Clement Gehring
Clement Gehring
The links you posted seem to give me a 404 error. > The ability to work with stochastic iterable functions. After a lot of discussion with Clément, that just means...
> In my design, I chose to put all of the other parameters to two_phase_solver into the param_func object to keep things simple. I think you're essentially describing an interface...
It's been made easier in #26 and #30 but there are still some things to do like properly support pytrees and add support to stochastic vjps.
As far as `two_phase_solve` is concerned, you are always solving fixed-point problems regardless of forwards/backwards/other. It formulates reverse differentiation as a fixed-point problem and recursively calls itself with one fewer...
I think I understand what you want. I don't know if I want to change the whole "differentiate a fixed-point by solving a fixed-point" paradigm of Christianson's two-phase method. Although...
Thanks for pointing that out. I think the problem is that the use of the word "solution" is ambiguous given we define a `FixedPointSolution` namedtuple. Those docs were written when...
No, the docs still use the same language. We should change the word solution to something more specific.
Can you open an issue with a brief explanation of what the bug is. It's not obvious to me from your link. Could you elaborate on what is preventing us...
@pierrelux Have all your requested changes been addressed?
I'll have to think a bit more carefully about this but is using the two-phase method on non-fixed point problems advisable? At this point it might be better to use...