Make TOGL feature complete

[marcusblake]

Currently, the implementation of TOGL ignores the following features:

• handling higher-order information properly
• expanding simplicial complexes
• making use of the dimension of features

[Pseudomanifold]

Some of these are rather tough to handle; this should probably rather become a project on its own. In the original paper, for instance, we are not actually going beyond dimension 1 in terms of features, so I am somewhat loath to mix the re-implementation of TOGL with new feature development. A good starting point would be to see how to capture 1-dimensional information; in my current implementation, I figured that I could get this stuff much more easily by using a lower star filtration.

[uchukwu]

Hi @Pseudomanifold, is this a good place to pick up on where we left off in BorgwardtLab/TOGL#13, or should I open a new ticket?

[Pseudomanifold]

Sure!

[uchukwu]

Thanks! You mentioned that TOGL isn't complete yet during our discussion. What is missing? What features do you need added?

[Pseudomanifold]

Potentially some other aggregation functions would be useful; I also have not tried out how easy it is to include this type of layer into arbitrary GNNs. Except for the aggregation functions, I think this is relatively in sync with the other code, though.

[JerBaf]

Hello! I am also coming from the previous implementation of TOGL. I tried to carefully read the code but I struggle to identify what are the key differences between this implementation and the original one for TOGL. Would it be possible to have a few key points on the major differences between the two code bases?

Best,
Jeremy

[Pseudomanifold]

The main difference is that the current implementation only supports aggregation based on deep sets.

This is what works already:
- simple deep set layer
- simple TOGL implementation with deep set functions
- basic GCN with TOGL

This is what is currently ignored:
- handling higher-order information properly (described in TOGL but not yet implemented)
- expanding simplicial complexes (described in TOGL but not implemented)
- making use of the dimension of features (I'm only using dimension 0 at the moment)

Hope that helps!

[JerBaf]

Okay great thank you very much!

[cypans]

Hello, may I ask if we can backpropagate after continuous homology calculation? Especially on CPU, gudhi does not support Tensors

[cypans]

BorgwardtLab/TOGL#13 (comment)