About RealTime FCOS

[lkevinzc]

Hi there!

The FCOS-RT is really amazing! Thanks for the work!

A quick question regarding its configuration:MS_DLA_34_4x_syncbn.yaml:
The solver is set up like this:
SOLVER: STEPS: (300000, 340000) MAX_ITER: 360000

I just noticed that while 360K is 4x of 90K in the vanilla setting, 300K and 340K are not 4x of 60K and 80K.

Is there any particular reason for such two lr dropping points?

Looking forward to your reply!

Thanks!

[tianzhi0549]

@lkevinzc these changing points are chosen by following https://arxiv.org/abs/1811.08883.

[lkevinzc]

@tianzhi0549 Oh I see, "reduce the learning rate by 10× in the last 60k and last 20k iterations respectively, no matter how many total iterations"!

So I previously misunderstood the setting.

Thank you so much for pointing out!

[lkevinzc]

@lkevinzc these changing points are chosen by following https://arxiv.org/abs/1811.08883.

Hi there~ Sorry for raising new questions under this old closed issue.

I am wondering why you chose sync_bn and bn for training the RT-FCOS? Will GN work better for smaller batch size?

Thanks!

[tianzhi0549]

@lkevinzc Using GN in the backbone needs to pre-train the backbone (with GN) on ImageNet again.

[lkevinzc]

@lkevinzc Using GN in the backbone needs to pre-train the backbone (with GN) on ImageNet again.

Thank you so much for your prompt and helpful reply!