def fit_island_iteratively(self, img, isl, iter_ngmax=5, opts=None):
"""Fits an island iteratively.
For large islands, which can require many Gaussians to fit well,
it is much faster to fit a small number of Gaussians simultaneously
and iterate. However, this does usually result in larger residuals.
"""
import functions as func
sgaul = []; sfgaul = []
gaul = []; fgaul = []
if opts == None:
opts = img.opts
thresh_isl = opts.thresh_isl
thresh_pix = opts.thresh_pix
thresh = opts.fittedimage_clip
thr = isl.mean + thresh_isl * isl.rms
rms = isl.rms
if opts.verbose_fitting:
print 'Iteratively fitting island ', isl.island_id
gaul = []; fgaul = []
ffimg_tot = N.zeros(isl.shape, dtype=N.float32)
peak_val = N.max(isl.image - isl.islmean)
while peak_val >= thr:
sgaul, sfgaul = self.fit_island(isl, opts, img, ffimg=ffimg_tot, ngmax=iter_ngmax, ini_gausfit='simple')
gaul = gaul + sgaul; fgaul = fgaul + sfgaul
# Calculate residual image
if len(sgaul) > 0:
for g in sgaul:
gcopy = g[:]
gcopy[1] -= isl.origin[0]
gcopy[2] -= isl.origin[1]
S1, S2, Th = func.corrected_size(gcopy[3:6])
gcopy[3] = S1
gcopy[4] = S2
gcopy[5] = Th
A, C1, C2, S1, S2, Th = gcopy
shape = isl.shape
b = find_bbox(thresh*isl.rms, gcopy)
bbox = N.s_[max(0, int(C1-b)):min(shape[0], int(C1+b+1)),
max(0, int(C2-b)):min(shape[1], int(C2+b+1))]
x_ax, y_ax = N.mgrid[bbox]
ffimg = func.gaussian_fcn(gcopy, x_ax, y_ax)
ffimg_tot[bbox] += ffimg
peak_val_prev = peak_val
peak_val = N.max(isl.image - isl.islmean - ffimg_tot)
if func.approx_equal(peak_val, peak_val_prev):
break
else:
break
if len(gaul) == 0:
# Fitting iteratively did not work -- try normal fit
gaul, fgaul = self.fit_island(isl, opts, img, ini_gausfit='default')
return gaul, fgaul
def __init__(self, img, gaussian, isl_idx, g_idx, flag=0):
"""Initialize Gaussian object from fitting data
Parameters:
img: PyBDSM image object
gaussian: 6-tuple of fitted numbers
isl_idx: island serial number
g_idx: gaussian serial number
flag: flagging (if any)
"""
import functions as func
from const import fwsig
import numpy as N
use_wcs = True
self.gaussian_idx = g_idx
self.gaus_num = 0 # stored later
self.island_id = isl_idx
self.jlevel = img.j
self.flag = flag
self.parameters = gaussian
p = gaussian
self.peak_flux = p[0]
self.centre_pix = p[1:3]
size = p[3:6]
if func.approx_equal(size[0], img.pixel_beam()[0]*1.1) and \
func.approx_equal(size[1], img.pixel_beam()[1]) and \
func.approx_equal(size[2], img.pixel_beam()[2]+90.0) or \
img.opts.fix_to_beam:
# Check whether fitted Gaussian is just the distorted pixel beam
# given as an initial guess or if size was fixed to the beam. If so,
# reset the size to the undistorted beam.
# Note: these are sigma sizes, not FWHM sizes.
size = img.pixel_beam()
size = (size[0], size[1], size[2]+90.0) # adjust angle so that corrected_size() works correctly
size = func.corrected_size(size) # gives fwhm and P.A.
