def buildNoisePadImage(self, noise_pad_size, noise_pad, rng):
"""A helper function that builds the `pad_image` from the given `noise_pad` specification.
"""
# Make it with the same dtype as the image
pad_image = galsim.Image(noise_pad_size,
noise_pad_size,
dtype=self.image.dtype)
# Figure out what kind of noise to apply to the image
if isinstance(noise_pad, float):
noise = galsim.GaussianNoise(rng, sigma=np.sqrt(noise_pad))
elif isinstance(noise_pad,
galsim.correlatednoise._BaseCorrelatedNoise):
noise = noise_pad.copy(rng=rng)
elif isinstance(noise_pad, galsim.Image):
noise = galsim.CorrelatedNoise(noise_pad, rng)
elif self.use_cache and noise_pad in InterpolatedImage._cache_noise_pad:
noise = InterpolatedImage._cache_noise_pad[noise_pad]
if rng:
# Make sure that we are using a specified RNG by resetting that in this cached
# CorrelatedNoise instance, otherwise preserve the cached RNG
noise = noise.copy(rng=rng)
elif isinstance(noise_pad, str):
noise = galsim.CorrelatedNoise(galsim.fits.read(noise_pad), rng)
if self.use_cache:
InterpolatedImage._cache_noise_pad[noise_pad] = noise
else:
raise ValueError(
"Input noise_pad must be a float/int, a CorrelatedNoise, Image, or filename "
+ "containing an image to use to make a CorrelatedNoise!")
# Add the noise
pad_image.addNoise(noise)
return pad_image
def test_corr_padding_cf():
"""Test for correlated noise padding of InterpolatedImage."""
import time
t1 = time.time()
imgfile = 'fits_files/blankimg.fits'
orig_nx = 147
orig_ny = 124
orig_seed = 151241
rng = galsim.BaseDeviate(orig_seed)
# Make an ImageCorrFunc
cf = galsim.CorrelatedNoise(rng, galsim.fits.read(imgfile))
# first, make the base image
orig_img = galsim.ImageF(orig_nx, orig_ny, scale=1.)
gal = galsim.Gaussian(sigma=2.5, flux=100.)
gal.drawImage(orig_img, method='no_pixel')
for iter in range(n_iter):
# make it into an InterpolatedImage padded with cf
int_im = galsim.InterpolatedImage(orig_img, noise_pad=cf)
# do it again with a particular seed
int_im = galsim.InterpolatedImage(orig_img, rng = galsim.GaussianDeviate(orig_seed),
noise_pad = cf)
# repeat
int_im = galsim.InterpolatedImage(orig_img, rng = galsim.GaussianDeviate(orig_seed),
noise_pad = cf)
t2 = time.time()
print('time for %s = %.2f'%(funcname(),t2-t1))
def check_symm_noise(noise_image, msg):
# A helper funciton to see if a noise image has 4-fold symmetric noise.
im2 = noise_image.copy()
# Clear out any wcs to make the test simpler
im2.wcs = galsim.PixelScale(1.)
noise = galsim.CorrelatedNoise(im2)
cf = noise.drawImage(galsim.Image(bounds=galsim.BoundsI(-1,1,-1,1), scale=1))
# First check the variance
print('variance: ',cf(0,0), noise.getVariance())
np.testing.assert_almost_equal(cf(0,0)/noise.getVariance(), 1.0, decimal=VAR_NDECIMAL,
err_msg=msg + ':: noise variance is wrong.')
cf_plus = np.array([ cf(1,0), cf(-1,0), cf(0,1), cf(0,-1) ])
cf_cross = np.array([ cf(1,1), cf(-1,-1), cf(-1,1), cf(1,-1) ])
print('plus pattern: ',cf_plus)
print('diff relative to dc: ',(cf_plus-np.mean(cf_plus))/cf(0,0))
print('cross pattern: ',cf_cross)
print('diff relative to dc: ',(cf_cross-np.mean(cf_cross))/cf(0,0))
# For now, don't make these asserts. Just print whether they will pass or fail.
if True:
if np.all(np.abs((cf_plus-np.mean(cf_plus))/cf(0,0)) < 0.01):
print('plus test passes')
else:
print('*** FAIL ***')
print(msg + ': plus pattern is not constant')
if np.all(np.abs((cf_cross-np.mean(cf_cross))/cf(0,0)) < 0.01):
print('cross test passes')
else:
print('*** FAIL ***')
print(msg + ': cross pattern is not constant')
else:
np.testing.assert_almost_equal((cf_plus-np.mean(cf_plus))/cf(0,0), 0.0, decimal=2,
err_msg=msg + ': plus pattern is not constant')
np.testing.assert_almost_equal((cf_cross-np.mean(cf_cross))/cf(0,0), 0.0, decimal=2,
err_msg=msg + ': cross pattern is not constant')
def buildNoisePadImage(self, pad_factor, noise_pad, rng):
"""A helper function that builds the pad_image from the given noise_pad specification.
"""
import numpy as np
if pad_factor <= 0.:
pad_factor = galsim._galsim.getDefaultPadFactor()
padded_size = int(np.ceil(np.max(self.orig_image.array.shape) * pad_factor))
if isinstance(self.orig_image, galsim.BaseImageF):
pad_image = galsim.ImageF(padded_size, padded_size)
if isinstance(self.orig_image, galsim.BaseImageD):
pad_image = galsim.ImageD(padded_size, padded_size)
# Figure out what kind of noise to apply to the image
if isinstance(noise_pad, float):
noise = galsim.GaussianNoise(rng, sigma = np.sqrt(noise_pad))
elif isinstance(noise_pad, galsim.correlatednoise._BaseCorrelatedNoise):
noise = noise_pad.copy()
if rng: # Let a user supplied RNG take precedence over that in user CN
noise.setRNG(rng)
elif isinstance(noise_pad,galsim.BaseImageF) or isinstance(noise_pad,galsim.BaseImageD):
noise = galsim.CorrelatedNoise(rng, noise_pad)
elif self.use_cache and noise_pad in InterpolatedImage._cache_noise_pad:
noise = InterpolatedImage._cache_noise_pad[noise_pad]
if rng:
# Make sure that we are using a specified RNG by resetting that in this cached
# CorrelatedNoise instance, otherwise preserve the cached RNG
noise.setRNG(rng)
elif isinstance(noise_pad, str):
noise = galsim.CorrelatedNoise(rng, galsim.fits.read(noise_pad))
if self.use_cache:
InterpolatedImage._cache_noise_pad[noise_pad] = noise
else:
raise ValueError(
"Input noise_pad must be a float/int, a CorrelatedNoise, Image, or filename "+
"containing an image to use to make a CorrelatedNoise!")
pad_image.addNoise(noise)
return pad_image
def check_noise(noise_image, noise, msg):
# A helper function to check that the current noise in the image is properly described
# by the given CorrelatedNoise object
noise2 = galsim.CorrelatedNoise(noise_image)
print('noise = ',noise)
print('noise2 = ',noise2)
np.testing.assert_almost_equal(noise.getVariance(), noise2.getVariance(),
decimal=CHECKNOISE_NDECIMAL,
err_msg=msg + ': variance does not match.')
