Let's say I'm predicting weather, and I want to use 7 days of weather data (made up of 5 parameters) to predict the next day's temperature. So each training batch has a sequence of 7 timesteps (7 days of weather data) to make up the X data, i.e.:
[batch_size, 7, 5]
For the Y data I assume I provide just 1 value (the 8th day) for each sequence of the batch, i.e.:
[batch_size, 1]
?
To answer my own question: It depends on what I am trying to model. If I am interested in inputting 7 timesteps to predict the weather on the 8th day that is classed as a many-to-one modelling, and my proposed Y values are correct for that scenario.
The case where you may have many-to-many is language translation where each word in the input sequence is translated to corresponding words in the output.