The following repository contains a package 'phase_retrieval_code_demo' and scripts to replicate experiments of "Phase Retrieval with Holography and Untrained Priors: Tackling the Challenges of Low-Photon Nanoscale Imaging" Hannah Lawrence, David A. Barmherzig, Henry Li, Michael Eickenberg, Marylou Gabrié. The general package can be used for future research on phase retrieval. A demo notebook for usage of the package is also made available. arxiv:2012.07386

The package has been tested with

• matplotlib==3.1.0
• numpy==1.16.4
• Pillow==6.0.0
• scikit-image==0.17.2
• scipy==1.4.1
• torch==1.4.0
• torchvision==0.5.0

from the root of this repository run:

python setup.py install --user

The package in itself accepts some dictionaries to regroup arguments corresponding to different conceptual blocks. Dictionaries makes things clear thanks to the keys. However, in order to be able to run experiments from command lines, for the experiments like "run_reconstruction_holo_opt.py" the dictionaries have tuples equivalent. Here we describe the different possible combination of arguments, how to pass them as dictionaries to the functions and classes of the packages and finally what are their tuples equivalent for the experiment files.

For a noiseless simulation

args_noise = {'noise_type': None}
-an (None,)

For Poisson noise as in the paper: noise_n corresponds to N_p, the number of photons/pixel

args_noise = {'noise_type': 'poisson', 'noise_n': 1e3}
-an ('poisson',1e3)

For additive Gaussian noise noise_sig corresponds to variance of the noise

args_noise = {'noise_type': 'gaussian', 'noise_sig': 1e-3}
-an ('gaussian',1e-3)

For a simulation without reference - classical phase retrieval

args_ref = {'method': None}
-arf (None,)

For a simulation with a reference,

• 'pad' (int) corresponds to the position of the reference with respect to the signal, ref separated by as many blocks of zeros of the size of the signal
• 'method' correspond to the values

# binary random 0-1 entries, right next to the signal
args_ref = {'method': 'binary', 'pad': 0}
-arf ('binary',0)
# random uniform 0-1 entries, 
# separated by a block of zeros of the size of the signal 
args_ref = {'method': 'rand', 'pad': 1}
-arf ('rand',1)
# custom ref, where x_ref as same size as signal
args_ref = {'method': 'custom', 'x_ref'=x_ref}
## no-command line translation

For experiments small block reference is passed as custom to packages like the last line above. The experiment files create this custom ref from a tuple:

• prefix of first element fixes relative size (below 0.1)
• second element remains separation as above
• third argument gives relative position of small block ref in the frame after separation as been taken into account.

-arf ('0.1blockbinary',0,0.5)

#TODO

Fixing the optimizer and learning rate, supports SGD, Adam and LBFGS

args_opt = {'method': 'Adam', 'lr': 0.05}
-ao ('Adam',0.05)
args_opt={'method': 'LBFGS', 'lr': 0.01}
-ao ('LBFGS',0.01)
args_opt={'method': 'SGD', 'lr': 0.1}
-ao ('SGD',0.1)

For optimization directly on the pixels, Optionally inititalization can be chosen:

• fixing 'init_mode' -random uniform between 0-1 'rand' -random normal with variance 1 'randn' (default)
• giving 'x_init'

args_prior = {'method': None, 'init_mode': 'rand'}
-ap (None,'rand')
args_prior = {'method': None, 'init_mode': 'randn'}
-ap (None,'randn')
args_prior = {'method': None, 'x_init': 'rand'}
## no command line equivalent

For optimization with a deep decoder, Optionally: - depth and channels can be specified, (default 2-64) - need_sigmoid (default True)

args_prior = {'method': 'deepdecoder'}
-ap ('deepdecoder')
args_prior = {'method': 'deepdecoder', 'depth': 2}
-ap ('deepdecoder',2)
args_prior = {'method': 'deepdecoder', 'depth': 2, 'channels': 64}
-ap ('deepdecoder',2,62)
args_prior = {'method': 'deepdecoder', 'depth': 2, 'channels': 64, 'need_sigmoid': False}
-ap ('deepdecoder',2,62,'nosig')
args_prior = {'method': 'deepdecoder', 'depth': 2, 'channels': 64, 'need_sigmoid': True}
-ap ('deepdecoder',2,62,'sig')