Add object regularization for multi-slice reconstruction

[yiranlus]

I add object regularization which is mentioned in

Zhen Chen et al. ,Electron ptychography achieves atomic-resolution limits set by lattice vibrations.Science372,826-831(2021).DOI:10.1126/science.abg2533

I tested with the two random images and have the following results on phase (regularization = 0.05),

and before it was like this

For now, it only supports single mode and slice distances should be uniform.

However, the current code is very slow. I hope someone can give some hints how to make it faster.

[daurer]

@yiranlus thanks for your contribution. It's seems to make to add such an optional regularizer to the 3PIE engine. Here are a some suggestions to improve the code

1. Avoid hard-coding storage IDs like "Sscan_00G00" as they can be changed by the user script. Instead, you can get things like shape and psize from the first object view:

shape = list(self.ob.views.values())[0].shape
psize = list(self.ob.views.values())[0].psize

2. Instead of np.fft.fftn and np.fft.ifftn you can use the internal propagater which you can get from the first pod

fft  = list(P.pods.values())[0].fw
ifft = list(P.pods.values())[0].bw

and you should probably iterate over the pods when you apply the regularizer update (similar to what is done in the multislice update) - this way your code should also work for multiple probe modes.}

3. See other in-line comments below

[daurer]

It would probably make sense to move this a bit higher up to make sure that self.ob is calculated for plotting after the regulariser has been applied...

[daurer]

this part is basically just calculating some weights w and does not depend on any current update so can be moved into a separate function, e.g. initialize_regularizer and called once in the constructor. In this current implementation the weights are re-calculated for every iteration which seems unnecessary.