raoyongming / HorNet

[NeurIPS 2022] HorNet: Efficient High-Order Spatial Interactions with Recursive Gated Convolutions

Home Page:https://hornet.ivg-research.xyz/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

使用位置

jianglanlele opened this issue · comments

请问,gnconv可以用在3d unet的skip connection部分嘛,或者我应该加在网络的那一部分,谢谢

Hi, 感谢对我们工作的关注。gnconv可以用来替换模型中的包含空间建模的模块,比如3x3卷积,self-attention层等,我们之前的尝试都能有一个比较显著的提点。如果你指的skip connection中本身就包含空间卷积,那我觉得应该是可以用的。

Hi, 感谢对我们工作的关注。gnconv可以用来替换模型中的包含空间建模的模块,比如3x3卷积,self-attention层等,我们之前的尝试都能有一个比较显著的提点。如果你指的skip connection中本身就包含空间卷积,那我觉得应该是可以用的。

感谢您的回复,我还是不太理解,就是如果是最基础的unet网络,我该把gnconv放在网络的哪个或那些位置,以下图中的为例,希望可以得到解答,感谢
image

如果是这样的架构的话,比较好的使用方式是应该替代模型中的3x3 conv

@raoyongming Excellent work Yongming! I have a relevant question here. If I were to replace regular 3x3 convs in a Unet-like arch, should I replace all convs with gnconvs? I guess gnconv adds a bit of overhead that a regular conv so replacing all convs may not be a good idea. Which place in Unet do you think gnconv would play a key role, if I just want to change a few places?

Hi @liyubrt, interesting question! Replacing 3x3 convs in deep layers (lower spatial resolution, more channels) may be more helpful. In our design for HorFPN, we also use gnconv to fuse features from different stages, which may also be useful in U-Net.

@raoyongming Thank you for your advise. I will definitely try it out.