Siddhant Chaudhary

Hi @CarloLucibello. Is this a documentation issue? I looked at the implementation in `src/layers/heteroconv.jl`, and it seems like the implementation takes care of all standard graph conv layers (in the...

Hi @CarloLucibello, another small question: for layers like `AGNNConv` which involve self loops; such layers will then have to operate on two types of edges, right? Because, we'll have to...

If this looks good, I'll go ahead and add some tests and add an example in the documentation as well.

The documentation CI is failing for an `RNN`. I assume even there the output is incorrect? It's probably showing the `state` parameter as non-trainable (which is trainable, right?)

> > It's probably showing the state parameter as non-trainable (which is trainable, right?) > > That's `Recur.state`, which should be non-trainable. Note how only `cell` is included below: >...

> Yes, tied parameters are tricky as we found out while working on Optimisers.jl. Sometimes it feels like a philosophical question. Do we consider array wrappers like `Adjoint` and `Transpose`...

> I believe that'd run into the same problem with shared params across nominally different layers. Maybe one idea would be to separately count the number of shared params and...

> ```julia > d1 !== d2 > ``` I see, yes, that makes sense. I think, any solution which counts distinct (or shared) parameters in a model must use some...

Hello. I'm trying to implement a few convenience functions, but I had a small question: are `first` and `last` supposed to take in a `DTable`, and return a `DTable` with...

@jpsamaroo how does this implementation of `first` look? If this is okay, I'll implement `last` similarly.