cf_im1 = galsim.Image(8,8, wcs=noise_image.wcs)
cf_im2 = cf_im1.copy()
noise.drawImage(image=cf_im1)
noise2.drawImage(image=cf_im2)
np.testing.assert_almost_equal(cf_im1.array, cf_im2.array,
decimal=CHECKNOISE_NDECIMAL,
err_msg=msg + ': image of cf does not match.')
def make_cf(args):
bls = Table.read(args.blank_file, format='ascii.basic')
for f, filt in enumerate(args.filter_names):
print "Making Correlation function for ", filt
q, = np.where(bls['FILTER'] == filt)
bl = q[0]
xmin, xmax = bls['XMIN'][bl], bls['XMAX'][bl]
ymin, ymax = bls['YMIN'][bl], bls['YMAX'][bl]
hdu = pyfits.open(bls['FILE'][bl])
bl_dat = hdu[0].data[ymin:ymax, xmin:xmax]
hdu.close()
rng = galsim.BaseDeviate(123456)
img = galsim.Image(bl_dat)
cn = galsim.CorrelatedNoise(img, rng=rng, scale=0.03)
image = galsim.ImageD(80, 80, scale=0.03)
cf = cn.drawImage(image)
name = args.out_path + args.out_name.replace('filter',
args.file_filter_name[f])
print "Savings fits file at ", name
pyfits.writeto(name, cf.array, clobber=True)
print "Unpadded convolved image bounds = " + str(convimage1.bounds)
convimage3_padded = galsim.ImageD(
int(largeim_size * size_factor) + 32,
int(largeim_size * size_factor) + 32)
# Set the scales of convimage2 & 3 to be 0.18 so that addNoise() works correctly
convimage2.scale = 0.18
convimage3_padded.scale = 0.18
print "Padded convolved image bounds = " + str(convimage3_padded.bounds)
print ""
# We draw, calculate a correlation function for the resulting field, and repeat to get an
# average over nsum_test trials
cimobj.draw(convimage1, dx=0.18, normalization='sb')
cn_test1 = galsim.CorrelatedNoise(gd,
convimage1,
dx=0.18,
correct_periodicity=False,
subtract_mean=True,
correct_sample_bias_prototype=False)
testim1 = galsim.ImageD(smallim_size, smallim_size)
cn_test1.draw(testim1, dx=0.18)
convimage2.addNoise(
conv_cn
) # Now we make a comparison by simply adding noise from conv_cn
cn_test2 = galsim.CorrelatedNoise(gd,
convimage2,
dx=0.18,
correct_periodicity=False,
subtract_mean=True,
correct_sample_bias_prototype=False)
testim2 = galsim.ImageD(smallim_size, smallim_size)
cn2 = galsim.getCOSMOSNoise(ud, CFIMFILE_UNS, dx_cosmos=1.)
testim1 = galsim.ImageD(7, 7)
testim2 = galsim.ImageD(7, 7)
var1 = 0.
var2 = 0.
noisearrays = cPickle.load(open(NOISEIMFILE, 'rb'))
for noisearray, i in zip(noisearrays, range(len(noisearrays))):
noise1 = galsim.ImageViewD((noisearray.copy()).astype(np.float64), scale=1.)
noise2 = galsim.ImageViewD((noisearray.copy()).astype(np.float64), scale=1.)
cn1.applyWhiteningTo(noise1)
cn2.applyWhiteningTo(noise2)
var1 += noise1.array.var()
var2 += noise2.array.var()
cntest1 = galsim.CorrelatedNoise(ud, noise1)
cntest2 = galsim.CorrelatedNoise(ud, noise2)
cntest1.draw(testim1, dx=1., add_to_image=True)
cntest2.draw(testim2, dx=1., add_to_image=True)
print "Done "+str(i + 1)+"/"+str(len(noisearrays))
testim1 /= len(noisearrays)
testim2 /= len(noisearrays)
var1 /= len(noisearrays)
var2 /= len(noisearrays)
print var1, var2
delx = np.arange(7) - 3.
import matplotlib.pyplot as plt
A simple script used to quickly generate plots for the discussion of which interpolant should be
used to describe correlation functions. See the discussion at
https://github.com/GalSim-developers/GalSim/pull/452#discussion-diff-5701561
"""
import numpy as np
import matplotlib.pyplot as plt
import galsim
CFSIZE=9
UPSAMPLING = 3
rng = galsim.BaseDeviate(752424)
gd = galsim.GaussianNoise(rng)
noise_image = galsim.ImageD(CFSIZE, CFSIZE)
noise_image.addNoise(gd)
cn = galsim.CorrelatedNoise(rng, noise_image, x_interpolant=galsim.Nearest(tol=1.e-4), dx=1.)
test_image = galsim.ImageD(
2 * UPSAMPLING * CFSIZE, 2 * UPSAMPLING * CFSIZE, scale=1. / float(UPSAMPLING))
cn.applyRotation(30. * galsim.degrees)
cn.draw(test_image)
plt.clf()
plt.pcolor(test_image.array)
plt.colorbar()
plt.savefig('cf_nearest.png')
plt.show()
def __init__(self,
real_galaxy_catalog,
index=None,
id=None,
random=False,
rng=None,
x_interpolant=None,
k_interpolant=None,
flux=None,
pad_factor=0,
noise_pad=False,
pad_image=None,
use_cache=True,
gsparams=None):
import pyfits
import numpy as np
# Code block below will be for galaxy selection; not all are currently implemented. Each
# option must return an index within the real_galaxy_catalog.
if index is not None:
if id is not None or random is True:
raise AttributeError(
'Too many methods for selecting a galaxy!')
use_index = index
elif id is not None:
if random is True:
raise AttributeError(
'Too many methods for selecting a galaxy!')
use_index = real_galaxy_catalog._get_index_for_id(id)
elif random is True:
if rng is None:
uniform_deviate = galsim.UniformDeviate()
elif isinstance(rng, galsim.BaseDeviate):
uniform_deviate = galsim.UniformDeviate(rng)
else:
raise TypeError(
"The rng provided to RealGalaxy constructor is not a BaseDeviate"
)
use_index = int(real_galaxy_catalog.nobjects * uniform_deviate())
else:
raise AttributeError('No method specified for selecting a galaxy!')
