Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

Hi <@U0512USU2KF>! In every training iteration the data loader samples `batch_size` time-series from the dataset. Then, the model samples `windows_batch_size` individual windows (of size `input_size` + `h`) from the time-series in the batch.

Usually `windows_batch_size` is more important to tune since it ultimately determines the size of the batch for the model for gradient computation.

So if I have window batch size of 128 per say it means it has 128 windows in each batch right