Eelco Hoogendoorn

Just looking at the levit implementation; one apparent difference from the paper I noticed relates to this 'Following [54], each batch normalization weight parameter that joins up with a residual...

Yeah but its (dim_in, dim_in) in the first strided conv layer. Dont know if its a massive difference in practice, but its not according to the definition in the original...

Yeah not a depthwise-conv expert by any means; infact this was the first time i dug into the construct cause I got annoyed with the PiT paper not really explaining...

Frankly, I think supporting numpy arrays of object dtype is something we should have excluded in the first place. Numpy bugs or not, this is very hard to get right...

I agree of course; the deserialized datastructure should ideally be indistinguishable from the original; and I dont think its technically impossible; but I really dont see it being very easy....

I see some work has been done on NumpyNDArrayHandlerBinary already; a few notes: - Adding the tolist() case could be implemented more elegantly by defering to the superclass; it handles...

https://en.wikipedia.org/wiki/Vector_space Basically I just mean addition of two objects of equal type, and scalar multiplication; __add__(self, other), and __mul__(self, scalar). These should be well defined operations for all the classes...

Right; it makes sense that the objects as named are intended to be normalized quantities; so indeed it makes sense not to provide operations that take them outside their designated...

> Do let me know if you make a more general package though, sounds like something that could be useful to me. :) https://github.com/EelcoHoogendoorn/numga There you go!

I did a bit of benchmarking for those interested: ``` from jaxlie import SE3, SO2, SO3 import jax from jax import vmap, make_jaxpr import jax.numpy as jnp import jax.random key...