Comparing Treex with Equinox

I think it is natural to compare Treex with Equinox, as both are PyTree-based libraries. The README currently says

Other Pytree-based approaches like Parallax and Equinox do not have a total state management solution to handle complex states as encountered in Flax. Treex has the Filter and Update API, which is very expressive and can effectively handle systems with a complex state.

I assume the total state management solution refers to the Kind system in Treeo. However, the recent RFC indicates that we cannot use that with higher-level frameworks like Elegy. Suppose I want to use Elegy for training loop automation, is there any reason I should prefer Treex over Equinox?

[cgarciae]

Stateful layers like BatchNorm are experimental in Equinox, but I think this itself is not a problem. The real issue is that Equinox is designed to work with arbitrary pytrees so its filters can really only take into account leaf values, this means you don’t have access any metadata including the field’s name, which in turn means you probably cant express something like “select all the batch stats and cache leaves”.

Treeo filters on the other hand operate over FieldInfo which have access to the following information:

class FieldInfo:
    name: Optional[str]
    value: Any
    kind: Type[Any]
    module: Optional[Tree]

So if leaf is part of a treeo.Tree the filter has access to field’s name, kind, and the actual module instance (tree would’ve been a better name for this), and of course the leaf value itself. This makes treeo/treex filters a bit more powerful.

[cgarciae]

Its a very good question. I’ve been following Equinox closely as Treex was inspired by it, and talk every one in a while with Patrick. Ideally we only have a single Pytree Module library but as it stands here is the situation:

• Treex main contribution was to add Kinds to Pytrees that are able to emulate Flax collections. This is needed for the advanced lifted transformations as you see on Flax: nn.{scan, vmap, etc}, it also simplifies a lot of code since e.g. selecting all BatchStats become a trivial filter.
• Problem is that Kinds are a non-standard mechanism in Pytrees which could be called “tagged leaves”. The problem is that since not all Pytree have this capability it become harder to interop with Pytrees from other libraries. For Elegy this meant that you had to create wrapper objects that correctly annotated the Kinds so Elegy could query them. Equinox has focused on staying within the current limitations of Pytree’s, I think is very valuable to explore how far you can get this way for a NN library.

If you don’t need BatchNorm or other potentially stateful layers (anything that needs a 'cache') then Equinox is a good choice, I highly encourage you to try it out, else Treex was designed to try deal with these cases (although the state management situation for pytrees is not perfect). I will definitely try to add Equinox support for Elegy.