Margaret Duff
Margaret Duff
After discussion with Edo and Gemma we decided that this should go in but that we would still further investigate #1863
@gfardell - do you want to check the adjoint code and approve if you are happy?
We need to decide the default. If no `out` is provided to direct(adjoint) but it should return a 1 by 1 block data container, should this be a data container...
Discussed with @samdporter! We should make this a priority before the SIRF challenge and the SIRF/CIL training at PSMR
@paskino suggested looking at his PR https://github.com/TomographicImaging/CIL/pull/1766
> I'm doing some synergistic deconvolution. I'm working with SIRF which means I need a `BlockDataContainer `for multiple channels. The problem comes when I define a operator like this: >...
> might want to check: > > * https://github.com/TomographicImaging/Hackathon-000-Stochastic-QualityMetrics/blob/main/img_quality_cil_stir/image_quality_callback.py > * https://github.com/SyneRBI/PETRIC/blob/main/petric.py In the Petric challenge they have the callback: ```python class SaveIters(Callback): """Saves `algo.x` as "iter_{algo.iteration:04d}.hv" and `algo.loss` in...
> I'll open a follow-up PR to fix this (#1913) as well as #1892 later. Fine for now. Thanks @casperdcl - is it the naming that you want to change?
I think in the paper they show optimal sampling when they know their functions are strongly convex. We don't yet have strong convexity options in SPDHG, although maybe we should!