graph-learn功能空缺点

[bookug]

请问现在graph-learn是否支持spark分布式训练？多机多GPU分布式训练？
是否支持TransformerConv，KGCN等较新的算法？
egograph是否可以对接pyg，或者像subgraph那样通过GSL转换数据集从而支持pyg?
训练/预测时各个点是独立训练的么，除了需要共享参数？
每个点维护自身的k跳邻居信息，这种做法是否会导致较大的内存开销？（是否有量化的统计数据）
PS: github上有独立的graph-learn仓库，这两边有区别么？会实时同步？

[sighingnow]

请问现在graph-learn是否支持spark分布式训练？

graph-learn provides efficient support for sampling (graphscope.learning.Graph()), and tensorflow is the builtin supported ML framework, and definitely other engines can be leveraged on top of graphlearn.

多机多GPU分布式训练？

Yes, distributed training is supported and multi-gpu the train depends on the framework ifself. It would require certain setup effort, IMO.

是否支持TransformerConv，KGCN等较新的算法？

GraphScope (as well as graphlearn) doesn't have those builtin algorihtms, but users definitely can implement such ones on top of current framework.

egograph是否可以对接pyg，或者像subgraph那样通过GSL转换数据集从而支持pyg?

graphlearn hasn't been integrated with pyg yet.

训练/预测时各个点是独立训练的么，除了需要共享参数？

Yes.

每个点维护自身的k跳邻居信息，这种做法是否会导致较大的内存开销？（是否有量化的统计数据）

We haven't profiling yet and cannot tell such numbers currently.

PS: github上有独立的graph-learn仓库，这两边有区别么？会实时同步？

GraphScope requires the graphscope branch of the graph-learn repo, and depends it as a submodule.