网络参数重组经典论文

1. RepVGG

RepVGG: Making VGG-style ConvNets Great Again详解
提出了一种替代卷积核：使用3x3conv+1x1conv+identity代替原本单一的3x3conv

2. ACNet

ACNet: Strengthening the Kernel Skeletons for Powerful CNN via Asymmetric Convolution Blocks详解
提出了一种替代卷积核：使用1x3conv+3x1conv+3x3conv代替原本单一的3x3conv

3. DBB

Diverse Branch Block: Building a Convolution as an Inception-like Unit详解
总结了6种可以进行参数重组的变换：

提出了一种替代卷积核：使用1x1conv + (1x1conv+kxkconv) + (1x1conv+AVG) + KxKconv代替原本单一的KxKconv

4. RepMLP

Re-parameterizing Convolutions into Fully-connected Layers for Image Recognition详解
提出了如下所示的MLP结构：

5. SpineNet-Seg

ResRep: Lossless CNN Pruning via Decoupling Remembering and Forgetting详解
提出了如下所示的无损剪枝结构

6. RepNAS

RepNAS: Searching for Efficient Re-parameterizing Blocks将参数重组与网络结构搜索相结合。