Compute_stepsize function in proximal_grad

Question

Compute_stepsize function in proximal_grad

mrocklin opened this issue 7 years ago · comments

Currently the stepsize computation in proximal_grad is still embedded within the code, which makes it hard to manage it with Dask. If the proximal_grad method is useful then it would be nice to pull this function out.

cc @moody-marlin

Chris White · Answer 1 · Sat Jan 28 2017 03:17:08 GMT+0800 (China Standard Time)

So this is where the regularization stuff gets hairy; proximal_grad is an algorithm for regularized logistic problems, which means the function evaluations within compute_stepsize need to be a modified loglike function; this in turn makes jitting compute_stepsize more complicated because it now needs a function argument.

Matthew Rocklin · Answer 2 · Sat Jan 28 2017 03:18:17 GMT+0800 (China Standard Time)

I'm not looking to jit, I'm looking to wrap with dask.delayed. It is fine to take a function as an input in this case.

…

On Fri, Jan 27, 2017 at 2:17 PM, Chris White ***@***.***> wrote: So this is where the regularization stuff gets hairy; proximal_grad is an algorithm for regularized logistic problems, which means the function evaluations within compute_stepsize need to be a *modified* loglike function; this in turn makes jitting compute_stepsize more complicated because it now needs a function argument. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#17 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AASszJQ9eTsV_z0alnbd6SqUCwM1LWcZks5rWkK0gaJpZM4LwDyU> .