Kyle Daruwalla
Kyle Daruwalla
That's because of the `cache` in `fmap`, right? `fmap` isn't even used here, so I don't think it is affected. But that speaks to a wider problem about lack of...
Yeah that PR really should have added the example as a test, but yes, the higher-order stuff was only put in place to potentially support them in the future prior...
Works for me
Seems like a good idea to me. If this is something we want, can I suggest kwargs again for the internal parameters like LR? Not necessary to achieve this, but...
Maybe `state_tree` and `state_leaf` are the right balance? I feel like having the tree mental model is important for understanding how this repo works.
We discussed this on call. There was agreement that the structure is no longer a tree, so calling it a tree is confusing. But after comparing to other libraries like...
We discussed this during the call. In general, all the options outlined here don't seem worth it. There is very little value-add, and we are either breaking/bending existing semantics or...
Ideally, I think the implementation would underly `update` (i.e. `update` is `maptrainable` with `f` specialized to call `apply`).
Agreed, with the ability to add more trees to call as described above.
Keeping differentiability aside, is `fmapstructure` not sufficient because of how vectors are handled (e.g. `layers` in `Chain`)?