self.size_pix = size # FWHM in pixels and P.A. CCW from +y axis
# Use img.orig_beam for flux calculation and deconvolution on wavelet
# images, as img.beam has been altered to match the wavelet scale.
# Note: these are all FWHM sizes.
if img.waveletimage:
bm_pix = N.array(img.beam2pix(img.orig_beam))
else:
bm_pix = N.array(img.beam2pix(img.beam))
# Calculate fluxes, sky sizes, etc. All sizes are FWHM.
tot = p[0]*size[0]*size[1]/(bm_pix[0]*bm_pix[1])
if flag == 0:
# These are good Gaussians
errors = func.get_errors(img, p+[tot], img.islands[isl_idx].rms)
self.centre_sky = img.pix2sky(p[1:3])
self.centre_skyE = img.pix2coord(errors[1:3], self.centre_pix, use_wcs=use_wcs)
self.size_sky = img.pix2gaus(size, self.centre_pix, use_wcs=use_wcs) # FWHM in degrees and P.A. east from north
self.size_sky_uncorr = img.pix2gaus(size, self.centre_pix, use_wcs=False) # FWHM in degrees and P.A. east from +y axis
self.size_skyE = img.pix2gaus(errors[3:6], self.centre_pix, use_wcs=use_wcs)
self.size_skyE_uncorr = img.pix2gaus(errors[3:6], self.centre_pix, use_wcs=False)
gaus_dc, err = func.deconv2(bm_pix, size)
self.deconv_size_sky = img.pix2gaus(gaus_dc, self.centre_pix, use_wcs=use_wcs)
self.deconv_size_sky_uncorr = img.pix2gaus(gaus_dc, self.centre_pix, use_wcs=False)
self.deconv_size_skyE = img.pix2gaus(errors[3:6], self.centre_pix, use_wcs=use_wcs)
self.deconv_size_skyE_uncorr = img.pix2gaus(errors[3:6], self.centre_pix, use_wcs=False)
else:
# These are flagged Gaussians, so don't calculate sky values or errors
errors = [0]*7
self.centre_sky = [0., 0.]
self.centre_skyE = [0., 0.]
self.size_sky = [0., 0., 0.]
self.size_sky_uncorr = [0., 0., 0.]
self.size_skyE = [0., 0.]
self.size_skyE_uncorr = [0., 0., 0.]
self.deconv_size_sky = [0., 0., 0.]
self.deconv_size_sky_uncorr = [0., 0., 0.]
self.deconv_size_skyE = [0., 0., 0.]
self.deconv_size_skyE_uncorr = [0., 0., 0.]
self.total_flux = tot
self.total_fluxE = errors[6]
self.peak_fluxE = errors[0]
self.total_fluxE = errors[6]
self.centre_pixE = errors[1:3]
self.size_pixE = errors[3:6]
self.rms = img.islands[isl_idx].rms
self.mean = img.islands[isl_idx].mean
self.total_flux_isl = img.islands[isl_idx].total_flux
self.total_flux_islE = img.islands[isl_idx].total_fluxE
def __call__(self, img):
mylog = mylogger.logging.getLogger("PyBDSM."+img.log+"Preprocess")
bstat = func.bstat
if img.opts.kappa_clip is None:
kappa = -img.pixel_beamarea()
else:
kappa = img.opts.kappa_clip
if img.opts.polarisation_do:
pols = ['I', 'Q', 'U', 'V']
ch0images = [img.ch0_arr, img.ch0_Q_arr, img.ch0_U_arr, img.ch0_V_arr]
img.clipped_mean_QUV = []
img.clipped_rms_QUV = []
else:
pols = ['I'] # assume I is always present
ch0images = [img.ch0_arr]
if hasattr(img, 'rms_mask'):
mask = img.rms_mask
else:
mask = img.mask_arr
opts = img.opts
for ipol, pol in enumerate(pols):
image = ch0images[ipol]
### basic stats
mean, rms, cmean, crms, cnt = bstat(image, mask, kappa)
if cnt > 198: cmean = mean; crms = rms
if pol == 'I':
if func.approx_equal(crms, 0.0, rel=None):
raise RuntimeError('Clipped rms appears to be zero. Check for regions '\
'with values of 0 and\nblank them (with NaNs) '\
'or use trim_box to exclude them.')
