Measure spectral error in samples

[krisrs1128]

More than just the raw images error, we should consider the consistency between the learned power spectra and the ground truth.

Assignment:

• Use the 2d fourier transform, as described in Tianle's message,

The other metric I mentioned during yesterday’s call is the power spectrum or spectral density. It’s basically the square of the 2d Fourier transform an image. We can take the radial profile of the 2d power spectrum and obtain 1d spectrum if we wish. The spectra can be compared from modeled and real images. We can therefore compare if the modeled image have spatial spectral consistency with real ones.

[krisrs1128]

@mustafaghali -- I remember seeing this in your thesis, but I can't find the code for it anywhere. Can you share it with us when you have the chance?

[mustafaghali]

@krisrs1128 I used this script https://github.com/krisrs1128/clouds_dist/commit/57d5be4e15c3d4171fda7a58832225eb289f3b8c

[krisrs1128]

Ah I totally missed that, thank you for pointing me to it!