# read in the galaxy, PSF images; for now, rely on pyfits to make I/O errors. Should
# consider exporting this code into fits.py in some function that takes a filename and HDU,
# and returns an ImageView
gal_image = real_galaxy_catalog.getGal(use_index)
PSF_image = real_galaxy_catalog.getPSF(use_index)
# choose proper interpolant
if x_interpolant is None:
lan5 = galsim.Lanczos(5, conserve_flux=True, tol=1.e-4)
self.x_interpolant = galsim.InterpolantXY(lan5)
else:
self.x_interpolant = galsim.utilities.convert_interpolant_to_2d(
x_interpolant)
if k_interpolant is None:
self.k_interpolant = galsim.InterpolantXY(
galsim.Quintic(tol=1.e-4))
else:
self.k_interpolant = galsim.utilities.convert_interpolant_to_2d(
k_interpolant)
# read in data about galaxy from FITS binary table; store as normal attributes of RealGalaxy
# save any other relevant information as instance attributes
self.catalog_file = real_galaxy_catalog.file_name
self.index = use_index
self.pixel_scale = float(real_galaxy_catalog.pixel_scale[use_index])
# handle padding by an image
specify_size = False
padded_size = gal_image.getPaddedSize(pad_factor)
if pad_image is not None:
specify_size = True
if isinstance(pad_image, str):
pad_image = galsim.fits.read(pad_image)
if (not isinstance(pad_image, galsim.BaseImageF)
and not isinstance(pad_image, galsim.BaseImageD)):
raise ValueError(
"Supplied pad_image is not one of the allowed types!")
# If an image was supplied directly or from a file, check its size:
# Cannot use if too small.
# Use to define the final image size otherwise.
deltax = ((1 + pad_image.getXMax() - pad_image.getXMin()) -
(1 + gal_image.getXMax() - gal_image.getXMin()))
deltay = ((1 + pad_image.getYMax() - pad_image.getYMin()) -
(1 + gal_image.getYMax() - gal_image.getYMin()))
if deltax < 0 or deltay < 0:
raise RuntimeError("Image supplied for padding is too small!")
if pad_factor != 1. and pad_factor != 0.:
import warnings
msg = "Warning: ignoring specified pad_factor because user also specified\n"
msg += " an image to use directly for the padding."
warnings.warn(msg)
else:
if isinstance(gal_image, galsim.BaseImageF):
pad_image = galsim.ImageF(padded_size, padded_size)
if isinstance(gal_image, galsim.BaseImageD):
pad_image = galsim.ImageD(padded_size, padded_size)
# Set up the GaussianDeviate if not provided one, or check that the user-provided one
# is of a valid type. Note if random was selected, we can use that uniform_deviate safely.
if random is True:
gaussian_deviate = galsim.GaussianDeviate(uniform_deviate)
else:
if rng is None:
gaussian_deviate = galsim.GaussianDeviate()
elif isinstance(rng, galsim.BaseDeviate):
# Even if it's already a GaussianDeviate, we still want to make a new Gaussian
# deviate that would generate the same sequence, because later we change the sigma
# and we don't want to change it for the original one that was passed in. So don't
# distinguish between GaussianDeviate and the other BaseDeviates here.
gaussian_deviate = galsim.GaussianDeviate(rng)
else:
raise TypeError(
"rng provided to RealGalaxy constructor is not a BaseDeviate"
)
# handle noise-padding options
try:
noise_pad = galsim.config.value._GetBoolValue(noise_pad, '')
except:
pass
if noise_pad:
self.pad_variance = float(real_galaxy_catalog.variance[use_index])
# Check, is it "True" or something else? If True, we use Gaussian uncorrelated noise
# using the stored variance in the catalog. Otherwise, if it's a CorrelatedNoise we use
# it directly; if it's an Image of some sort we use it to make a CorrelatedNoise; if
# it's a string, we read in the image from file and make a CorrelatedNoise.
if type(noise_pad) is not bool:
if isinstance(noise_pad,
galsim.correlatednoise._BaseCorrelatedNoise):
cn = noise_pad.copy()
if rng: # Let user supplied RNG take precedence over that in user CN
cn.setRNG(gaussian_deviate)
# This small patch may have different overall variance, so rescale while
# preserving the correlation structure by default
cn.setVariance(self.pad_variance)
elif (isinstance(noise_pad, galsim.BaseImageF)
or isinstance(noise_pad, galsim.BaseImageD)):
cn = galsim.CorrelatedNoise(gaussian_deviate, noise_pad)
elif use_cache and noise_pad in RealGalaxy._cache_noise_pad:
cn = RealGalaxy._cache_noise_pad[noise_pad]
# Make sure that we are using the desired RNG by resetting that in this cached
# CorrelatedNoise instance
if rng:
cn.setRNG(gaussian_deviate)
# This small patch may have different overall variance, so rescale while
# preserving the correlation structure
cn.setVariance(self.pad_variance)
elif isinstance(noise_pad, str):
tmp_img = galsim.fits.read(noise_pad)
cn = galsim.CorrelatedNoise(gaussian_deviate, tmp_img)
if use_cache:
RealGalaxy._cache_noise_pad[noise_pad] = cn
# This small patch may have different overall variance, so rescale while
# preserving the correlation structure
cn.setVariance(self.pad_variance)
else:
raise RuntimeError(
"noise_pad must be either a bool, CorrelatedNoise, Image, "
+ "or a filename for reading in an Image")
# Set the GaussianDeviate sigma
gaussian_deviate.setSigma(np.sqrt(self.pad_variance))
# populate padding image with noise field
if type(noise_pad) is bool:
pad_image.addNoise(galsim.DeviateNoise(gaussian_deviate))
else:
pad_image.addNoise(cn)
else:
self.pad_variance = 0.
# Now we have to check: was the padding determined using pad_factor? Or by passing in an
# image for padding? Treat these cases differently:
# (1) If the former, then we can simply have the C++ handle the padding process.
# (2) If the latter, then we have to do the padding ourselves, and pass the resulting image
# to the C++ with pad_factor explicitly set to 1.
if specify_size is False:
# Make the SBInterpolatedImage out of the image.
self.original_image = galsim.SBInterpolatedImage(
gal_image,
xInterp=self.x_interpolant,
kInterp=self.k_interpolant,
dx=self.pixel_scale,
pad_factor=pad_factor,
pad_image=pad_image,
gsparams=gsparams)
else:
# Leave the original image as-is. Instead, we shift around the image to be used for
# padding. Find out how much x and y margin there should be on lower end:
x_marg = int(np.round(0.5 * deltax))
y_marg = int(np.round(0.5 * deltay))
# Now reset the pad_image to contain the original image in an even way
pad_image = pad_image.view()
pad_image.setScale(self.pixel_scale)
pad_image.setOrigin(gal_image.getXMin() - x_marg,
gal_image.getYMin() - y_marg)
# Set the central values of pad_image to be equal to the input image
pad_image[gal_image.bounds] = gal_image
self.original_image = galsim.SBInterpolatedImage(
pad_image,
xInterp=self.x_interpolant,
kInterp=self.k_interpolant,
dx=self.pixel_scale,
pad_factor=1.,
gsparams=gsparams)
