Giters

kkhoot / PAA

A PyTorch implementation of the paper `Probabilistic Anchor Assignment with IoU Prediction for Object Detection` ECCV 2020 (https://arxiv.org/abs/2007.08103)

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Loss weight

YangHai-1218 opened this issue · comments

commented

Hi, thanks for your work. But I wonder if the result is sensitive to the loss weight (1.0, 1.3, 0.5). As we know, in ATSS, the loss weight is (1.0, 2.0, 1.0) and in FCOS, the loss weight is (1.0, 1.0, 1.0). Have you tried other loss weight combinations？ Thanks!

commented

Hi @NaCl-Ocean , thank you for your interest. The loss weights were set to make the magnitudes of the three losses similar when they are converged. I also tested the weights (1.0, 1.0, 1.0), and I don't remember exactly but the results on Res50 were similar. There may be better loss weights that would give better results, but I don't think it significantly affects the model performance.

commented

Thanks for your reply. I tested (1.0, 2.0, 1.0)(ATSS) loss weight on PAA Res50 by mmdetection, and got 40.1 mAP. Maybe the loss weight will slightly affect the model performance. It may also due to be unstable performance under 1x setting.

commented

Thank you for sharing your result. Yes I think you are right. The loss weights can slightly affect the model performance (if not, we don't need to set the weights :)). It is also correct that the 1x setting doesn't make the training converged, so there are fluctuations in the results.