SRM net

[EvgenyKashin]

Hi, cool repository! Can you consider adding to comparison SRM net from "SRM : A Style-based Recalibration Module for Convolutional Neural Networks" (https://arxiv.org/abs/1903.10829) which looks very promising. You can use SRM block implementation from this repository https://github.com/EvgenyKashin/SRMnet

[implus]

Thanks for suggestions! We will work on it later.

[hyunjaelee410]

Hi, thank you for sharing your work!

When training SRMNet on ImageNet, I found it useful not to apply weight decay on SRM parameters. Since SRM module learns to how to utilize mean and std statistics, it is better not to impose weight decay regularization.

You can refer to https://github.com/hyunjaelee410/style-based-recalibration-module/blob/master/imagenet.py#L327 https://github.com/hyunjaelee410/style-based-recalibration-module/blob/master/models/recalibration_modules.py#L17-L19

and @EvgenyKashin thanks for your interest in our work and sharing your implementation!

[EvgenyKashin]

@hyunjaelee410 it is very cool insight about weight decay. Thank you for sharing official implementation of your work!

[shiyuanyin]

@implus Hi, great repository!
I'd like to ask some questions. Have you tried other backbone networks as a comparison? I want to know its universality and apply to other network? for example, the SEnet can be used in many scenarios and networks