# also make the original PSF image, with far less fanfare: we don't need to pad with
# anything interesting.
self.original_PSF = galsim.SBInterpolatedImage(
PSF_image,
xInterp=self.x_interpolant,
kInterp=self.k_interpolant,
dx=self.pixel_scale,
gsparams=gsparams)
# recalculate Fourier-space attributes rather than using overly-conservative defaults
self.original_image.calculateStepK()
self.original_image.calculateMaxK()
self.original_PSF.calculateStepK()
self.original_PSF.calculateMaxK()
if flux != None:
self.original_image.setFlux(flux)
self.original_image.__class__ = galsim.SBTransform # correctly reflect SBProfile change
self.original_PSF.setFlux(1.0)
self.original_PSF.__class__ = galsim.SBTransform # correctly reflect SBProfile change
# Calculate the PSF "deconvolution" kernel
psf_inv = galsim.SBDeconvolve(self.original_PSF, gsparams=gsparams)
# Initialize the SBProfile attribute
GSObject.__init__(
self,
galsim.SBConvolve([self.original_image, psf_inv],
gsparams=gsparams))
print "Unpadded convolved image bounds = " + str(convimage1.bounds)
convimage3_padded = galsim.ImageD(
int(largeim_size * size_factor) + 32,
int(largeim_size * size_factor) + 32)
# Set the scales of convimage2 & 3 to be 0.18 so that addNoise() works correctly
convimage2.scale = 0.18
convimage3_padded.scale = 0.18
print "Padded convolved image bounds = " + str(convimage3_padded.bounds)
print ""
# We draw, calculate a correlation function for the resulting field, and repeat to get an
# average over nsum_test trials
cimobj.draw(convimage1, dx=0.18, normalization='sb')
cn_test1 = galsim.CorrelatedNoise(convimage1,
gd,
dx=0.18,
correct_periodicity=False,
subtract_mean=True)
testim1 = galsim.ImageD(smallim_size, smallim_size)
cn_test1.draw(testim1, dx=0.18)
convimage2.addNoise(
conv_cn
) # Now we make a comparison by simply adding noise from conv_cn
cn_test2 = galsim.CorrelatedNoise(convimage2,
gd,
dx=0.18,
correct_periodicity=False,
subtract_mean=True)
testim2 = galsim.ImageD(smallim_size, smallim_size)
cn_test2.draw(testim2, dx=0.18)
def test_corr_padding():
"""Test for correlated noise padding of InterpolatedImage."""
import time
t1 = time.time()
# Set up some defaults for tests.
decimal_precise=4
decimal_coarse=2
imgfile = 'fits_files/blankimg.fits'
orig_nx = 187
orig_ny = 164
big_nx = 319
big_ny = 322
orig_seed = 151241
# Read in some small image of a noise field from HST.
# Rescale it to have a decently large amplitude for the purpose of doing these tests.
im = 1.e2*galsim.fits.read(imgfile)
# Make a CorrrlatedNoise out of it.
cn = galsim.CorrelatedNoise(im, galsim.BaseDeviate(orig_seed))
# first, make a noise image
orig_img = galsim.ImageF(orig_nx, orig_ny, scale=1.)
orig_img.addNoise(cn)
# make it into an InterpolatedImage with some zero-padding
# (note that default is zero-padding, by factors of several)
int_im = galsim.InterpolatedImage(orig_img)
# draw into a larger image
big_img = galsim.ImageF(big_nx, big_ny)
int_im.draw(big_img, scale=1.)
# check that variance is diluted by expected amount - should be exact, so check precisely!
big_var_expected = np.var(orig_img.array)*float(orig_nx*orig_ny)/(big_nx*big_ny)
np.testing.assert_almost_equal(np.var(big_img.array), big_var_expected, decimal=decimal_precise,
err_msg='Variance not diluted by expected amount when zero-padding')
# make it into an InterpolatedImage with noise-padding
int_im = galsim.InterpolatedImage(orig_img, rng = galsim.GaussianDeviate(orig_seed),
noise_pad = im, noise_pad_size = max(big_nx,big_ny))
# draw into a larger image
big_img = galsim.ImageF(big_nx, big_ny)
int_im.draw(big_img, scale=1.)
# check that variance is same as original - here, we cannot be too precise because the padded
# region is not huge and the comparison will be, well, noisy.
np.testing.assert_almost_equal(np.var(big_img.array), np.var(orig_img.array),
decimal=decimal_coarse,
err_msg='Variance not correct after padding image with correlated noise')
# check that if we pass in a RNG, it is actually used to pad with the same noise field
# basically, redo all of the above steps and draw into a new image, make sure it's the same as
# previous.
int_im = galsim.InterpolatedImage(
orig_img, rng=galsim.GaussianDeviate(orig_seed), noise_pad=cn,
noise_pad_size = max(big_nx,big_ny))
big_img_2 = galsim.ImageF(big_nx, big_ny)
int_im.draw(big_img_2, scale=1.)
np.testing.assert_array_almost_equal(big_img_2.array, big_img.array, decimal=decimal_precise,
err_msg='Cannot reproduce correlated noise-padded image with same choice of seed')
# Finally, check inputs:
# what if we give it a screwy way of defining the image padding?
try:
np.testing.assert_raises(ValueError,galsim.InterpolatedImage,orig_img,noise_pad=-1.)
except ImportError:
print 'The assert_raises tests require nose'
# also, check that whether we give it a string, image, or cn, it gives the same noise field
# (given the same random seed)
infile = 'fits_files/blankimg.fits'
inimg = galsim.fits.read(infile)
incf = galsim.CorrelatedNoise(inimg, galsim.GaussianDeviate()) # input RNG will be ignored below
int_im2 = galsim.InterpolatedImage(orig_img, rng=galsim.GaussianDeviate(orig_seed),
noise_pad=inimg, noise_pad_size = max(big_nx,big_ny))
int_im3 = galsim.InterpolatedImage(orig_img, rng=galsim.GaussianDeviate(orig_seed),
noise_pad=incf, noise_pad_size = max(big_nx,big_ny))
big_img2 = galsim.ImageF(big_nx, big_ny)
big_img3 = galsim.ImageF(big_nx, big_ny)
int_im2.draw(big_img2, scale=1.)
int_im3.draw(big_img3, scale=1.)
np.testing.assert_equal(big_img2.array, big_img3.array,
err_msg='Diff ways of specifying correlated noise give diff answers')
t2 = time.time()
print 'time for %s = %.2f'%(funcname(),t2-t1)
# function that is NPIX by NPIX
if not os.path.isfile(
CFIMFILE
): # If the CFIMFILE already exists skip straight through to the plots
# Read in the pickled images
noiseims = cPickle.load(open(NOISEIMFILE, 'rb'))
# Loop through the images and sum the correlation functions
hst_ncf = None
bd = galsim.BaseDeviate(12345) # Seed is basically unimportant here
for noiseim in noiseims:
noiseim = noiseim.astype(np.float64)
if hst_ncf is None:
# Initialize the HST noise correlation function using the first image
hst_ncf = galsim.CorrelatedNoise(bd,
galsim.ImageViewD(noiseim),
correct_periodicity=True,
subtract_mean=SUBTRACT_MEAN)
else:
hst_ncf += galsim.CorrelatedNoise(bd,
galsim.ImageViewD(noiseim),
correct_periodicity=True,
subtract_mean=SUBTRACT_MEAN)
hst_ncf /= float(len(noiseims))
# Draw and plot an output image of the resulting correlation function
cfimage = galsim.ImageD(NPIX, NPIX)
hst_ncf.draw(cfimage, dx=1.)
