test of GPU version for CDI

[heidemeissner]

The reconstruction algorithm RAAR (= Relaxed Averaged Alternating Reflection) used in Jena for coherent diffraction imaging exist in python and in pytorch. Pytorch is beneficial when using GPUs, since some optimization features are automatically applied. In Jena, differences between the results where observed although started with the same seed. Only for experimental data sets. And no obvious difference in the code was visible. That means the accuracy needs to be tested.

[heidemeissner]

https://github.com/Ptychography-4-0/ptychography/tree/master/raar_pytorch

[heidemeissner]

Performance: for one example (1024x1024 pixel) it is 3.6 s on GPU (CUDA) and 161s on CPU

[heidemeissner]

Test of pytorch version and our python RAAR version reveals differences for experimental test set. Found three differences in code:

• RAAR (pytorch) returns object not yet filtered by support
• last iteration of pure error reduction iterations are missing
• most important: in the pytorch version, the support is defined and applied for real AND imaginary space.