Mmdetection: How to use the custom optimizer?

[sunjie97]

I want to customize the optimizer(RAdam etc.), but I find that the optimizer in mmdetection is imported from torch.optim(only a few of optimizer are implemented), can you provide an official example to do this? Thank you

[hellock]

We will introduce an optimizer registry like dataset or model registry to support customizing the optimizer. As a walkaround now, you may just modify this line and get your own optimizer class.

[sunjie97]

OK, I'll try, thanks

[hiyyg]

It seems the custom optimizer documentation is out of date. How to use custom optimizer after the cleaning of optimizer code?

[hiyyg]

I figured out using this way:

from mmcv.runner.optimizer import build_optimizer, OPTIMIZERS
from my_optimizer import MyOptimizer
OPTIMIZERS.register_module()(MyOptimizer)

optim_cfg = dict(type="MyOptimizer", lr=0.001)
optimizer = build_optimizer(model, optim_cfg)