# Save this to the output filename specified in the script header
cfimage.write(CFIMFILE)
else:
cfimage = galsim.fits.read(CFIMFILE)
failed1 = 0
failed2 = 0
failed3 = 0
failed4 = 0
failed5 = 0
failed6 = 0
failed7 = 0
TESTSTART = 25
TESTEND = 125
for dim in range(TESTSTART, TESTEND):
noise_array = np.random.randn(dim, dim)
noise_image = galsim.ImageViewD(noise_array)
cn1 = galsim.CorrelatedNoise(rng, noise_image)
cn2 = galsim.CorrelatedNoise(rng, noise_image, correct_periodicity=False)
# First try with (default), then without, periodicity correction
test_image1.setZero()
try:
cn1.applyTo(test_image1)
except RuntimeError:
failed1 += 1
test_image2.setZero()
try:
cn2.applyTo(test_image2)
except RuntimeError:
failed2 += 1
# Then try calculating the PS by hand, in the same manner as the CorrelatedNoise internals
noiseft = np.fft.fft2(noise_array)
ps = np.abs(noiseft)**2
def test_metacal_tracking():
"""Test that the noise tracking works for the metacal use case involving deconvolution and
reconvolution by almost the same PSF.
This test is similar to the above test_uncorrelated_noise_tracking, except the modifications
are based on what is done for the metacal procedure.
"""
import math
def check_noise(noise_image, noise, msg):
# A helper function to check that the current noise in the image is properly described
# by the given CorrelatedNoise object
noise2 = galsim.CorrelatedNoise(noise_image)
print('noise = ', noise)
print('noise2 = ', noise2)
np.testing.assert_almost_equal(noise.getVariance(),
noise2.getVariance(),
decimal=CHECKNOISE_NDECIMAL,
err_msg=msg +
': variance does not match.')
cf_im1 = galsim.Image(8, 8, wcs=noise_image.wcs)
cf_im2 = cf_im1.copy()
noise.drawImage(image=cf_im1)
noise2.drawImage(image=cf_im2)
np.testing.assert_almost_equal(cf_im1.array,
cf_im2.array,
decimal=CHECKNOISE_NDECIMAL,
err_msg=msg +
': image of cf does not match.')
def check_symm_noise(noise_image, msg):
# A helper funciton to see if a noise image has 4-fold symmetric noise.
im2 = noise_image.copy()
# Clear out any wcs to make the test simpler
im2.wcs = galsim.PixelScale(1.)
noise = galsim.CorrelatedNoise(im2)
cf = noise.drawImage(
galsim.Image(bounds=galsim.BoundsI(-1, 1, -1, 1), scale=1))
# First check the variance
print('variance: ', cf(0, 0), noise.getVariance())
np.testing.assert_almost_equal(cf(0, 0) / noise.getVariance(),
1.0,
decimal=VAR_NDECIMAL,
err_msg=msg +
':: noise variance is wrong.')
cf_plus = np.array([cf(1, 0), cf(-1, 0), cf(0, 1), cf(0, -1)])
cf_cross = np.array([cf(1, 1), cf(-1, -1), cf(-1, 1), cf(1, -1)])
print('plus pattern: ', cf_plus)
print('diff relative to dc: ', (cf_plus - np.mean(cf_plus)) / cf(0, 0))
print('cross pattern: ', cf_cross)
print('diff relative to dc: ',
(cf_cross - np.mean(cf_cross)) / cf(0, 0))
# For now, don't make these asserts. Just print whether they will pass or fail.
if True:
if np.all(np.abs((cf_plus - np.mean(cf_plus)) / cf(0, 0)) < 0.01):
print('plus test passes')
else:
print('*** FAIL ***')
print(msg + ': plus pattern is not constant')
if np.all(
np.abs((cf_cross - np.mean(cf_cross)) / cf(0, 0)) < 0.01):
print('cross test passes')
else:
print('*** FAIL ***')
print(msg + ': cross pattern is not constant')
else:
np.testing.assert_almost_equal(
(cf_plus - np.mean(cf_plus)) / cf(0, 0),
0.0,
decimal=2,
err_msg=msg + ': plus pattern is not constant')
np.testing.assert_almost_equal(
(cf_cross - np.mean(cf_cross)) / cf(0, 0),
0.0,
decimal=2,
err_msg=msg + ': cross pattern is not constant')
seed = 1234567 # For use as a unit test, we need a specific seed
#seed = 0 # During testing, it's useful to see how numbers flop around to know if
# something is systematic or random.
rng = galsim.BaseDeviate(seed)
noise_var = 1.3
im_size = 1024
dg = 0.1 # This is bigger than metacal would use, but it makes the test easier.
# Use a non-trivial wcs...
#wcs = galsim.JacobianWCS(0.26, 0.04, -0.04, -0.26) # Rotation + flip. No shear.
wcs = galsim.JacobianWCS(0.26, 0.03, 0.08, -0.21) # Fully complex
#dudx = 0.12*0.26
#dudy = 1.10*0.26
#dvdx = -0.915*0.26
#dvdy = -0.04*0.26
#wcs = galsim.JacobianWCS(dudx, dudy, dvdx, dvdy) # Even more extreme
# And an asymmetric PSF
#orig_psf = galsim.Gaussian(fwhm=0.9).shear(g1=0.05, g2=0.03)
# This one is small enough not to be fully Nyquist sampled, which makes things harder.
orig_psf = galsim.Gaussian(fwhm=0.7).shear(g1=0.05, g2=0.03)
# pixel is the pixel in world coords
pixel = wcs.toWorld(galsim.Pixel(scale=1))
pixel_inv = galsim.Deconvolve(pixel)
psf_image = orig_psf.drawImage(nx=im_size, ny=im_size, wcs=wcs)
# Metacal only has access to the PSF as an image, so use this from here on.
psf = galsim.InterpolatedImage(psf_image)
psf_nopix = galsim.Convolve([psf, pixel_inv])
psf_inv = galsim.Deconvolve(psf)
# Not what is done currently, but using a smoother target PSF helps make sure the
# zeros in the deconvolved PSF get adequately zeroed out.
def get_target_psf(psf):
dk = 0.1 # The resolution in k space for the KImage
small_kval = 1.e-2 # Find the k where the given psf hits this kvalue
smaller_kval = 3.e-3 # Target PSF will have this kvalue at the same k
kim = psf.drawKImage(scale=dk)
karr_r = kim.real.array
# Find the smallest r where the kval < small_kval
nk = karr_r.shape[0]
kx, ky = np.meshgrid(np.arange(-nk / 2, nk / 2),
np.arange(-nk / 2, nk / 2))
ksq = (kx**2 + ky**2) * dk**2
ksq_max = np.min(ksq[karr_r < small_kval * psf.flux])
# We take our target PSF to be the (round) Gaussian that is even smaller at this ksq
# exp(-0.5 * ksq_max * sigma_sq) = smaller_kval
sigma_sq = -2. * np.log(smaller_kval) / ksq_max
return galsim.Gaussian(sigma=np.sqrt(sigma_sq))
