Francesco Farina
Francesco Farina
> I think it is a great idea to have support for GNNs in MLX. However, I think that it is better to implement GNN-related stuff in a separate library...
> When it comes to tests, you could test whether neighborhood aggregation is correctly implemented. Yes that makes sense, but right now that part is not exposed (just used in...
Yes, closing now.
Hi @angeloskath, great - we're iterating on exposing each `scatter_{op}`. Do you think having also the generic `scatter` with the `mode` (or `reduce`) parameter (as @gboduljak implemented in https://github.com/francescofarina/mlx/pull/1) would...
Looks like this one can be safely closed and we can move with https://github.com/ml-explore/mlx/pull/394 which has a different API. @angeloskath @gboduljak ?
That makes sense, but I would argue that having to use protected members may result in a design/behavioral pattern that's hard to enforce in a "soft" way. Throwing an error...
Just to add on the comment I made here https://github.com/ml-explore/mlx/pull/380#issuecomment-1879761512 . It seems we're going down the road of implementing scatter and sparse ops in the mlx core, so maintaining...
> The fastest path to get this going is an external repo (outside ml-explore that is). If you all are interested in making one I think that is great and...
> I assume the message passing base class is what you are referring to? Are there others you were thinking of? @awni yes I was thinking of the MPNN and...
@TristanBilot sure, handler in my profile.