Michael O'Brien
Michael O'Brien
This is a good suggestion. My users typically need to reference individual pytree leaves for fine control of which parameters are transformed via vmap, grad, etc (we’ve spoken about this...
Though to avoid confusion with pytree leaves rather than using a property this is easily solved by using a `get_*` method!
I’m not sure I fully understand what you’re trying to do with this vmap, but depending on what that is the discussion in #618 could be useful.
This makes sense. This and the nan option both seem like they would have their use cases!
This is okay, makes sense! It is probably best practice to be explicit about the out_axes anyway. The best replacement for the behavior is probably to just create an out_axes...
> 1. I think that different termination criteria could in principle be supported - this is a use case [@bagibence](https://github.com/bagibence) is also interested in, if I recall correctly? This should...
On the termination criteria, one thought is that `norm` is somewhat superfluous with having a function for the criteria itself. The issue though would be that it is far simpler...
> First off, as a clarifying question: If you have a scalar loss, why is squaring it undesirable? Due you think that this is due to the noisiness, and would...
I was reading up on some literature on this problem and have been looking through `scipy`’s implementation of LM. It is based on the work of J. J. More, “The...
Thanks for the thoughtful reply on this! Long reply incoming… so please no rush to get back to me! > Do you propose making a change in this line? >...