expects sequences of

feature dimension 1.

Error in Aws_LSTM_PhaseMeasure (line 51)

net = trainNetwork(X_train, y_train, layers, options);

Answer 1

[Matt J]

fn=@(z) reshape( num2cell(permute(z,[1,3,2]),2)  ,[],1);
X_train=fn(X_train);
y_train=fn(y_train);
X_test=fn(X_test);
y_test=fn(y_test);

options = trainingOptions('adam', ...
    'MaxEpochs', 50, ...
    'MiniBatchSize', 32, ...
    'ValidationData', {X_test, y_test}, ...
    'Plots', 'training-progress', ...
    'InitialLearnRate', 0.01);

% Train the LSTM network

net = trainNetwork(X_train, y_train, layers, options);

[Leostar90]

although it worked but after training I got this error.

Error using DAGNetwork/predict (line 166)

The prediction sequences are of feature dimension 1 991 but the input layer expects sequences

of feature dimension 1.

Error in SeriesNetwork/predict (line 279)

Y = this.UnderlyingDAGNetwork.predict(X, varargin{:});

Error in Aws_LSTM_PhaseMeasure (line 64)

estimated_Q_signals = predict(net, new_X_sequences);

[Matt J]

Put X in the same format as X_train.

K>> X_train

X_train =

792×1 cell array

{[ 1 0.9980 0.9921 0.9823 0.9685 0.9510 0.9296 0.9046 0.8761 0.8440]}

{[ 0.9980 0.9921 0.9823 0.9685 0.9510 0.9296 0.9046 0.8761 0.8440 0.8086]}

{[ 0.9921 0.9823 0.9685 0.9510 0.9296 0.9046 0.8761 0.8440 0.8086 0.7701]}

{[ 0.9823 0.9685 0.9510 0.9296 0.9046 0.8761 0.8440 0.8086 0.7701 0.7285]}

{[ 0.9685 0.9510 0.9296 0.9046 0.8761 0.8440 0.8086 0.7701 0.7285 0.6840]}

{[ 0.9510 0.9296 0.9046 0.8761 0.8440 0.8086 0.7701 0.7285 0.6840 0.6367]}

{[ 0.9296 0.9046 0.8761 0.8440 0.8086 0.7701 0.7285 0.6840 0.6367 0.5870]}

{[ 0.9046 0.8761 0.8440 0.8086 0.7701 0.7285 0.6840 0.6367 0.5870 0.5350]}

{[ 0.8761 0.8440 0.8086 0.7701 0.7285 0.6840 0.6367 0.5870 0.5350 0.4808]}

{[ 0.8440 0.8086 0.7701 0.7285 0.6840 0.6367 0.5870 0.5350 0.4808 0.4248]}

{[ 0.8086 0.7701 0.7285 0.6840 0.6367 0.5870 0.5350 0.4808 0.4248 0.3670]}

...

[Leostar90]

You mean I should make changes in here?

X = I_signal(1:end-1); % I signal as input

y = Q_signal(2:end); % Corresponding Q signal as target output

Like this

fn=@(z) reshape( num2cell(permute(z,[1,3,2]),2) ,[],1);

X=fn(X);

y=fn(y);

[Matt J]

Yes. Did you try it?

[Leostar90]

yes but there was an error.

Error using buffer

Signal X must be numeric, logical, or a character array.

Error in Aws_LSTM_PhaseMeasure (line 20)

X_sequences = buffer(X, sequence_length, sequence_length-1, 'nodelay').';

[Matt J]

The reformatting fn(X) was to meet the requirements of predict(). Other commands like buffer() may require a different format.

[Leostar90]

I am trying to find format for buffer on google but failed.

[Matt J]

I get no errors with,

fn=@(z) reshape( num2cell(permute(z,[1,3,2]),2)  ,[],1);
X_train=fn(X_train);
y_train=fn(y_train);
X_test=fn(X_test);
y_test=fn(y_test);

options = trainingOptions('adam', ...
    'MaxEpochs', 50, ...
    'MiniBatchSize', 32, ...
    'ValidationData', {X_test, y_test}, ...
    'Plots', 'training-progress', ...
    'InitialLearnRate', 0.01);

% Train the LSTM network

net = trainNetwork(X_train, y_train, layers, options);

% Use the trained model to estimate Q signals from new I signals
new_I_signals = cos(2 * pi * signal_frequency * time); % Replace with your new I signals
new_X_sequences = buffer(new_I_signals, sequence_length, sequence_length-1, 'nodelay').';
new_X_sequences = reshape(new_X_sequences, [1, size(new_X_sequences, 1), sequence_length]);

estimated_Q_signals = predict(net, fn(new_X_sequences));

[Leostar90]

Problem is here actually.

Q_signal is 1000x1 double

and

estimated_Q_signals is 991x1 cell

and every cell is 1x10 row vector.

so I got this error.

Error using plot

Invalid data argument.

Error in Aws_LSTM_PhaseMeasure (line 69)

plot(time(2:end), estimated_Q_signals(2:end), 'DisplayName', 'Estimated Q Signal');

% Use the trained model to estimate Q signals from new I signals

new_I_signals = cos(2 * pi * signal_frequency * time); % Replace with your new I signals

new_X_sequences = buffer(new_I_signals, sequence_length, sequence_length-1, 'nodelay').';

new_X_sequences = reshape(new_X_sequences, [1, size(new_X_sequences, 1), sequence_length]);

estimated_Q_signals = predict(net, fn(new_X_sequences));

% Plot the original Q signals and estimated Q signals

figure;

plot(time(2:end), Q_signal(2:end), '--', 'DisplayName', 'Original Q Signal');

hold on;

plot(time(2:end), estimated_Q_signals(2:end), 'DisplayName', 'Estimated Q Signal');

[Matt J]

It's not clear why that's a problem. Did you expect 991 estimates or not? Isn't the sequence length supposed to be 10? That's what you said in your code.

% Define sequence length (number of time steps in each sequence)
sequence_length = 10;  % You can adjust this as needed

[Leostar90]

Actually, size of estimated_Q_signals and Q-signal should be same ... in my case size of Q-signal is 1000x1 double.

plot(time(2:end), Q_signal(2:end), '--', 'DisplayName', 'Original Q Signal');

hold on;

plot(time(2:end), estimated_Q_signals(2:end), 'DisplayName', 'Estimated Q Signal');

y = Q_signal(2:end); % Corresponding Q signal as target output

[Matt J]

Actually, size of estimated_Q_signals and Q-signal should be same ... in my case size of Q-signal is 1000x1 double.

Well, that can't happen because new_X_sequences is not that size,

K>> whos new_X_sequences

Name Size Bytes Class Attributes

new_X_sequences 1x991x10 79280 double

[Leostar90]

I think I chose wrong way to solve my problem ..... seems like I should solve with

"Sequence-to-One Regression Using Deep Learning". I have to predict one signal from other signal ....and for training I have synthetic data x signal and y signal (targets) .... I am new to MATLAB Deep Learning so I think for training we should pass sequences to model.

After training when I should pass one signal as input to PREDICT function and it should result in same diemension of y-signal (target) in my problem Q-signal 1000x1 double.So I think "Sequence to one regression using deep learning in MATLAB could work here.