Hey guys, I have two questions, [1] What's the difference between the '`batch_size`' and '`window_batch_size`' in the configuration? I know the batch_size but does know what's the 'window_batch_size' is used for.. [2] In the document saying the '`fit`' is iterative sliding windows (by adjusting step_size), I wonder if the '`predict`' can have outputs just like the '`cross_validation`' does. Because I do not only want the last predicted sequence but also want every predicted sequence.

Hi @王梦石!

batch_size

corresponds to the number of time series in the batch, and

windows_batch_size

to the number of windows in the batch. During each training iteration, the loader first sample

batch_size

series, then the model selects

windows_batch_size

windows at random starting at different timestamps.

Thank you very much Cristian!, Also, a quick question, suppose I want to implement some other models (such as TiDE) with neuralforecast framework. What I need to to is to write the 'Model' and 'AutoModel' inherit from the exisitng 'BaseWindow' and 'BaseAuto'. then theoretically it should work?

Hey. We have a guide for this. Please let us know if something is missing from it

And also this one can be useful: https://nixtla.github.io/neuralforecast/examples/neuralforecast_map.html