Suggested Architecture for Image Restoration #11
benjaminpope
started this conversation in
Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
We should probably look at how MPoL, MIRA, and SQUEEZE parametrize things.
For a quick start, I think we basically want to assume the model is in zodiax, and we have a
path
to animage
which is our pixel or other basis representation of the target image. This is one of the parameters of azodiax
function that gives our forwards model to match our data, and we have a likelihood function.Then basically what we need to write is
regularizers
and
optimization schedules, including wrappers that know (preloaded?) good weights to attach to each of these terms, in Adam, maybe with convenience functions to instantiate everything.
We should expect to interface with this in
optax
for now, but leave the option open fornumpyro
with probabilistic priors down the track.We will have to be careful about centering in cases (eg with interferometric data) where we don't have an absolute phase centre. This could just be by setting one phase to zero but we don't want the instability where the whole image flies off in some direction!
Beta Was this translation helpful? Give feedback.
All reactions