hustvl/QueryInst

which learning rate is true?

howardchina opened this issue · 2 comments

howardchina commented

Hi,

I find the latest configuration file from mmdetection official released version is different from here.
Which one is better in your experiment?

here:

QueryInst/configs/queryinst/queryinst_r50_fpn_1x_coco.py

Line 121 in a10cca0

optimizer = dict(_delete_=True, type='AdamW', lr=0.000025, weight_decay=0.0001)

or official released version from mmdetection:
https://github.com/open-mmlab/mmdetection/blob/master/configs/queryinst/queryinst_r50_fpn_1x_coco.py#L127

looking for your reply!

howardchina commented

'grad_clip' setting is also different between these two sources.

vealocia commented

Hi, @howardchina!
Thanks for your interest in our work.
This repo is the original queryinst implementation. All settings are kept the same as in our paper.
But we find that modifying learning rate and grad_clip leads to a better result under ResNet-101 so in mmdetection, we provide an augmented queryinst as you see.
Both implementations are OK but I suggest you to use the augmented one (mmdetection) for better results.

  • 👍2