Missing dependency on dgl

[ndickson-nvidia]

Installing via pip as specified in the README, i.e. pip install dgllife , there's no dependency on any dgl package, so without manually installing a dgl package, dgllife fails to initialize at runtime. It is mentioned earlier in the README to install dgl, but it's easy to miss when working with code that imports dgllife directly, but not dgl. Also, without having a versioned dependency, compatibility issues may result in the future if people are using incompatible versions.

On the other hand, the dgl package naming with respect to CPU vs. CUDA 10.2/11.3/11.6 may make it challenging to add such a dependency, so it's not quite as simple as it seems. I just wanted to mention it because I ran into a case where dgl wasn't installed and there was some confusion among multiple people regarding dependencies. It worked after I ensured that a dgl package would be explicitly installed.

[mufeili]

Thank you for pointing out that. Currently I have no ideas about a good fix. Let's keep the issue open and see if there will be other suggestions.