Jonas Gehring

Thanks for weighing in here, @goldsborough. Here's my take on this, @ebetica can maybe elaborate a bit more. > Compile time type safety being one of the main advantages provided...

> So we already have no compile time guarantees. This PR makes it easier to do things at run time. Yes, that was my point more or less. I don't...

Looks good; for the sake of completeness you might want to declare a default assignment operator although that's done by the compiler automatically anyway.

> but I much dislike putting more boilerplate on users (in having to destructure the variant in the function that calls the typed forward) I'm pretty sure we could do...

I think a low-overhead way would be to offer a saveParameters/loadParameters function in the container base class and leave it to users to make sure they can reproduce their model...

Note that you can serialize custom data in your model by implementing a custom `serialize` or `load/save` function.

This is a good idea! The git backend uses git-rev-list from plumbing, so the information from .mailmap is not pulled in automatically.

Ok, thanks! I think the issue is that the `ValueMapper` trait for smart pointers simply forwards the underlying type. I'd guess that the `from` functions would need to be reimplemented,...

I see; I rolled something on my own for now but it would be nice to have a more principled solution in ATen itself, taking into account backends, shared storage,...

IIRC this was excluded by design, I'm not sure what was the exact reason though.