# The target PSF dilates the part without the pixel, but reconvolve by the real pixel.
#psf_target_nopix = psf_nopix.dilate(1. + 2.*dg)
#psf_target_nopix = orig_psf.dilate(1. + 4.*dg)
psf_target_nopix = get_target_psf(psf_nopix.shear(g1=dg))
print('PSF target HLR = ', psf_target_nopix.calculateHLR())
print('PSF target FWHM = ', psf_target_nopix.calculateFWHM())
psf_target = galsim.Convolve([psf_target_nopix, pixel])
# Make an image of pure (white) Gaussian noise
# Normally, there would be a galaxy in this image, but for the tests, we just have noise.
obs_image = galsim.Image(im_size, im_size, init_value=0, wcs=wcs)
obs_image.addNoise(
galsim.GaussianNoise(rng=rng, sigma=math.sqrt(noise_var)))
# The noise on this image should be describable as an UncorrelatedNoise object:
noise = galsim.UncorrelatedNoise(variance=noise_var, wcs=wcs)
check_noise(obs_image, noise, 'initial UncorrelatedNoise model is wrong')
# Make an InterpolatedImage profile to use for manipulating this image
# We can get away with no padding here, since our image is so large, but normally, you would
# probably want to pad this with noise padding.
ii = galsim.InterpolatedImage(obs_image, pad_factor=1)
ii.noise = noise
# If we draw it back, the attached noise attribute should still be correct
check_noise(ii.drawImage(obs_image.copy(), method='no_pixel'), ii.noise,
'noise model is wrong for InterpolatedImage')
# Here is the metacal process for which we want to understand the noise.
# We'll try a few different methods.
shear = galsim.Shear(g1=dg)
sheared_obj = galsim.Convolve(ii, psf_inv).shear(shear)
final_obj = galsim.Convolve(psf_target, sheared_obj)
final_image = final_obj.drawImage(obs_image.copy(), method='no_pixel')
try:
check_symm_noise(final_image, 'Initial image')
#didnt_fail = True
# This bit doesn't work while we are not actually raising exceptions in check_symm_noise
# So we expect to see **FAIL** text at this point.
print(
'The above tests are expected to **FAIL**. This is not a problem.'
)
didnt_fail = False
except AssertionError as e:
print('As expected initial image fails symmetric noise test:')
print(e)
didnt_fail = False
if didnt_fail:
assert False, 'Initial image was expected to fail symmetric noise test, but passed.'
if True:
print('\n\nStrategy 1:')
# Strategy 1: Use the noise attribute attached to ii and use it to either whiten or
# symmetrize the noise in the final image.
# Note: The check_noise tests fail. I think because the convolve and deconvolve impose
# a maxk = that of the psf. Which is too small for an accurate rendering of the
# correlation function (here just an autocorrelation of a Pixel.
# The whiten tests kind of work, but they add a lot of extra noise. Much more than
# strategy 4 below. So the level of correlation remaining is pretty well below the
# dc variance. Symmetrize doesn't add much noise, but the residual correlation is about
# the same, which means it doesn't pass the test relative to the lower dc variance.
# First, deconvolve and reconvolve by the same PSF:
test_obj = galsim.Convolve([ii, psf, psf_inv])
# This fails...
if False:
check_noise(
test_obj.drawImage(obs_image.copy(),
method='no_pixel'), test_obj.noise,
'noise model is wrong after convolve/deconvolve by psf')
# Now use a slightly dilated PSF for the reconvolution:
test_obj = galsim.Convolve([ii, psf_target, psf_inv])
if False:
check_noise(
test_obj.drawImage(obs_image.copy(), method='no_pixel'),
test_obj.noise, 'noise model is wrong for dilated target psf')
# Finally, include the shear step. This was done above with sheared_obj, final_obj.
if False:
check_noise(final_image, final_obj.noise,
'noise model is wrong when including small shear')
# If we whiten using this noise model, we should get back to white noise.
t3 = time.time()
final_image2 = final_image.copy() # Don't clobber the original
final_var = final_image2.whitenNoise(final_obj.noise)
t4 = time.time()
print('Noise tracking method with whiten: final_var = ', final_var)
print('Check: direct variance = ', np.var(final_image2.array))
check_symm_noise(final_image2, 'noise whitening does not work')
print('Time for noise tracking with whiten = ', t4 - t3)
# Using symmetrizeNoise should add less noise, but also work.
t3 = time.time()
final_image2 = final_image.copy()
final_var = final_image2.symmetrizeNoise(final_obj.noise)
t4 = time.time()
print('Noise tracking method with symmetrize: final_var = ', final_var)
print('Check: direct variance = ', np.var(final_image2.array))
check_symm_noise(final_image2, 'noise symmetrizing does not work')
print('Time for noise tracking with symmetrize = ', t4 - t3)
if True:
print('\n\nStrategy 2:')
# Strategy 2: Don't trust the noise tracking. Track a noise image through the same process
# and then measure the noise from that image. Use it to either whiten or
# symmetrize the noise in the final image.
# Note: This method doesn't work any better. The added noise for whitening is even more
# than strategy 1. And the remaining correlations are still similarly significant for the
# symmetrize version. A little smaller than strategy 1, but not enough to pass our tests.
# Make another noise image, since we don't actually have access to a pure noise image
# for real objects. But we should be able to estimate the variance in the image.
t3 = time.time()
noise_image = galsim.Image(im_size, im_size, init_value=0, wcs=wcs)
noise_image.addNoise(
galsim.GaussianNoise(rng=rng, sigma=math.sqrt(noise_var)))
noise_ii = galsim.InterpolatedImage(noise_image, pad_factor=1)
sheared_noise_obj = galsim.Convolve(noise_ii, psf_inv).shear(shear)
final_noise_obj = galsim.Convolve(psf_target, sheared_noise_obj)
final_noise_image = final_noise_obj.drawImage(obs_image.copy(),
method='no_pixel')
# Use this to construct an appropriate CorrelatedNoise object
noise = galsim.CorrelatedNoise(final_noise_image)
t4 = time.time()
final_image2 = final_image.copy()
final_var = final_image2.whitenNoise(noise)
t5 = time.time()
check_noise(
final_noise_image, noise,
'noise model is wrong when direct measuring the final noise image')
print('Direct noise method with whiten: final_var = ', final_var)
