how to build one to many and many to many LSTM in matlab?

I don't know that is possible to build these kind of networks in matlab? If it is possible, can I use these kind of network in simulink?

Answer 1

[Yash Sharma]

To build one-to-many and many-to-many LSTM models in MATLAB, you can use the Deep Learning Toolbox. You need to change number of input features and number of output features to build different LSTM architectures. For example:

One-to-Many LSTM:

In a one-to-many LSTM, a single input sequence is used to predict multiple outputs.

numFeatures = 1;  % Single input feature
numResponses = 10;  % Number of output time steps
numHiddenUnits = 100;

% Define the network architecture
layers = [
    sequenceInputLayer(numFeatures)
    lstmLayer(numHiddenUnits, 'OutputMode', 'sequence')
    fullyConnectedLayer(numResponses)
    regressionLayer];

Many-to-Many LSTM:

In a many-to-many LSTM, multiple input sequences are used to predict multiple output sequences.

numFeatures = 5;  % Number of input features
numResponses = 5;  % Number of output features
numHiddenUnits = 100;
% Define the network architecture
layers = [
sequenceInputLayer(numFeatures)
lstmLayer(numHiddenUnits, 'OutputMode', 'sequence')
fullyConnectedLayer(numResponses)
regressionLayer];

You can make other LSTM architectures similarly.

Hope it helps!

Answer 2

[Krishna]

Hi,

I understand that you're interested in building various sequence-to-sequence models using LSTM networks. This can be effectively achieved in MATLAB using dlnetworks. Please refer to the following documentation to learn more about 'dlnetworks':

https://in.mathworks.com/help/deeplearning/ref/dlnetwork.html

Additionally, you can go through the following example to learn more about using dlnetworks for sequence modeling and to get started.

https://in.mathworks.com/help/deeplearning/ug/time-series-forecasting-using-deep-learning.html

To build many to one architecture you have to set the property of 'lstmLayer' as 'last'. To build this type of architecture, please go through the following documentation,

https://in.mathworks.com/help/deeplearning/ug/classify-sequence-data-using-lstm-networks.html

To build many to many architectures you have to set the property of 'lstmLayer' as 'sequence'. To build this type of architecture, please go through the following documentation,

https://in.mathworks.com/help/deeplearning/ug/sequence-to-sequence-classification-using-deep-learning.html

Now I don't think there is any point of building one to one architecture as it is just Lstm network with 1 time step. It is same as feed forward network. To build this just use 'fullyConnectedLayer' instead of 'lstmLayer'. Go through this documentation to learn more regarding ‘fullyConnectedLayer’,

https://in.mathworks.com/help/deeplearning/ref/nnet.cnn.layer.fullyconnectedlayer.html

Creating architectures for one-to-many tasks can be challenging. In this type of setup, the model receives a single input and must generate an entire sequence based on that input. Typically, the initial input is provided first, and then the complete sequence is generated iteratively using the model's own outputs as subsequent inputs. To handle this, you need to define a custom training loop where the network's output is fed back as input during training. This iterative process forms the basis of training for such tasks in MATLAB.

Please go through the following documentation to learn more regarding custom training loop,

https://in.mathworks.com/help/deeplearning/ug/train-network-using-custom-training-loop.html

Hope this helps.