graph-learn
graph-learn copied to clipboard
alibaba
An Industrial Graph Neural Network Framework
1. Trained with GraphLearn-Training: implement model with u2i2i and i2i2i meta-path; 2. Infered user embedding with Dynamic-Graph-Service: - Add getQuery http request; - Refine EgoGraph/EgoTensor for parsing sampled fbs records;...
1. The new server impl is based on hiactor, using this impl will override the default thead pool impl. 2. Configure with CMake option `-DWITH_HIACTOR=ON` to build with hiactor. 3....
GraphLearn mainly provides the ability of large-scale distributed **offline training**, but lacks support for **online inference** (most GNN frameworks do not support it), while industrial use often requires online inference,...
用户文档上写了： 分布式场景下，即存在多个GraphLearn Server时，构图时会自动进行图的partition, 将图分布式存储。默认partition是按照src_id % server数进行节点和边的分配。 想问一下这个partition方法是在源码的什么位置呢，我在源码中没有找到。我目前是想重写自己的partition策略并加进去，期待回答谢谢。
gcn 报错
world_size: 1 rank: 0 Traceback (most recent call last): File "train.py", line 228, in run(args) File "train.py", line 151, in run g.init(task_index=args.rank, hosts=thg.get_cluster_spec()) AttributeError: module 'graphlearn.python.nn.pytorch' has no attribute 'get_cluster_spec'
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: graphscope中GLE就对应graph-learn，这两边有区别么？会实时同步？
Pip install is not working. I also tried the solution you have given but I still get this error: ERROR: Could not find a version that satisfies the requirement graph-learn...
