Confusion about DGL and different backends

[xnhp]

Hey there,

I just recently started having a first look into DGL.

I am intruiged by the fact that it comes with quite a number of models/layers already implemented.

In some flavour text, it is suggested that DGL is independent of the backend (PyTorch / Tensorflow), etc.

Does this mean that any pre-implemented layer can be executed on either of the two backends? As an example, take simple graph convolution – could I execute this while using TensorFlow as the backend as described here?

My confusion comes from the fact that in the docs there is a distinction being made between PyTorch, MXNet and Tensorflow NN modules.

Best Regards,
Ben

[mufeili]

The NN modules for different backends are independent. DGL has implementation of GraphConv for both PyTorch and TensorFlow though.

[xnhp]

Awesome, thanks a lot for the clarification!