1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
|
### Config file for Purify
Version: 4.2.0
GeneralConfiguration:
InputOutput:
input:
# If the following parameter is "measurements", only the measurements block should
# be provided, and the measurements will be read from file. If it's "simulation", only the
# simulation block should be provided, and random measurements of the given
# image will be generated with the requested noise.
source: measurements # one from measurements, simulation
measurements:
measurements_files:
- /path/to/measurment/set # path to the measurement set
measurements_polarization: I # one from I Q V XX LL
measurements_units: radians # one from lambda, radians, pixels
measurements_sigma: 1 # the uncertainty of a visibility (RMS noise of the real or imaginary part of the visibility in Jy)
warm_start: /path/to/warm/start/image #path to fits file image of previous purify solution (assumig the same cell size), otherwise leave as an empty quotation ""
#simulation:
# skymodel: /path/to/sky/image # path to the sky model image file, the ground truth image for the simulation
# signal_to_noise: 30 # signal-to-noise ratio to be added to the sky model
# number_of_measurements: 10000 # number of measurements per a node
# w_rms: 30. # RMS w value in random uvw coverage
# sim_J: 8 # the minimum support for the kernel when simulating measurements
# coverage_files: #path to converage units. This is only used if it is not an empty string
# - path/to/coverage/measurement/file
# coverage_units: lambda # one from lambda, radians, pixels
output_prefix: /path/to/output/dir # results directory will be created inside this path, and called output_<timestamp>
### Execution parameters
logging: debug # How much logging you want (debug, critical, ..). debug maybe the most useful in some cases because it gives the most output.
iterations: 100 # number of iterations (>= 0). More iterations can give better results.
epsilonScaling: 1 # factor to scale measurements_sigma by. We suggest 1, but it can be useful to change.
########## Measure Operators ##########
MeasureOperators:
pixelSize: #pixel size in arcseconds. This should be estimated from the longest baseline (PURIFY should print out a suggestion)
cellsizex: 1
cellsizey: 1
imageSize: # for the moment: power of two . For a simulation or when using a warm start, this must be same image size as the ground truth
width: 1024
height: 1024
J: # Interpolation pixels for the anti-aliasing kernel (default 4x4). Input either a value or pair (J == 6 is the only option for pswf)
Jx: 4
Jy: 4
Jw: 30 #Maximum size of w kernel
kernel: kb # kernel, choose between: kb, Gauss, box, pswf
oversampling: 2 # value > 1. Value of 2 is the standard
gpu: False #This can be used when compiled with arrayfire gpu library
powermethod:
iters: 100 # value > 0. This is the maximum number of iterations used with the power method for calculating the measurement operator norm.
tolerance: 1e-4 # value > 0. This is the tolerance for convergence of the operator norm
eigenvector: #Here we can choose real and imaginary fits images of the operator eigenvector associated with the operator norm. This speeds up convergence for repeat runs.
real: ""
imag: ""
wide-field:
wprojection: False # using radially symmetric w projection kernel
mpi_wstacking: False # applies average w-stack correction on each node (always True with wprojection)
mpi_all_to_all: False # performs all to all operation of the grid to even out computation. Highly recommended when using MPI for wide-field imaging!
conjugate_w: True #reflects measurements onto the positive w-domain (can reduce computation)
kmeans_iterations: 100 #number of iterations in w-stacking clustering algorithm
########## SARA ##########
SARA:
# wavelet bases, to choose from: Dirac, DB1, ..., DBn (where n=0 is dirac, and 1 <= n <= 35 is DB basis)
wavelet_dict: 0..8 # sequence of desired basis: 0,23,31 or 0,12..16,20 . We typically use 0..8, this gives good results but will be expensive for large images.
wavelet_levels: 4 # maximum number of wavelet levels for each dictionary. More levels can be good for increasing quality of large scale structures in an image (we recommend 4 to 6).
realValueConstraint: True # Boolean, we recommend this to be True for Stokes I
positiveValueConstraint: True # Boolean, we recommend this to be True for Stokes I
AlgorithmOptions:
algorithm: primaldual # will just read the options of that one. We recommend the primal dual algorithm, especially for large image sizes.
########## PADMM ##########
padmm: #solve the constrained problem
# Following is only accepted when MPI is used
# none - no MPI used, used when on a laptop or no w-stacking
# serial-equivalent - gridding and degridding and visibilities split across nodes
mpiAlgorithm: serial-equivalent # one of none, serial-equivalent
# Be careful changing the following unless you know what you are doing
relVarianceConvergence: 1e-3 # (>0) relative converegence tolerance of objective function (we suggest at most 1e-3, but smaller might produce better results)
epsilonConvergenceScaling: 1 # (>0) l2 convergence tolerance for residuals (we suggest 1)
dualFBVarianceConvergence: 1e-3 # (>0) relative convergence tolerance of l1 proximal
stepsize: # Stepsize, choose value between: default, variable, 100, x
update_tolerance: 1e-1 #when predicted stepsize is greater than this value, it will be updated
update_iters: 0 #how many iterations to allow updating of step size (we recommend 0 to 50)
fb: #solve the unconstrained problem
mpiAlgorithm: serial-equivalent # one of none, serial-equivalent
regularisation_parameter: 1
stepsize: 1
relVarianceConvergence: 1e-3 # (>0) relative convergence of the objective function
dualFBVarianceConvergence: 1e-3 # (>0) relative convergence tolerance of l1 proximal
primaldual: #solve the constrained problem
# Following is only accepted when MPI is used
mpiAlgorithm: serial-equivalent # one of none, serial-equivalent, random-updates
# Be careful changing the following unless you know what you are doing
relVarianceConvergence: 1e-3 # (>0) relative converegence tolerance of objective function (we suggest at most 1e-3, but smaller might produce better results)
epsilonConvergenceScaling: 1 # (>0) l2 convergence tolerance for residuals (we suggest 1)
precondition_iters: 0 #use sampling density (based on field of view) to accelerate primal dual (we recommend 1 or 0)
stepsize: # Stepsize, choose value between: default, variable, 100, x
update_tolerance: 1e-1 #when predicted stepsize is greater than this value, it will be updated
update_iters: 0 #how many iterations to allow updating of step size (we recommend 0 to 50)
|
