Closed mo-kli closed 1 year ago
Thank you for your comment. Actually, this is intentional because in training we can access the ground truth and we assume the scorer gives the correct mode closest to the ground truth, but in testing we take the best mode from the trained scorer instead of "cheating". However, your comment is undoubtedly reasonable and you can modify the code to your idea.
There might be a bug in
select_future
https://github.com/MCZhi/DIPP/blob/main/utils/train_utils.py#L68. This function uses the global variablebest_mode
, which is computed earlier inMFMA_loss
using the ground truth. However,select_future
is also used by the planner to selectplan_init
and the predicition mode considered during planning. This looks like cheating, insteadscores
or the reference path should be used in my opinion byselect_future
.