Gemaxis / GAT

MIT License
1 stars 1 forks source link

关于FC-HAT #1

Closed hetang-wang closed 6 months ago

hetang-wang commented 6 months ago

抱歉,打扰了,没有其他渠道能够沟通。请问你最后解决了FC-HAT git lfs文件的问题了吗?fork是不是也没有办法获取到?同研究生,期待你的回复呀~

Gemaxis commented 6 months ago

不好意思,没有解决,我的研究方向是 dMRI & HyperGraph,我当时是想用他的批处理超图的方法,最后自己写了dataLoader.

hetang-wang commented 6 months ago

不好意思,没有解决,我的研究方向是 dMRI & HyperGraph,我当时是想用他的批处理超图的方法,最后自己写了dataLoader.

好的,感谢你的回复~

hetang-wang commented 6 months ago

抱歉,又打扰你了,FC-HAT的train.py文件中的内容train_loader = Data.DataLoader(data_loader.brains_loader(idx_train), batch_size=1, shuffle=True,num_workers=0, pin_memory=True, drop_last=False)似乎表明了作者将batch设置为1了,这样网络的性能应该不太稳定。我想请教一下有没有可以参考的超图批处理方法?另外,一次对一张图进行分类,计算损失函数但并不反向传播,累积例如32张图的32个损失函数求平均进行反向传播和优化器的step的方法你觉得会是有效的吗?

Gemaxis commented 6 months ago

Sorry, I couldn't find any code in the hypergraph related to batching, I solved it as you said, which was also confirmed by the author of HGNN. Based on my experiments, implementing batching has indeed accelerated the process.

hetang-wang commented 6 months ago

感谢你的回复!

Gemaxis commented 2 months ago

不好意思,我当时的回复是错误的。我今天又看了下你的回复,我是一次对一批超图进行分类,然后将loss叠加,原理是类似你所说的,如果你还对这个有兴趣或者找到了更好的方法,欢迎加我的微信m872727085一起探讨交流。

hetang-wang commented 2 months ago

我暂时没再使用这种方法了,所以也没有更好的想法