Graph Neural Networks in Keras

[Patrick Reiser]

Hello,

I am working with Graph Neural Networks in science and research.

Graph neural networks have gained a lot of attention in recent years. 

I have started to work with batched graphs in the style of keras by creating kgcnn (https://github.com/aimat-lab/gcnn_keras) some time ago.

Of course this was my first attempt in preparing general modules in a library style and there is much I would improve on in a second iteration, especially on the backend and data loading part.

I would like to ask the keras developers, if there is a Keras-Graph or Keras-GNN similar to Keras-CV or Keras-NLP planned in the future?

An implementation within Keras-Core sounds very intriguing. 

If yes then I am happy to contribute.

Thanks

[Francois Chollet]

Hi Patrick,

Nice work with KGCNN!

> I would like to ask the keras developers, if there is a Keras-Graph or Keras-GNN similar to Keras-CV or Keras-NLP planned in the future?

There's TF-GNN, which might as well be called Keras-GNN since it's entirely based on the Keras API.

> An implementation within Keras-Core sounds very intriguing.

TF-GNN is likely going to be TF-only for the foreseeable future, due to its fundamental reliance on `tf.ExtensionTypes` for custom graph types. If you want to start something that works with JAX and PyTorch as well, there's definitely an opportunity there.

Best,

Francois

--
You received this message because you are subscribed to the Google Groups "Keras-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to keras-users...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/keras-users/c14cedbd-3c61-42e1-88d8-4cf48fd948dan%40googlegroups.com.

[Patrick Reiser]

Hello,

thank you very much for your fast answer.

Okay I understand, I think I would like to start something that works with JAX and PyTorch as well.

Thanks
Patrick

[Patrick Reiser]

Hello,

I just saw that keras is merging NLP and CV into a models repo. 

If there is interest from the keras team to also include GNNs, I am happy to contribute or become involved.

I reworked KGCNN for keras 3 and it works nicely. Ideally I would have removed lots of the clutter but I tried to stay backward compatible.

Just using BatchDataLoader compatible to PyG and disjoint graph representation seems to be the most elegant approach also for keras.

Best regards,

Patrick