Jonathan Godwin comments

Results 13 comments of


                                            Jonathan Godwin

Creating Own Dataset For 'Learning to simulate complex physics with Graph networks'

Hi there, we are happy to help if we can. Could you provide some more information on what your issues are?

Added from_scipy_sparse_matrix function

Thanks for bearing with me! I think this looks good, if you could resolve the final conflict I'll accept it :)

How to handle heterogeneous graph features?

Hey, Sorry for the delay in getting back to you. This is a question we've thought about quite a bit and at the moment there is no neat way to...

Adjacency matrix to GraphsTuple

Hey Salfaris, We don't currently have a function that does this, but we would be very happy to receive pull requests :)

What's the purpose of empty graph in padding?

Hey! Good question - we connect all edges to a single node (the first one) and then all other nodes are disconnected. We use this because then with `unpad_with_graphs` we...

Reference code for NLNN with multi-head attention?

Hey, A transformer is effectively a Graph Attention Network applied to a fully connected graph. In order to use Jraph to accomplish such a model you would: - create a...

Reference code for NLNN with multi-head attention?

Thanks! I've updated the example. You could also decoder the attention_logit_fn with @jax.vmap and then keep the np.dot product. The vmap simplifies some logic but can make things a little...

Reference code for NLNN with multi-head attention?

Hey, That is a little concerning. Do you have a code gist or something I could look at? I think it could be a few different things: - is the...

Reference code for NLNN with multi-head attention?

Hey, I would recommend using either the tensorflow dataloader or the pytorch data loader, with a wrapper that converts DGL (or equivalent) to GraphsTuple. I think, if you are using...

Feature engineering methods

Hey INF800, Unfortunately we don't - this is a library for GNNs. Network X is a nice library for this: https://networkx.org/documentation/stable/tutorial.html