clovaai / ECLIPSE

(CVPR 2024) ECLIPSE: Efficient Continual Learning in Panoptic Segmentation with Visual Prompt Tuning
Other
32 stars 4 forks source link

delta settings are different in training and evaluation #5

Closed wendyhuang2021 closed 2 months ago

wendyhuang2021 commented 2 months ago

Hi! I found delta used in training and evaluation sometimes sat different. Such as delta in ade_ps/100_5.sh, ade_ps/100_50.sh.

May I ask why? Different delta settings would result in very different accuracy between the last training step and the evaluation.

qjadud1994 commented 2 months ago

As explained in our paper, the delta is a post-processing parameter required to be tuned after training. Also, since the delta directly manipulates the output logits, it should be carefully set according to each scenario.

image

wendyhuang2021 commented 2 months ago

Thanks for your reply!