I do think that these could be further improved, but not sure if we are willing to put the time in right now. Right now there is still no basis for a user to decide which network would be the most suitable for them.

I agree, but since we have not yet implemented these networks and many lack detailed references, providing more specific information might be challenging. The key aspect for the user is understanding whether a network is suitable for regression, classification, or both, and this information is there. Given the lack of a one-size-fits-all architecture, experimenting with different networks is often necessary (more detailed descriptions and prior applications can of course be useful but they don’t guarantee optimal performance anyway).

Also, ideally, we would start adding the missing type annotations (for which the linting rules are currently suppressed in these modules). When I implemented ruff, I found that it was a lot of hassle to do so, and realize that this is still true. As above, not sure if it's worth spending more time figuring out what the types are for each element in each of these functions/classes.

While this would be of course a nice improvement, I don’t think it’s worth investing more time in this at this stage.