Why is learning rate in R50 logs half of the config lr?

[NaomiEX]

When looking at your R50 logs and configs, I noticed that the lr in the logs is 3e-4 after warmup:

whereas in the config the lr should be 6e-4:

Is this the intended behaviour? This doesn't show up in the R101 config and I was wondering why it behaves like this

[NaomiEX]

Nevermind, it turns out that the TextLoggerHook prints out the lr of the first param group, which happens to be the img_backbone in this case, and its lr is halved.

Although as a suggestion to the authors, you could change the order of param groups, like create self.head or self.neck first in sparse4d.py instead of self.img_backbone to avoid this confusion as I spent an hour trying to figure out why my lr was getting halved

[linxuewu]

Nevermind, it turns out that the TextLoggerHook prints out the lr of the first param group, which happens to be the img_backbone in this case, and its lr is halved.

Although as a suggestion to the authors, you could change the order of param groups, like create self.head or self.neck first in sparse4d.py instead of self.img_backbone to avoid this confusion as I spent an hour trying to figure out why my lr was getting halved

This is a design flaw in mmcv, changing the model order cannot fundamentally solve the problem, for example, when lr scale is also set for the head or neck.