/MPRG

Primary LanguagePython

This repository contains the code for the paper:

Misspecified Phase Retrieval with Generative Priors

Dependencies

  • Python 3.6

  • Tensorflow 1.5.0

  • Scipy 1.1.0

  • PyPNG

Running the code

We provide the guideline to run our method MPRG and to compare it with several methods on the MNIST and CelebA datasets.

(1) Run experiments on MNIST dataset

(1-1) MNIST image recovery from measurement abs(Ax)+eta

python mnist_main_mpr.py --nonlinear-model 'abs(Ax)+eta' --num-outer-measurement-ls 200 300 400 500 600 700 800 --max-update-iter 120 --method-ls MPRS PPower MPRG_Step2 APPGD MPRG --noise-std-ls 0.0 0.01 0.1 0.5 1.0

(1-2) MNIST image recovery from measurement abs(Ax)+2tanh(abs(Ax))+eta

python mnist_main_mpr.py --nonlinear-model 'abs(Ax)+2tanh(abs(Ax))+eta' --num-outer-measurement-ls 200 300 400 500 600 700 800 --method-ls MPRS PPower MPRG_Step2 APPGD MPRG --noise-std-ls 0.0 0.01 0.1 0.5 1.0

(1-3) MNIST image recovery from measurement 2sq(Ax)+3sin(abs(Ax))+eta

python mnist_main_mpr.py --nonlinear-model '2sq(Ax)+3sin(abs(Ax))+eta' --num-outer-measurement-ls 200 300 400 500 600 700 800 --method-ls MPRS PPower MPRG_Step2 APPGD MPRG --noise-std-ls 0.0 0.01 0.1 0.5 1.0

(2) Run experiments on CelebA dataset (2-1) CelebA image recovery from measurement abs(Ax+eta)+5tanh(abs(Ax))

python celebA_main_mpr.py --nonlinear-model 'abs(Ax+eta)+5tanh(abs(Ax))' --num-outer-measurement-ls 2000 4000 6000 8000 --method-ls PPower APPGD MPRG --noise-std-ls 0.05 0.1 0.2 0.3 0.4 0.5

(3) Run experiments using the architecture of gernerative prior in [3]

python mnist_main_mpr_relu.py --nonlinear-model 'abs(Ax)' --num-outer-measurement-ls 200 300 400 500 600 700 800 --max-update-iter 120 --method-ls MPRG --noise-std-ls 0.0

References

Large parts of the code are derived from Bora et al., Hyder et al., Liu et al.