Support input coefficients in Pyramid.recon_pyr

[berthyf96]

Is it possible to support custom pyramid coefficients as input to recon_pyr? I believe this could apply to all the Pyramid modules except GaussianPyramid. My particular use case is implementing phase-based motion magnification, where the coefficients of the complex CSP are modified before reconstruction. This may be useful for other applications, as well.

I've been using a modified version of this repo with this feature, so I'd be happy to start a PR.

[pehf]

Thanks for opening this issue Berthy. I am very curious about phase-based motion magnification!
You are right that currently recon_pyr does not take coefficients as argument. Instead, it uses the attribute pyr_coeffs of a previously initialized pyramid class. What modification did you use?

Here is the way I would go about it (if I understand you correctly): to reconstruct from modified coefficients, overwrite the pyr_coeffs attribute of the pyramid. For example, adding some noise in one of the channels and computing the corresponding reconstruction (assuming we do not want to keep the original coefficients):

oim = plt.imread('../DATA/Einstein.pgm') / 255.
pyr = pt.pyramids.SteerablePyramidFreq(oim)
pyr.pyr_coeffs[1,2] += np.std(pyr.pyr_coeffs[1,2]) * np.random.normal(size=pyr.pyr_coeffs[1,2].shape)
recon = pyr.recon_pyr()
pt.image_compare(oim, recon)

What do you think?