Hi, I am XING LI, a researcher from Sansan DSOC Data Analysis Group.
This is the article of Day 9 of Sansan Advent Calendar 2020.
Last time, we talked about some common tasks in deep graph learning and built a toy network on Node Classification
task as a demo. We have known how to build, train and test a simple deep graph neural network by DGL. However, the graph dataset we used, Cora dataset, only has 2708 nodes and 5429 edges. The GNN's aggregation operation will not take a long time on such tiny graph. But in reality, we also face much larger graphs, for example, over million nodes and billion edges. Consider an -layer GCN with hidden state size , training on an -node graph. To store the intermediate hidden states requires memory, which is easily exceeding one GPU’s capacity with a large . So today we are going to talk about how to train a deep GNN on large graphs with help of DGL.
続きを読む