def _meisner_psf_models():
global plotslice
# Meisner's PSF models
# for band in [4]:
for band in [1, 2, 3, 4]:
print()
print('W%i' % band)
print()
#pix = fitsio.read('wise-psf-avg-pix.fits', ext=band-1)
pix = fitsio.read('wise-psf-avg-pix-bright.fits', ext=band - 1)
fit = fits_table('wise-psf-avg.fits', hdu=band)
#fit = fits_table('psf-allwise-con3.fits', hdu=band)
scale = 1.
print('Pix shape', pix.shape)
h, w = pix.shape
xx, yy = np.meshgrid(np.arange(w), np.arange(h))
cx, cy = np.sum(xx * pix) / np.sum(pix), np.sum(yy * pix) / np.sum(pix)
print('Centroid:', cx, cy)
#S = 100
S = h / 2
slc = slice(h / 2 - S, h / 2 + S + 1), slice(w / 2 - S, w / 2 + S + 1)
plotss = 30
plotslice = slice(S - plotss,
-(S - plotss)), slice(S - plotss, -(S - plotss))
opix = pix
print('Original pixels sum:', opix.sum())
pix /= pix.sum()
pix = pix[slc]
print('Sliced pix sum', pix.sum())
psf = GaussianMixturePSF(fit.amp, fit.mean * scale, fit.var * scale**2)
psfmodel = psf.getPointSourcePatch(0., 0., radius=h / 2)
mod = psfmodel.patch
print('Orig mod sum', mod.sum())
mod = mod[slc]
print('Sliced mod sum', mod.sum())
print('Amps:', np.sum(psf.mog.amp))
_plot_psf(pix, mod, psf)
plt.suptitle('W%i: Orig' % band)
ps.savefig()
# Lanczos sub-sample
if band == 4:
lpix = _lanczos_subsample(opix, 2)
pix = lpix
h, w = pix.shape
#S = 140
slc = slice(h / 2 - S,
h / 2 + S + 1), slice(w / 2 - S, w / 2 + S + 1)
print('Resampled pix sum', pix.sum())
pix = pix[slc]
print('sliced pix sum', pix.sum())
psf = GaussianMixturePSF(fit.amp, fit.mean * scale,
fit.var * scale**2)
psfmodel = psf.getPointSourcePatch(0., 0., radius=h / 2)
mod = psfmodel.patch
print('Scaled mod sum', mod.sum())
mod = mod[slc]
print('Sliced mod sum', mod.sum())
_plot_psf(pix, mod, psf)
plt.suptitle('W%i: Scaled' % band)
ps.savefig()
plotslice = slice(S - plotss,
-(S - plotss)), slice(S - plotss, -(S - plotss))
psfx = GaussianMixturePSF.fromStamp(pix,
P0=(fit.amp, fit.mean * scale,
fit.var * scale**2))
psfmodel = psfx.getPointSourcePatch(0., 0., radius=h / 2)
mod = psfmodel.patch
print('From stamp: mod sum', mod.sum())
mod = mod[slc]
print('Sliced mod sum:', mod.sum())
_plot_psf(pix, mod, psfx)
plt.suptitle('W%i: fromStamp: %g = %s, res %.3f' %
(band, np.sum(psfx.mog.amp), ','.join(
['%.3f' % a
for a in psfx.mog.amp]), np.sum(pix - mod)))
ps.savefig()
print('Stamp-Fit PSF params:', psfx)
# class MyGaussianMixturePSF(GaussianMixturePSF):
# def getLogPrior(self):
# if np.any(self.mog.amp < 0.):
# return -np.inf
# for k in range(self.mog.K):
# if np.linalg.det(self.mog.var[k]) <= 0:
# return -np.inf
# return 0
#
# @property
# def amp(self):
# return self.mog.amp
#
# mypsf = MyGaussianMixturePSF(psfx.mog.amp, psfx.mog.mean, psfx.mog.var)
# mypsf.radius = sh/2
sh, sw = pix.shape
# Initialize from original fit params
psfx = psf
# Try concentric gaussian PSF
sigmas = []
for k in range(psfx.mog.K):
v = psfx.mog.var[k, :, :]
sigmas.append(np.sqrt(np.sqrt(np.abs(v[0, 0] * v[1, 1]))))
print('Initializing concentric Gaussian PSF with sigmas', sigmas)
gpsf = NCircularGaussianPSF(sigmas, psfx.mog.amp)
gpsf.radius = sh / 2
mypsf = gpsf
tim = Image(data=pix, invvar=1e6 * np.ones_like(pix), psf=mypsf)
tim.modelMinval = 1e-16
# xx,yy = np.meshgrid(np.arange(sw), np.arange(sh))
# cx,cy = np.sum(xx*pix)/np.sum(pix), np.sum(yy*pix)/np.sum(pix)
# print 'Centroid:', cx,cy
# print 'Pix midpoint:', sw/2, sh/2
cx, cy = sw / 2, sh / 2
src = PointSource(PixPos(cx, cy), Flux(1.0))
tractor = Tractor([tim], [src])
tim.freezeAllBut('psf')
# tractor.freezeParam('catalog')
src.freezeAllBut('pos')
