Aditya Singh

Thanks for clarifying this. How do you want to proceed with this? Also as a side note, it would be helpful to add examples using `fn`, `out` and `default` to...

I only made changes within `nn` and left out rest, like `experimental`. These (2) test cases fail - https://github.com/patrick-kidger/equinox/blob/a89d5b486d13588caffc095f172a2ec39fd68278/tests/test_filters.py#L116 ``` assert Linear(weight=None, bias=None, in_features=1, out_features=1, use_bias=True) == Linear(\n weight=None,\n bias=None,\n...

The more I am interacting with equinox jax/optax I am thinking having `static_fields()` is not that bad to have as default. Saves the trouble of repeating filtering at places.

Thanks for the suggestion. Possibly a method which works on a `PyTree`, `list of leaves` to be ignored and `list of overrides` can be used by a user for constructing...

@patrick-kidger having a `strict=True/False` like functionality akin to Pytorch's `load_state_dict` can also be useful in loading partial weights which returns the mismatched leaves along with partially loaded model?

Yes. Although, is it possible to work on a common prefix of the two trees?

I see, thanks. Happy to work on `strict` and/or extending `_default_deserialise_filter_spec` if you see the advantage of adding them.

Hi @patrick-kidger Just wanted to go over some design minutes for extensibility of `_default_...`. I am not familiar with `type piracy` and found resources only discussing it in the context...

Thanks for the detailed response on `type piracy`.

Hi @patrick-kidger Regarding the `strict` functionality to allow loading a prefix of the stored tree. It seems that the current code already does this. https://github.com/patrick-kidger/equinox/blob/a89d5b486d13588caffc095f172a2ec39fd68278/equinox/serialisation.py#L185-L187 If `like` is a prefix,...