[Question] Batch size, number samples, number Timeseries

[leyp1]

Hi,

I have a couple of questions and would be thankful for any aid to help me understand what is going on behind the scences.

From what I read on the NBEATS model description (https://unit8co.github.io/darts/generated_api/darts.models.forecasting.nbeats.html?highlight=batch%20size) the batch size changes the amount of samples in one batch however changes nothing about the number of Timeseries in each batch, is this correct? If yes, then is there a parameter that changes the number of Timeseries in a batch, since at the moment from what I have seen it always seems to take all of the Timeseries when calling fit(...).

Additonally I have noticed that there is a difference between passing a list of Timeseries to fit(...) as the training set or one multivariate Timeseries, but I do not understand why. The amount of batches changes when I do this.

[hrzn]

Hi, There are a few concepts to untangle in your question.

• training samples: the number of training samples (number of input/output examples used to train machine learning models in a supervised fashion) will depend on the number of TimeSeries you give to fit() as well as the input/output chunk lengths.
• batch sizes: the number of training samples is not impacted by the batch size. The batch size determines the number of training samples that are consumed in a given gradient update step when training the model. For instance your series are such that you have 1000 training samples and your batch size is 10, one epoch will result in 100 gradient updates.
• prediction samples: when using probabilistic forecasting models, you can provide num_samples to predict(). This will determine the number of Monte Carlo samples that will be used to describe the resulting stochastic TimeSeries(s) returned by predict().
• Difference between training on multiple TimeSeries -vs- one multivariate TimeSeries: one TimeSeries always has the same time axis, so the different components of a multivariate series always share the same time steps. Furthermore, a model trained on multivariate series will learn to capture all the components jointly. In contrast, training on multiple TimeSeries allows to train on series not necessarily sharing the same axes. Multiple TimeSeries will result in more input/output sub-slices training examples compared to stacking all the dimensions in a single series and slicing them all together.

Hope this helps.

[leyp1]

Yes ok, this makes sense. Thank you for clarifying and answering so quickly!