Clarification on Loss Calculation at Delay Point in Exp_Informer

[trunghieuvu02]

Hi, thank you for the great work on this project!

I have a question about the loss calculation in the train function of the Exp_Informer class:

subseasonal_toolkit/subseasonal_toolkit/models/informer/exp/exp_informer.py

Line 141 in cff00fc

loss = criterion(pred[:,delay,:], true[:,delay,:])

From what I understand, the model outputs a tensor of shape [batch_size, -pred_len:, :]. However, it seems that the loss is only being computed at a single delay point rather than across the entire pred_len sequence. Could you please clarify the reasoning behind this choice?

Intuitively, I would expect the loss to be computed over the full prediction window to better capture sequence-level accuracy. Is there a specific rationale—perhaps related to the task definition or evaluation criteria—that explains why loss is only calculated at one point?

Thanks in advance for your clarification!