How to use multiple data in LSTM?

In this example https://www.mathworks.com/help/deeplearning/ug/time-series-forecasting-using-deep-learning.html,

1. Discription said that is 'For single time step predictions, use the CPU.' I wonder how to do multi time step pridiction in Matlab.
2. Espacially, I would like to know about the way to use multi training data set for LSTM, not single training data set like this example.

That example used a double data(1xN), but I hope to enter multiple(M) double data(like MxN).

Please let me know some idea or give your knowledge.

Answer 1

[Pratyush Roy]

Hi Daerno.

The example mentioned in the question is used for finding temporal relation between 1-D input and 1 dimensional output. As mentioned in the code:

numFeatures = 1;
numResponses = 1;
numHiddenUnits = 200;

layers = [ ...
    sequenceInputLayer(numFeatures)
    lstmLayer(numHiddenUnits)
    fullyConnectedLayer(numResponses)
    regressionLayer];

The number of features is 1. Hence the data passed as input is 1-dimensional in nature.

In general, LSTMs are built to work for multi-dimensional data. We can change the numFeatures and numResponses value to map one single/multi-dimensional vector to another single/multi-dimensional vector. This doc link captures a example involving multi-dimensional vectors.

Hope this helps!

[daemo Lee]

First of all, thank you for helping me.

Unfortunately, my fundamental problem is I would like to use a dataset for just about one feature(numFeatures=1).

For example, if I want to predict the price of stock I could train the model through the below code

and assume X is the only feature like the price of the stock for 1~9 timestep(month) and Y is the price of the stock for 2~10 timestep(month).

XTrain = [1,2,3,4,5,6,7,8,9]
YTrain = [2,3,4,5,6,7,8,9,10]
[net,info] = trainNetwork(XTrain,YTrain,layers,options);

but I hope to know how to train multi-data sets (3 stock data & numFeatures=1) simultaneously when I take multi same feature like below.

XTrain1 = [1,2,3,4,5,6,7,8,9]
YTrain1 = [2,3,4,5,6,7,8,9,10]

XTrain2 = [2,2,4,5,6,8,10,11,12]
YTrain2 = [2,4,5,6,8,10,11,12,15]

XTrain3 = [5,3,2,1,1,1,6,9,10]
YTrain3 = [3,2,1,1,1,6,9,10,15]

[net,info] = trainNetwork(XTrain,YTrain,layers,options);

Please reply.

[Pratyush Roy]

Hi Daemo,

Since you have multiple datasets, you can train multiple LSTMs in parallel. Please refer to the doc link below for more details:

https://in.mathworks.com/help/deeplearning/ug/train-deep-learning-networks-in-parallel.html

Hope this helps!