I noticed that DGL as best open source and it is faster. Is there any comparison of PyG or StellarGraph or any other Graph Deep Learning model with DGL specific to Graph Representation learning. Can any one explain
Hey, glad to see you like DGL. I think in overall the area of graph representation learning is still quite fresh and there is hardly any standard benchmark like ImageNet in CV, or SQuAD in NLP. As a result, system comparison at the current stage could be biased from scenarios that have practical impact. As a result, DGL’s focus is to push graph representation learning out of research incubator into larger or more complicated applications. Keeping this in mind, you could find some comparison numbers in our latest blog post.
Minjie, nice to hear Kernel fusion in v0.3 and good initiative taken by DGL to boost its performance to next level. This seems to be a phenomenal change in Graph paradigm proposed by DGL and i’m curious to know the release dates as well!!!
But two things from my end
#1 understood its nascent nature but if the focus is specific to deep learning with GCN based learning model where DGL is doing good, if you can benchmark analysis with other Models/App as well. Again i’m focusing specific to models of GCN with Graph Classification, feature extraction/learning and embedding scenarios.
#2 if we adopt UDF based function based like in v0.2, do you recommend us to wait for v0.3 as it seems a phenomenal change in the message passing protocols.