img.raw_mean = mean
img.raw_rms = rms
img.clipped_mean= cmean
img.clipped_rms = crms
mylog.info('%s %.4f %s %.4f %s ' % ("Raw mean (Stokes I) = ", mean*1000.0, \
'mJy and raw rms = ',rms*1000.0, 'mJy'))
mylog.info('%s %.4f %s %s %.4f %s ' % ("sigma clipped mean (Stokes I) = ", cmean*1000.0, \
'mJy and ','sigma clipped rms = ',crms*1000.0, 'mJy'))
else:
img.clipped_mean_QUV.append(cmean)
img.clipped_rms_QUV.append(crms)
mylog.info('%s %s %s %.4f %s %s %.4f %s ' % ("sigma clipped mean (Stokes ", pol, ") = ", cmean*1000.0, \
'mJy and ','sigma clipped rms = ',crms*1000.0, 'mJy'))
image = img.ch0_arr
# Check if pixels are outside the universe
if opts.check_outsideuniv:
mylogger.userinfo(mylog, "Checking for pixels outside the universe")
noutside_univ = self.outside_univ(img)
img.noutside_univ = noutside_univ
frac_blank = round(float(noutside_univ)/float(image.shape[0]*image.shape[1]),3)
mylogger.userinfo(mylog, "Number of additional pixels blanked", str(noutside_univ)
+' ('+str(frac_blank*100.0)+'%)')
else:
noutside_univ = 0
# If needed, (re)mask the image
if noutside_univ > 0:
mask = N.isnan(img.ch0_arr)
masked = mask.any()
img.masked = masked
if masked:
img.mask_arr = mask
img.blankpix = N.sum(mask)
### max/min pixel value & coordinates
shape = image.shape[0:2]
if mask != None:
img.blankpix = N.sum(mask)
if img.blankpix == 0:
max_idx = image.argmax()
min_idx = image.argmin()
else:
max_idx = N.nanargmax(image)
min_idx = N.nanargmin(image)
img.maxpix_coord = N.unravel_index(max_idx, shape)
img.minpix_coord = N.unravel_index(min_idx, shape)
img.max_value = image.flat[max_idx]
img.min_value = image.flat[min_idx]
### Solid angle of the image
cdelt = N.array(img.wcs_obj.acdelt[:2])
img.omega = N.product(shape)*abs(N.product(cdelt))/(180.*180./pi/pi)
### Total flux in ch0 image
if 'atrous' in img.filename or img._pi or img.log == 'Detection image':
# Don't do this estimate for atrous wavelet images
# or polarized intensity image,
# as it doesn't give the correct flux. Also, ignore
# the flux in the detection image, as it's likely
# wrong (e.g., not corrected for the primary beam).
img.ch0_sum_jy = 0
else:
im_flux = N.nansum(image)/img.pixel_beamarea() # Jy
img.ch0_sum_jy = im_flux
mylogger.userinfo(mylog, 'Flux from sum of (non-blank) pixels',
'%.3f Jy' % (im_flux,))
### if image seems confused, then take background mean as zero instead
alpha_sourcecounts = 2.5 # approx diff src count slope. 2.2?
if opts.bmpersrc_th is None:
n = (image >= 5.*crms).sum()
if n <= 0:
n = 1
mylog.info('No pixels in image > 5-sigma.')
mylog.info('Taking number of pixels above 5-sigma as 1.')
img.bmpersrc_th = N.product(shape)/((alpha_sourcecounts-1.)*n)
mylog.info('%s %6.2f' % ('Estimated bmpersrc_th = ', img.bmpersrc_th))
else:
img.bmpersrc_th = opts.bmpersrc_th
mylog.info('%s %6.2f' % ('Taking default bmpersrc_th = ', img.bmpersrc_th))
confused = False
if opts.mean_map == 'default':
if opts.bmpersrc_th <= 25. or cmean/crms >= 0.1:
confused = True
img.confused = confused
mylog.info('Parameter confused is '+str(img.confused))
img.completed_Ops.append('preprocess')
return img
def __call__(self, img):
mylog = mylogger.logging.getLogger("PyBDSM." + img.log + "Preprocess")
bstat = func.bstat
if img.opts.kappa_clip is None:
kappa = -img.pixel_beamarea()
else:
kappa = img.opts.kappa_clip
if img.opts.polarisation_do:
pols = ['I', 'Q', 'U', 'V']
ch0images = [
img.ch0_arr, img.ch0_Q_arr, img.ch0_U_arr, img.ch0_V_arr
]
img.clipped_mean_QUV = []
img.clipped_rms_QUV = []
else:
pols = ['I'] # assume I is always present
ch0images = [img.ch0_arr]
if hasattr(img, 'rms_mask'):
mask = img.rms_mask
else:
mask = img.mask_arr
opts = img.opts
for ipol, pol in enumerate(pols):
image = ch0images[ipol]
### basic stats
mean, rms, cmean, crms, cnt = bstat(image, mask, kappa)
if cnt > 198:
cmean = mean
crms = rms
if pol == 'I':
if func.approx_equal(crms, 0.0, rel=None):
raise RuntimeError('Clipped rms appears to be zero. Check for regions '\
'with values of 0 and\nblank them (with NaNs) '\
'or use trim_box to exclude them.')
