DeepInverse integration for more algos, physics, models, datasets, losses...

[Andrewwango]

Hi there! Great work for setting up a diffusion benchmark for interesting inverse problems!

In this issue, I propose connection with DeepInverse to make use of the following features in DeepInverse:

• other SotA reconstruction algorithms in DeepInverse (variational optim framework, PnP/RED, unrolled, MCMC, self-sup learning, GANs...);
• more physics operators (deblurring, super-resolution, medical imaging, remote sensing, phase retrieval...);
• datasets + model training framework

Motivation

Reproducible research in inverse problems is accelerated by sharing ideas and methods across modalities, e.g. it's nice to see that diffusion methods proposed e.g. in MRI have been successfully applied in other scientific inverse problems. Along with deepinv/deepinv#990 (i.e. integrate inversebench into deepinverse), researchers from more communities can benefit from more baseline methods given an inverse problem, as well as more inverse problems given a reconstruction algorithm.

[devzhk]

Hi, thanks again for raising this and for the interest in InverseBench! Also, congrats to the achievement of DeepInverse!

I checked the cited issue in deepinv. It sounds like you are primarily interested in integrating full waveform inversion, Navier-Stokes, and inverse scattering problems into DeepInverse library. That said, I'm not sure if you are proposing to reimplement these inverse problems directly within DeepInverse, using InverseBench mainly as a reference? or shared interfaces so that DeepInverse solvers can directly run on InverseBench? How do you see the maintenance being handled in either case?

[Andrewwango]

Hi @devzhk , there is a potential integration in both directions - I'll have a look in the coming weeks if providing wrappers would be trivial in either direction. For example, it could be nice in inversebench to call an optimisation/sampling algorithm from DeepInverse.

Maintenance is an issue, thanks for raising it. We've observed with minimal wrappers that it doesn't provide much of a problem, as long as the wrapped functionalities are clearly documented and tested, with deprecations notified far enough in advance.