Is it possible to divide data from a very long time series into smaller pieces so that an LSTM network can be trained more easily?

Hi all,

I decided to use LSTM neural networks to solve a problem of identification of a nonlinear system.

In particular, the latter is a SISO process and as training data I have available very long time series (more than 400000 samples) both for the input and the output of the process.

I wonder, in this regard, if there is a function in trainingOptions (or any other methodology) that allows you to partition these time series into smaller parts (with fewer samples), in order to facilitate the training process of LSTM.

I thank you in advance for your support.

Marco

Answer 1

[aditi bagora]

You can utilize the 'MiniBatchSize' parameter in the “trainingOptions” function to define the desired size of mini batches. Batching will partition your data into smaller parts and can be used in training your network.

To know more details, refer to the below link: https://www.mathworks.com/help/deeplearning/ref/trainingoptions.html#bu59f0q_sep_bu59f0q-MiniBatchSize

I hope the provided information helps in resolving your query.