GuideWsp
GuideWsp
Yes, I had validated the two model(one is your pretrained, the other is my retrained), the accuracy of the pretrained is 81.472（same as the paper） , while my retrained is...
up to now, i have no imagenetv2 datasets. If possible, you can add my Wechat: AIWalker-Happy The accuracy of the latest retrained model is 85.16（use-ema). 
I meet same error, di you fix this bug?
> Great work！ Is the code coming soon？ you can get my reimpl from https://mp.weixin.qq.com/s/zzS16Zxc4WXbXtnH70uBnQ
我想了解一下如何将一个多分支相同通道数的网络按照你论文中算法1裁剪为多分支不同通道数的网络。 我认为该算法才是ScalaNet的核心。
谢谢反馈。我们的复现方式也是只考虑BN层中scale，对每个block中的scale按照大小排序选择。 但有一点没搞明白，那就是：每个block最终残留的总体通道数是如何确定的？这个参数直接决定了最终模型的大小。
or maybe it just a quantizer method? Just for the code, I thought it as a good Quantizer Method.