This message was deleted.

Hi @Weng Keong Lee 🙂. In each training iteration, the loader will first sample

batch_size

entire time series from the dataset. The model will then sample

windows_batch_size

windows starting at random timestamps with size `input_size`+`h` . The model will use the input part to predict the next

h

values.

Hi Christian, Thank you for the info, just to clarify, if I use batch_size=32, horizon=48, input_size =2x48, windows_batch_size=128, am I correct to interpret that per training iteration/time step, it will randomly sample 128 windows of (input_size+h) per batch 32 times?

No, the total number of sampled windows is 128. Those windows will be uniformly sampled from 32 random time series on your dataset (if your dataset has less time series than your batch size is will simply sample from all). So it will sample, for example, 5 windows from ts 1, 2 windows from ts 2, and so on... Think of batch size as a parameter controlling the mixing/variance of windows.