Open hansen7 opened 2 years ago
可以的。我们大概花2个月左右时间写这章。先在这里讨论下章节的内容。
GNN应用 计算范式: Message Passing 编程范式:Tensor-Centric(PyG), Graph-Centric(DGL), Vertex-Centric(MindSpore-GraphLearning) 整图学习系统: RoC 基于采样子图Mini-Batch学习系统: PyG, DGL
@luomai @hansen7 本章节需要覆盖的一些内容的proposal~
I don't agree that there are differences between the paradigms between DGL, PyG, and Seastar. All of them adopt vertice/edge-centric DSL. The only differences are how is the DSL compiled and executed, and the methodology converges as time went by.
I encourage paying more attention to distributed training of GNNs: how to optimize pipelining, networking, sampling, etc.
@yzh119 谢谢建议。图学习系统这块我们还没有开始梳理出来。很欢迎提建议和参与。我们最近在忙第一版的发布。忙完这阵就会开始思考图学习系统的章节了。
各位老师好,想请问图学习这一块进展如何了?对Graph+MLSys感兴趣,非常期待!
与@luomai 老师交流后,我很希望能够Co-Lead/参与到这一拓展章节的撰写过程中,我对Graph, GNN等比较熟悉,刚开始了解MLSys + Graph这一有趣的方向。