In DGL’s first release last December, we focused on usability by introducing a set of carefully designed, easy-to-use APIs that support a variety of model implementations of Graph Neural Networks. We decided to keep DGL framework-agnostic to engage with users from different platforms (PyTorch, MXNet…). As a result, in our earlier releases, we largely leveraged the available functionalities provided by these frameworks, and based on many valuable feedback from our users, we are well-aware of the room of improvement particularly on some new models defined on the sparse and irregular graphs.
This is a companion discussion topic for the original entry at https://www.dgl.ai/blog/2019/05/04/kernel.html