LSTM with vector as output for multi step ahead forecasting

Question

Alexander Stein 2023-4-19

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1949623-lstm-with-vector-as-output-for-multi-step-ahead-forecasting

评论： Alexander Stein 2023-4-25

采纳的回答： Katja Mogalle

在 MATLAB Online 中打开

Hello,

I would like to build a LSTM network that outputs a vector only on the last step (Outputmode = last).

Input-data: Sequence with a fixed size of [4,1344] (4 features, 1344 steps).

Output-data: Vector of size [96,1] (Output on last step)

My attempt so far:

numFeatures = 4;
numResponses = size(YTrain{1},1);
% Input layer
layers = sequenceInputLayer(numFeatures);
% variation of layers and hidden units
for i = 1:LSTMDepth-1
    layers = [layers;lstmLayer(numHiddenUnits,OutputMode="sequence")];
end
layers = [layers;lstmLayer(numHiddenUnits,OutputMode="last")];
% Output layers
layers = [ layers 
fullyConnectedLayer(numResponses)
regressionLayer];
% training options
maxEpochs = 300;
miniBatchSize = 20;
options = trainingOptions("adam", ...
ExecutionEnvironment="auto", ...
MaxEpochs=maxEpochs, ...
MiniBatchSize=miniBatchSize, ...
ValidationData={XValidation,YValidation}, ...
ValidationFrequency=30, ...
InitialLearnRate=params.InitialLearnRate, ...
LearnRateDropFactor=0.2, ...
LearnRateDropPeriod=15, ...
GradientThreshold=1, ...
Shuffle="never", ...
Verbose=true);
% Training: XTrain and YTrain are cell arrays
net = trainNetwork(XTrain,YTrain,layers,options);

Can someone help me how to build such a network?

Thanks in advance

2 个评论
显示无隐藏无

KSSV 2023-4-19

Attach your data.

Alexander Stein 2023-4-19

I attached it to the original post

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Katja Mogalle 2023-4-25

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1949623-lstm-with-vector-as-output-for-multi-step-ahead-forecasting#answer_1222664

Hello Alexander,

The network structure generally looks fine to me.

When I run your code I get the following error, is it the same as you're getting?

Error using trainNetwork

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.

To fix this error, you need to slightly reorganize your response data. Currently it's a NumObservations-by-1 cell array of NumResponses-by-1 matrices. However, you need to convert this to a large matrix of size NumObservations-by-NumResponses. To do this, you can use the following line of code:

>> YTrain = cat(2,YTrain{:})';

If you need more information, you can also look at the documentation page of trainNetwork: https://uk.mathworks.com/help/deeplearning/ref/trainnetwork.html?s_tid=doc_ta#mw_d0b3a2e4-09a0-42f9-a273-2bb25956fe66 (the link points to the explanation of the response data formats)

I hope this helps.

Katja