# Neither of these work currently, so maybe a bug in the whitening code?
# Or possibly in my logic here.
check_symm_noise(
final_image2,
'whitening the noise using direct noise model failed')
print('Time for direct noise with whitening = ', t5 - t3)
t6 = time.time()
final_image2 = final_image.copy()
final_var = final_image2.symmetrizeNoise(noise)
t7 = time.time()
print('Direct noise method with symmetrize: final_var = ', final_var)
check_symm_noise(
final_image2,
'symmetrizing the noise using direct noise model failed')
print('Time for direct noise with symmetrizing = ', t7 - t6 + t4 - t3)
if False:
print('\n\nStrategy 3:')
# Strategy 3: Make a noise field and do the same operations as we do to the main image
# except use the opposite shear value. Add this noise field to the final
# image to get a symmetric noise field.
# Note: This method works! But only for square pixels. However, they may be rotated
# or flipped. Just not sheared.
# Update: I think this method won't ever work for non-square pixels. The reason it works
# for square pixels is that in that case, it is equivalent to strategy 4.
t3 = time.time()
# Make another noise image
rev_image = galsim.Image(im_size, im_size, init_value=0, wcs=wcs)
rev_image.addNoise(
galsim.GaussianNoise(rng=rng, sigma=math.sqrt(noise_var)))
rev_ii = galsim.InterpolatedImage(rev_image, pad_factor=1)
rev_sheared_obj = galsim.Convolve(rev_ii, psf_inv).shear(-shear)
rev_final_obj = galsim.Convolve(psf_target, rev_sheared_obj)
rev_final_image = rev_final_obj.drawImage(obs_image.copy(),
method='no_pixel')
# Add the reverse-sheared noise image to the original image.
final_image2 = final_image + rev_final_image
t4 = time.time()
# The noise variance in the end should be 2x as large as the original
final_var = np.var(final_image2.array)
print('Reverse shear method: final_var = ', final_var)
check_symm_noise(final_image2, 'using reverse shear does not work')
print('Time for reverse shear method = ', t4 - t3)
if True:
print('\n\nStrategy 4:')
# Strategy 4: Make a noise field and do the same operations as we do to the main image,
# then rotate it by 90 degress and add it to the final image.
# This method works! Even for an arbitrarily sheared wcs.
t3 = time.time()
# Make another noise image
noise_image = galsim.Image(im_size, im_size, init_value=0, wcs=wcs)
noise_image.addNoise(
galsim.GaussianNoise(rng=rng, sigma=math.sqrt(noise_var)))
noise_ii = galsim.InterpolatedImage(noise_image, pad_factor=1)
noise_sheared_obj = galsim.Convolve(noise_ii, psf_inv).shear(shear)
noise_final_obj = galsim.Convolve(psf_target, noise_sheared_obj)
noise_final_image = noise_final_obj.drawImage(obs_image.copy(),
method='no_pixel')
# Rotate the image by 90 degrees
rot_noise_final_image = galsim.Image(
np.ascontiguousarray(np.rot90(noise_final_image.array)))
# Add the rotated noise image to the original image.
final_image2 = final_image + rot_noise_final_image
t4 = time.time()
# The noise variance in the end should be 2x as large as the original
final_var = np.var(final_image2.array)
print('Rotate image method: final_var = ', final_var)
check_symm_noise(final_image2,
'using rotated noise image does not work')
print('Time for rotate image method = ', t4 - t3)
if False:
print('\n\nStrategy 5:')
# Strategy 5: The same as strategy 3, except we target the effective net transformation
# done by strategy 4.
# I think this strategy probably can't work for non-square pixels, because the shear
# happens before the convolution by the PSF. And if the wcs is non-square, then the
# PSF is sheared relative to the pixels. That shear isn't being accounted for here,
# so the net result isn't equivalent to rotating by 90 degrees at the end.
t3 = time.time()
# Make another noise image
rev_image = galsim.Image(im_size, im_size, init_value=0, wcs=wcs)
rev_image.addNoise(
galsim.GaussianNoise(rng=rng, sigma=math.sqrt(noise_var)))
rev_ii = galsim.InterpolatedImage(rev_image, pad_factor=1)
# Find the effective transformation to apply in sky coordinates that matches what
# you would get by applying the shear in sky coords, going to image coords and then
# rotating by 90 degrees.
#
# If J is the jacobian of the wcs, and S1 is the applied shear, then we want to find
# S2 such that J^-1 S2 = R90 J^-1 S1
jac = wcs.jacobian()
J = jac.getMatrix()
Jinv = jac.inverse().getMatrix()
S1 = shear.getMatrix()
R90 = np.array([[0, -1], [1, 0]])
S2 = J.dot(R90).dot(Jinv).dot(S1)
scale, rev_shear, rev_theta, flip = galsim.JacobianWCS(
*S2.flatten()).getDecomposition()
# Flip should be False, and scale should be essentially 1.0.
assert flip == False
assert abs(scale - 1.) < 1.e-8
rev_sheared_obj = galsim.Convolve(
rev_ii, psf_inv).rotate(rev_theta).shear(rev_shear)
rev_final_obj = galsim.Convolve(psf_target, rev_sheared_obj)
rev_final_image = rev_final_obj.drawImage(obs_image.copy(),
method='no_pixel')
# Add the reverse-sheared noise image to the original image.
final_image2 = final_image + rev_final_image
t4 = time.time()
# The noise variance in the end should be 2x as large as the original
final_var = np.var(final_image2.array)
print('Alternate reverse shear method: final_var = ', final_var)
check_symm_noise(final_image2,
'using alternate reverse shear does not work')
print('Time for alternate reverse shear method = ', t4 - t3)
def __init__(self, image, x_interpolant = None, k_interpolant = None, normalization = 'flux',
dx = None, flux = None, pad_factor = 0., noise_pad = 0., rng = None,
pad_image = None, calculate_stepk=True, calculate_maxk=True,
use_cache=True, use_true_center=True, gsparams=None):
import numpy as np
# first try to read the image as a file. If it's not either a string or a valid
# pyfits hdu or hdulist, then an exception will be raised, which we ignore and move on.
try:
image = galsim.fits.read(image)
except:
pass
# make sure image is really an image and has a float type
if not isinstance(image, galsim.BaseImageF) and not isinstance(image, galsim.BaseImageD):
raise ValueError("Supplied image is not an image of floats or doubles!")
# it must have well-defined bounds, otherwise seg fault in SBInterpolatedImage constructor
if not image.getBounds().isDefined():
raise ValueError("Supplied image does not have bounds defined!")
