[datanecdote]

How does Jax lose composability or introspection?

[krastanov]

E.g. jax does not autodifferentiate anything that is not jax (scipy ode solvers, special functions, image processing libraries, special number types (mpmath), domain-specific libraries). Compare that to Zygote.jl

[datanecdote]

It is true that Jax cannot differentiate through C code. But it can differentiate through python code that was written to accept Numpy.

[amkkma]

Which is extremely limited compared to Zygote which can do custom types, dicts, custom arrays, complex type system and multiple dispatch uses etc

[datanecdote]

Try reading the docs before making sweeping negative comments about what a piece of software can and cannot do.

https://jax.readthedocs.io/en/latest/notebooks/autodiff_cook...

[krastanov]

Are you talking about "Differentiating with respect to nested lists, tuples, and dicts" from that page? The comment to which you are responding covers quite a bit more. The jax documentation specifically says "standard Python containers". Zygote.jl and other less stable Julia auto-diff libraries go far beyond the built-ins and can work with structures defined by packages never designed to be used with automatic differentiation. Of course, there are limitations, but quite a bit less severe than the one in jax (and again, I am saying this while being a big jax fan).

[datanecdote]

As the document I linked to says, Jax autograd supports custom data types and custom gradients.

It’s honestly exhausting arguing with all you Julia boosters. You can down vote me to hell, I don’t care. I’m done engaging with this community.

You all are not winning over any market share from Python with your dismissive, arrogant, closed minded culture.

[amkkma]

I don't see anywhere about a better type system or multiple dispatch. Try being less salty

In what language are you defining these custom arrays or types? certainly not in python, or they'll be too slow to be worthwhile.

[krastanov]

Indeed, but "python code written to accept numpy" is a pretty restrictive subset (comparatively; I do still enjoy using python). It does not even cover most of scipy, let alone the domain specific libraries, which frequently end up using cython or C for their tightest loops.