Clear Filters
Clear Filters

How to combine multiple net in LSTM

16 views (last 30 days)
Luc Xuan Bui
Luc Xuan Bui on 8 Apr 2024
Commented: Luc Xuan Bui on 14 Apr 2024
I intend to train three sequences using LSTM, then combine them into one 'net' for prediction to speed up the training process. However, I'm facing difficulties in achieving this.

Answers (1)

Ben
Ben on 9 Apr 2024
You can combine 3 separate LSTM-s into one network by adding them to a dlnetwork object and hooking up the outputs. Note that if the LSTM-s have OutputMode="sequence" then either you need all input sequences to have the same length, or have some layer(s) that can manage the data with different sequence lengths.
Here's an example with OutputMode="last"
inputSizes = [1,2,3];
outputSize = 4;
lstmHiddenSize = 5;
hiddenSize = 10;
sequenceLengths = [6,7,8];
x1 = dlarray(rand(inputSizes(1),sequenceLengths(1)),"CT");
x2 = dlarray(rand(inputSizes(2),sequenceLengths(2)),"CT");
x3 = dlarray(rand(inputSizes(3),sequenceLengths(3)),"CT");
layers = [
sequenceInputLayer(inputSizes(1))
lstmLayer(lstmHiddenSize,OutputMode="last")
concatenationLayer(1,3,Name="cat")
fullyConnectedLayer(hiddenSize)
reluLayer
fullyConnectedLayer(outputSize)];
net = dlnetwork(layers,Initialize=false);
net = addLayers(net,[sequenceInputLayer(inputSizes(2));lstmLayer(lstmHiddenSize,OutputMode="last",Name="lstm2")]);
net = addLayers(net,[sequenceInputLayer(inputSizes(3));lstmLayer(lstmHiddenSize,OutputMode="last",Name="lstm3")]);
net = connectLayers(net,"lstm2","cat/in2");
net = connectLayers(net,"lstm3","cat/in3");
net = initialize(net);
y = predict(net,x1,x2,x3)
  1 Comment
Luc Xuan Bui
Luc Xuan Bui on 14 Apr 2024
Sorry for making it seem like I didn't state the question clearly. I have a data string 0-t, but I want to split this string into 3 strings 0-t1, t1-t2, t2-t. Then I train these 3 sequences with 3 different parameters on the LSTM network, but I want to make sure the output is only 1. That is, I have 3 inputs separated from 1 continuous sequence, the first 3 nets. output after training, then combine the information learned from these 3 nets into 1 for prediction. Your method is 3 inputs predicting 3 outputs. Sorry for this misunderstanding.

Sign in to comment.

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!