Open mrava87 opened 2 years ago
Looks very promising. Would be interested to see if this means we can relax having to know the maximum eigenvalue. Maybe setting a sufficiently large value is enough under this adaptive scheme. Looking forward to seeing this!
Good point! We haven’t looked into this but definitely for some problem where you can’t tell upfront the eigenvalue, not having to know it with high accuracy would be a bit plus :)
Now that #90 is closed, it would be interesting to convert the AFISTA code into a class-based solver: https://github.com/PyLops/pylops_transform2022/blob/main/afista.py
@mrava87 @NickLuiken
Hi @cako @mrava87, If this issue is not touched upon, would like to work on it.
Hi, this issue is already taken, we just need to move codes around so no point for you to look into it.
Feature
Recently an Adaptive-FISTA algorithm was introduced in https://www.earthdoc.org/content/papers/10.3997/2214-4609.202229010. This algorithm shows faster convergence for problems where the model is expected to be very sparse (as it approximates the
L0
norm).Implementation
@NickLuiken will implement this in
pylops.optimization.sparsity.ISTA
andpylops.optimization.sparsity.FISTA