Question - Prior - Training vs Testing
Closed this issue · 2 comments
Hi @karttikeya
I noticed that during the training that a standard normal prior is assumed to compute the KL Loss. However, in the testing/evaluation, I see this hyperparameter for sigma (value set to 1.3) https://github.com/HarshayuGirase/PECNet/blob/4c342597ca4310ef4976412dd8c829ad60a5465a/utils/models.py#L127
Two questions -
a) Why this difference? i.e why not use 1.3 during training as well to compute the KL ?
b) How did you find 1.3?
Regards & thanks
Kapil
Please see Section 3.1 of the paper.
In contrast, in situations where a high number of predictions are to be generated (such as K = 20, a standard
setting on benchmarks) we propose to use σT > 1 with no truncation. We posit that this procedure allows simple adjustment of prediction diversity in favor of overall performance for different K, thereby providing a simple method of
achieving good performance in all settings without requiring any retraining.
Thanks. Missed this part.