img.raw_mean = mean
img.raw_rms = rms
img.clipped_mean = cmean
img.clipped_rms = crms
mylog.info('%s %.4f %s %.4f %s ' % ("Raw mean (Stokes I) = ", mean*1000.0, \
'mJy and raw rms = ',rms*1000.0, 'mJy'))
mylog.info('%s %.4f %s %s %.4f %s ' % ("sigma clipped mean (Stokes I) = ", cmean*1000.0, \
'mJy and ','sigma clipped rms = ',crms*1000.0, 'mJy'))
else:
img.clipped_mean_QUV.append(cmean)
img.clipped_rms_QUV.append(crms)
mylog.info('%s %s %s %.4f %s %s %.4f %s ' % ("sigma clipped mean (Stokes ", pol, ") = ", cmean*1000.0, \
'mJy and ','sigma clipped rms = ',crms*1000.0, 'mJy'))
image = img.ch0_arr
# Check if pixels are outside the universe
if opts.check_outsideuniv:
mylogger.userinfo(mylog,
"Checking for pixels outside the universe")
noutside_univ = self.outside_univ(img)
img.noutside_univ = noutside_univ
frac_blank = round(
float(noutside_univ) / float(image.shape[0] * image.shape[1]),
3)
mylogger.userinfo(
mylog, "Number of additional pixels blanked",
str(noutside_univ) + ' (' + str(frac_blank * 100.0) + '%)')
else:
noutside_univ = 0
# If needed, (re)mask the image
if noutside_univ > 0:
mask = N.isnan(img.ch0_arr)
masked = mask.any()
img.masked = masked
if masked:
img.mask_arr = mask
img.blankpix = N.sum(mask)
### max/min pixel value & coordinates
shape = image.shape[0:2]
if mask != None:
img.blankpix = N.sum(mask)
if img.blankpix == 0:
max_idx = image.argmax()
min_idx = image.argmin()
else:
max_idx = N.nanargmax(image)
min_idx = N.nanargmin(image)
img.maxpix_coord = N.unravel_index(max_idx, shape)
img.minpix_coord = N.unravel_index(min_idx, shape)
img.max_value = image.flat[max_idx]
img.min_value = image.flat[min_idx]
### Solid angle of the image
cdelt = N.array(img.wcs_obj.acdelt[:2])
img.omega = N.product(shape) * abs(
N.product(cdelt)) / (180. * 180. / pi / pi)
### Total flux in ch0 image
if 'atrous' in img.filename or img._pi or img.log == 'Detection image':
# Don't do this estimate for atrous wavelet images
# or polarized intensity image,
# as it doesn't give the correct flux. Also, ignore
# the flux in the detection image, as it's likely
# wrong (e.g., not corrected for the primary beam).
img.ch0_sum_jy = 0
else:
im_flux = N.nansum(image) / img.pixel_beamarea() # Jy
img.ch0_sum_jy = im_flux
mylogger.userinfo(mylog, 'Flux from sum of (non-blank) pixels',
'%.3f Jy' % (im_flux, ))
### if image seems confused, then take background mean as zero instead
alpha_sourcecounts = 2.5 # approx diff src count slope. 2.2?
if opts.bmpersrc_th is None:
n = (image >= 5. * crms).sum()
if n <= 0:
n = 1
mylog.info('No pixels in image > 5-sigma.')
mylog.info('Taking number of pixels above 5-sigma as 1.')
img.bmpersrc_th = N.product(shape) / (
(alpha_sourcecounts - 1.) * n)
mylog.info('%s %6.2f' %
('Estimated bmpersrc_th = ', img.bmpersrc_th))
else:
img.bmpersrc_th = opts.bmpersrc_th
mylog.info('%s %6.2f' %
('Taking default bmpersrc_th = ', img.bmpersrc_th))
confused = False
if opts.mean_map == 'default':
if opts.bmpersrc_th <= 25. or cmean / crms >= 0.1:
confused = True
img.confused = confused
mylog.info('Parameter confused is ' + str(img.confused))
img.completed_Ops.append('preprocess')
return img