# check what normalization was specified for the image: is it an image of surface
# brightness, or flux?
if not normalization.lower() in ("flux", "f", "surface brightness", "sb"):
raise ValueError(("Invalid normalization requested: '%s'. Expecting one of 'flux', "+
"'f', 'surface brightness', or 'sb'.") % normalization)
# set up the interpolants if none was provided by user, or check that the user-provided ones
# are of a valid type
if x_interpolant is None:
self.x_interpolant = galsim.InterpolantXY(galsim.Quintic(tol=1e-4))
else:
self.x_interpolant = galsim.utilities.convert_interpolant_to_2d(x_interpolant)
if k_interpolant is None:
self.k_interpolant = galsim.InterpolantXY(galsim.Quintic(tol=1e-4))
else:
self.k_interpolant = galsim.utilities.convert_interpolant_to_2d(k_interpolant)
# Check for input dx, and check whether Image already has one set. At the end of this
# code block, either an exception will have been raised, or the input image will have a
# valid scale set.
if dx is None:
dx = image.scale
if dx == 0:
raise ValueError("No information given with Image or keywords about pixel scale!")
else:
if type(dx) != float:
dx = float(dx)
# Don't change the original image. Make a new view if we need to set the scale.
image = image.view()
image.setScale(dx)
if dx == 0.0:
raise ValueError("dx may not be 0.0")
# Set up the GaussianDeviate if not provided one, or check that the user-provided one is
# of a valid type.
if rng is None:
gaussian_deviate = galsim.GaussianDeviate()
elif isinstance(rng, galsim.BaseDeviate):
# Even if it's already a GaussianDeviate, we still want to make a new Gaussian deviate
# that would generate the same sequence, because later we change the sigma and we don't
# want to change it for the original one that was passed in. So don't distinguish
# between GaussianDeviate and the other BaseDeviates here.
gaussian_deviate = galsim.GaussianDeviate(rng)
else:
raise TypeError("rng provided to InterpolatedImage constructor is not a BaseDeviate")
# decide about deterministic image padding
specify_size = False
padded_size = image.getPaddedSize(pad_factor)
if pad_image:
specify_size = True
if isinstance(pad_image, str):
pad_image = galsim.fits.read(pad_image)
if ( not isinstance(pad_image, galsim.BaseImageF) and
not isinstance(pad_image, galsim.BaseImageD) ):
raise ValueError("Supplied pad_image is not one of the allowed types!")
# If an image was supplied directly or from a file, check its size:
# Cannot use if too small.
# Use to define the final image size otherwise.
deltax = (1+pad_image.getXMax()-pad_image.getXMin())-(1+image.getXMax()-image.getXMin())
deltay = (1+pad_image.getYMax()-pad_image.getYMin())-(1+image.getYMax()-image.getYMin())
if deltax < 0 or deltay < 0:
raise RuntimeError("Image supplied for padding is too small!")
if pad_factor != 1. and pad_factor != 0.:
import warnings
msg = "Warning: ignoring specified pad_factor because user also specified\n"
msg += " an image to use directly for the padding."
warnings.warn(msg)
elif noise_pad:
if isinstance(image, galsim.BaseImageF):
pad_image = galsim.ImageF(padded_size, padded_size)
if isinstance(image, galsim.BaseImageD):
pad_image = galsim.ImageD(padded_size, padded_size)
# now decide about noise padding
# First, see if the input is consistent with a float.
# i.e. it could be an int, or a str that converts to a number.
try:
noise_pad = float(noise_pad)
except:
pass
if isinstance(noise_pad, float):
if noise_pad < 0.:
raise ValueError("Noise variance cannot be negative!")
elif noise_pad > 0.:
# Note: make sure the sigma is properly set to sqrt(noise_pad).
gaussian_deviate.setSigma(np.sqrt(noise_pad))
pad_image.addNoise(galsim.DeviateNoise(gaussian_deviate))
else:
if isinstance(noise_pad, galsim.correlatednoise._BaseCorrelatedNoise):
cn = noise_pad.copy()
if rng: # Let a user supplied RNG take precedence over that in user CN
cn.setRNG(gaussian_deviate)
elif isinstance(noise_pad,galsim.BaseImageF) or isinstance(noise_pad,galsim.BaseImageD):
cn = galsim.CorrelatedNoise(gaussian_deviate, noise_pad)
elif use_cache and noise_pad in InterpolatedImage._cache_noise_pad:
cn = InterpolatedImage._cache_noise_pad[noise_pad]
if rng:
# Make sure that we are using a specified RNG by resetting that in this cached
# CorrelatedNoise instance, otherwise preserve the cached RNG
cn.setRNG(gaussian_deviate)
elif isinstance(noise_pad, str):
cn = galsim.CorrelatedNoise(gaussian_deviate, galsim.fits.read(noise_pad))
if use_cache:
InterpolatedImage._cache_noise_pad[noise_pad] = cn
else:
raise ValueError(
"Input noise_pad must be a float/int, a CorrelatedNoise, Image, or filename "+
"containing an image to use to make a CorrelatedNoise!")
pad_image.addNoise(cn)
# Now we have to check: was the padding determined using pad_factor? Or by passing in an
# image for padding? Treat these cases differently:
# (1) If the former, then we can simply have the C++ handle the padding process.
# (2) If the latter, then we have to do the padding ourselves, and pass the resulting image
# to the C++ with pad_factor explicitly set to 1.
if specify_size is False:
# Make the SBInterpolatedImage out of the image.
sbinterpolatedimage = galsim.SBInterpolatedImage(
image, xInterp=self.x_interpolant, kInterp=self.k_interpolant,
dx=dx, pad_factor=pad_factor, pad_image=pad_image, gsparams=gsparams)
self.x_size = padded_size
self.y_size = padded_size
else:
# Leave the original image as-is. Instead, we shift around the image to be used for
# padding. Find out how much x and y margin there should be on lower end:
x_marg = int(np.round(0.5*deltax))
y_marg = int(np.round(0.5*deltay))
# Now reset the pad_image to contain the original image in an even way
pad_image = pad_image.view()
pad_image.setScale(dx)
pad_image.setOrigin(image.getXMin()-x_marg, image.getYMin()-y_marg)
# Set the central values of pad_image to be equal to the input image
pad_image[image.bounds] = image
sbinterpolatedimage = galsim.SBInterpolatedImage(
pad_image, xInterp=self.x_interpolant, kInterp=self.k_interpolant,
dx=dx, pad_factor=1., gsparams=gsparams)
self.x_size = 1+pad_image.getXMax()-pad_image.getXMin()
self.y_size = 1+pad_image.getYMax()-pad_image.getYMin()
# GalSim cannot automatically know what stepK and maxK are appropriate for the
# input image. So it is usually worth it to do a manual calculation here.
if calculate_stepk:
sbinterpolatedimage.calculateStepK()
if calculate_maxk:
sbinterpolatedimage.calculateMaxK()
# If the user specified a flux, then set to that flux value.
if flux != None:
if type(flux) != flux:
flux = float(flux)
sbinterpolatedimage.setFlux(flux)
# If the user specified a flux normalization for the input Image, then since
# SBInterpolatedImage works in terms of surface brightness, have to rescale the values to
# get proper normalization.
elif flux is None and normalization.lower() in ['flux','f'] and dx != 1.:
sbinterpolatedimage.scaleFlux(1./(dx**2))
# If the input Image normalization is 'sb' then since that is the SBInterpolated default
# assumption, no rescaling is needed.
# Initialize the SBProfile
GSObject.__init__(self, sbinterpolatedimage)
# Fix the center to be in the right place.
# Note the minus sign in front of image.scale, since we want to fix the center in the
# opposite sense of what the draw function does.
if use_true_center:
prof = self._fix_center(image, -image.scale)
GSObject.__init__(self, prof.SBProfile)