huawei-noah / Efficient-AI-Backbones

Efficient AI Backbones including GhostNet, TNT and MLP, developed by Huawei Noah's Ark Lab.
4.05k stars 707 forks source link

visualization in VIG #131

Open touristourist opened 2 years ago

touristourist commented 2 years ago

Hi, in the visualization part of VIG, I noticed that the neighbors are different for the same node in the 1st and the 12th block. Does this mean the adjacency matrix different between different blocks? As a beginner in GNN, this puzzles me. Looking for your reply, thanks! image

iamhankai commented 2 years ago

Thanks for the attention. The graph is dynamically generated in grapher module. By default, we use MRConv in this repo: https://github.com/lightaime/deep_gcns_torch.

Idiom1999 commented 2 years ago

我想问一下,源码里DeepGCN类中:HW = 224 // 4 * 224// 4的4是不是每个patch的边长为4,然后有HW=3136个patch的意思?

iamhankai commented 2 years ago

是3136个patch,不过stem是几层卷积,每个patch之间有一定overlap,边长为7

Idiom1999 commented 2 years ago

真的十分感谢您的解答!!! 我看到有DownSample进行了下采样操作,具体是使用大小为3,步长为2,padding=1的卷积核进行完成的。 这样做在CNN层面上是2倍下采样。那么对于整个图来讲,节点数就减少了4倍。这是以一种什么方式进行节点数的减少呢?就好比GraphUNet是采用topK的方法进行图的下采样,那么DownSample对于节点数的减少又是什么原理呢?

iamhankai commented 2 years ago

Downsample这里没有太多特殊设计,就是参考CNN的

Idiom1999 commented 2 years ago

非常感谢您这段时间的答复! 请问您有了解过Slic超像素分割吗?对基于Slic超像素分割作为初始节点嵌入来做图像语义分割,有什么推荐的思路和方法吗?

shubhrsingh22 commented 2 years ago

Hi,

I was wondering if you could guide me to the code for visualization. I have been trying to find it but couldnt.

iamhankai commented 2 years ago

非常感谢您这段时间的答复! 请问您有了解过Slic超像素分割吗?对基于Slic超像素分割作为初始节点嵌入来做图像语义分割,有什么推荐的思路和方法吗?

你这个思路挺好的,建议继续深挖下去!