[Feature]training timm model with frozen_stages and norm eval tricks

[Bo396543018]

I found use frozen_stages&norm eval may improve transfer learning performance.
Currently，the timm wrapper not support these tricks.

[Ezra-Yu]

Yes, they do not implement related functions. Every backbone has a different implement. Currently, there is no way to modify this in general.

what backbone do you want to use?

[Bo396543018]

I use timm's resnet50 model to compare 21k &1k pretrained improvement.

[mzr1996]

So why don't just use mmcls ResNet implementation

[Bo396543018]

Use timm warpper could avoid the problem for 21k pretrained weights inconsistent naming.

[mzr1996]

Well, the freeze stages function depends on the detailed network and we cannot provide a unified API except timm provides it.

[Bo396543018]

Ok，Thank you for your reply，I'll try to convert weights for comparison.

[Ezra-Yu]

@Bo396543018 We are considering adding relevant pre-trained models to mmcls, and we will thank you if you complete the relevant work, put relevant weights or scripts in mmcls.

[Bo396543018]

@ Ezra-Yu OK，When I complete，I'll share the converted weights