BriefGPT.xyz
Jul, 2020
WeightNet: 重新审视权重网络的设计空间
WeightNet: Revisiting the Design Space of Weight Networks
HTML
PDF
Ningning Ma, Xiangyu Zhang, Jiawei Huang, Jian Sun
TL;DR
WeightNet是一种用于生成卷积权重的框架，将当前两种最有效的SENet和CondConv方法结合在同一框架下，通过添加一个群组全连接层来实现，由此直接输出卷积权重，而不是特征空间，训练容易且节省空间，具有优异性能。
Abstract
We present a conceptually simple, flexible and effective framework for
weight generating networks
. Our approach is general that unifies two current distinct and extremely effective
senet
and
→