How to use LSTM to solve seq2seq problem in MATLAB?

I’m struggling with a seq2seq problem. That is: input 200 past values and use LSTM network to predict next 10 values at one time(winthout using closed loop forecasting).And I tried writing the code as shown below. But it will report an error: Invalid training data. For regression tasks, responses must be a vector, a matrix, or a 4-D array of real numeric responses. Responses must not contain NaNs.
inputSize = 5; % Number of input features
numTimeSteps = 200; % Number of time steps
numSamples = 100; % Number of samples

% Create random input data, a cell array of size [5, 200]
X = cell(numSamples, 1);
for i = 1:numSamples
X{i} = rand(inputSize, numTimeSteps);
end

outputSize = 1; % Number of output features for each time step
outputTimeSteps = 10; % Number of output time steps
numSamples = 100; % Number of samples

% Create random output data, a cell array of size [1, 10]
Y = cell(numSamples, 1);
for i = 1:numSamples
Y{i} = rand(outputSize, outputTimeSteps);
end

numHiddenUnits1 = 128; % Number of hidden units in the first LSTM layer
numHiddenUnits2 = 64; % Number of hidden units in the second LSTM layer
outputSize = 1; % Number of output features

layers = [
sequenceInputLayer(inputSize) % Input layer, number of input features is 5
lstmLayer(numHiddenUnits1, ‘OutputMode’, ‘last’) % First LSTM layer, outputs the last time step
fullyConnectedLayer(numHiddenUnits2) % Fully connected layer, transforms the input dimensions for the second LSTM layer
functionLayer(@(X) repmat(X, [1, outputTimeSteps]), ‘Name’, ‘replicate10’) % Expand the output of the last time step to 10 time steps
lstmLayer(numHiddenUnits2, ‘OutputMode’, ‘sequence’) % Second LSTM layer, outputs a sequence of 10 time steps
fullyConnectedLayer(outputSize) % Fully connected layer, number of output features is 1
regressionLayer]; % Regression layer, used for regression tasks

% Training options
options = trainingOptions(‘adam’, …
‘MaxEpochs’, 50, …
‘MiniBatchSize’, 32, …
‘Shuffle’, ‘every-epoch’, …
‘Verbose’, false);

% Train the network
net = trainNetwork(X, Y, layers, options);
Now I have two questions:
(1) What data format should be used for the input and output of the model during training? Cell arrays or 3D arrays?
(2) How to control the time expansion steps of the input and output for each layer (input layer, LSTM layer, fully connected layer) in an LSTM network?I’m struggling with a seq2seq problem. That is: input 200 past values and use LSTM network to predict next 10 values at one time(winthout using closed loop forecasting).And I tried writing the code as shown below. But it will report an error: Invalid training data. For regression tasks, responses must be a vector, a matrix, or a 4-D array of real numeric responses. Responses must not contain NaNs.
inputSize = 5; % Number of input features
numTimeSteps = 200; % Number of time steps
numSamples = 100; % Number of samples

% Create random input data, a cell array of size [5, 200]
X = cell(numSamples, 1);
for i = 1:numSamples
X{i} = rand(inputSize, numTimeSteps);
end

outputSize = 1; % Number of output features for each time step
outputTimeSteps = 10; % Number of output time steps
numSamples = 100; % Number of samples

% Create random output data, a cell array of size [1, 10]
Y = cell(numSamples, 1);
for i = 1:numSamples
Y{i} = rand(outputSize, outputTimeSteps);
end

numHiddenUnits1 = 128; % Number of hidden units in the first LSTM layer
numHiddenUnits2 = 64; % Number of hidden units in the second LSTM layer
outputSize = 1; % Number of output features

% Training options
options = trainingOptions(‘adam’, …
‘MaxEpochs’, 50, …
‘MiniBatchSize’, 32, …
‘Shuffle’, ‘every-epoch’, …
‘Verbose’, false);

% Train the network
net = trainNetwork(X, Y, layers, options);
Now I have two questions:
(1) What data format should be used for the input and output of the model during training? Cell arrays or 3D arrays?
(2) How to control the time expansion steps of the input and output for each layer (input layer, LSTM layer, fully connected layer) in an LSTM network? I’m struggling with a seq2seq problem. That is: input 200 past values and use LSTM network to predict next 10 values at one time(winthout using closed loop forecasting).And I tried writing the code as shown below. But it will report an error: Invalid training data. For regression tasks, responses must be a vector, a matrix, or a 4-D array of real numeric responses. Responses must not contain NaNs.
inputSize = 5; % Number of input features
numTimeSteps = 200; % Number of time steps
numSamples = 100; % Number of samples

% Create random input data, a cell array of size [5, 200]
X = cell(numSamples, 1);
for i = 1:numSamples
X{i} = rand(inputSize, numTimeSteps);
end

outputSize = 1; % Number of output features for each time step
outputTimeSteps = 10; % Number of output time steps
numSamples = 100; % Number of samples

% Create random output data, a cell array of size [1, 10]
Y = cell(numSamples, 1);
for i = 1:numSamples
Y{i} = rand(outputSize, outputTimeSteps);
end

numHiddenUnits1 = 128; % Number of hidden units in the first LSTM layer
numHiddenUnits2 = 64; % Number of hidden units in the second LSTM layer
outputSize = 1; % Number of output features

% Training options
options = trainingOptions(‘adam’, …
‘MaxEpochs’, 50, …
‘MiniBatchSize’, 32, …
‘Shuffle’, ‘every-epoch’, …
‘Verbose’, false);

Cart

Cart