# src.freezeAllBut('brightness')
# tractor.freezeParam('images')
# tractor.optimize_forced_photometry()
# tractor.thawParam('images')
# print 'Source flux after forced-photom fit:', src.getBrightness()
print('Optimizing Params:')
tractor.printThawedParams()
for i in range(200):
#dlnp,X,alpha = tractor.optimize(damp=0.1)
dlnp, X, alpha = tractor.optimize(damp=1)
print(i, 'dlnp %.3g' % dlnp, 'psf', gpsf)
if dlnp < 1e-6:
break
tractor.freezeParam('catalog')
gpsf.sigmas.stepsize = len(gpsf.sigmas) * [1e-6]
gpsf.weights.stepsize = len(gpsf.sigmas) * [1e-6]
for i in range(200):
#dlnp,X,alpha = tractor.optimize(damp=0.1)
dlnp, X, alpha = tractor.optimize(damp=1)
print(i, 'dlnp %.3g' % dlnp, 'psf', gpsf)
if dlnp < 1e-6:
break
print('PSF3(opt): flux', src.brightness)
print('PSF amps:', np.sum(mypsf.amp))
print('PSF amps * Source brightness:',
src.brightness.getValue() * np.sum(mypsf.amp))
print('pix sum:', pix.sum())
print('Optimize source:', src)
print('Optimized PSF:', mypsf)
mod = tractor.getModelImage(0)
print('Mod sum:', mod.sum())
_plot_psf(pix, mod, mypsf, flux=src.brightness.getValue())
plt.suptitle('W%i psf3 (opt): %g = %s, resid %.3f' %
(band, np.sum(mypsf.amp), ','.join(
['%.3f' % a for a in mypsf.amp]), np.sum(pix - mod)))
ps.savefig()
# Write before normalizing!
T = fits_table()
T.amp = mypsf.mog.amp
T.mean = mypsf.mog.mean
T.var = mypsf.mog.var
T.writeto('psf3-w%i.fits' % band)
T.writeto('psf3.fits', append=(band != 1))
mypsf.weights.setParams(
np.array(mypsf.weights.getParams()) /
sum(mypsf.weights.getParams()))
print('Normalized PSF weights:', mypsf)
mod = tractor.getModelImage(0)
print('Mod sum:', mod.sum())
_plot_psf(pix, mod, mypsf, flux=src.brightness.getValue())
plt.suptitle('W%i psf3 (opt): %g = %s, resid %.3f' %
(band, np.sum(mypsf.amp), ','.join(
['%.3f' % a for a in mypsf.amp]), np.sum(pix - mod)))
ps.savefig()
class MyGaussianPSF(NCircularGaussianPSF):
'''
A PSF model with strictly positive weights that sum to
unity.
'''
def __init__(self, sigmas, weights):
ww = np.array(weights)
ww = np.log(ww)
super(MyGaussianPSF, self).__init__(sigmas, ww)
@staticmethod
def getNamedParams():
return dict(sigmas=0, logweights=1)
def __str__(self):
return ('MyGaussianPSF: sigmas [ ' +
', '.join(['%.3f' % s for s in self.mysigmas]) +
' ], weights [ ' +
', '.join(['%.3f' % w for w in self.myweights]) + ' ]')
@property
def myweights(self):
ww = np.exp(self.logweights.getAllParams())
ww /= ww.sum()
return ww
@property
def weights(self):
ww = np.exp(self.logweights.getParams())
wsum = np.sum(np.exp(self.logweights.getAllParams()))
return ww / wsum
if band == 4:
# HACK
mypsf = MyGaussianPSF([1.7, 6.4, 15.0], [0.333, 0.666, 0.1])
else:
mypsf = MyGaussianPSF(gpsf.sigmas, gpsf.amp)
mypsf.radius = sh / 2
tim.psf = mypsf
print('Optimizing Params:')
tractor.printThawedParams()
for i in range(200):
#dlnp,X,alpha = tractor.optimize(damp=0.1)
dlnp, X, alpha = tractor.optimize(damp=1)
print(i, 'dlnp %.3g' % dlnp, 'psf', mypsf)
if dlnp < 1e-6:
break
mypsf.sigmas.stepsize = len(mypsf.sigmas) * [1e-6]
mypsf.logweights.stepsize = len(mypsf.sigmas) * [1e-6]
for i in range(200):
#dlnp,X,alpha = tractor.optimize(damp=0.1)
dlnp, X, alpha = tractor.optimize(damp=1)
print(i, 'dlnp %.3g' % dlnp, 'psf', mypsf)
if dlnp < 1e-6:
break
print('PSF amps:', np.sum(mypsf.amp))
print('pix sum:', pix.sum())
print('Optimize source:', src)
print('Optimized PSF:', mypsf)
mod = tractor.getModelImage(0)
print('Mod sum:', mod.sum())
_plot_psf(pix, mod, mypsf, flux=src.brightness.getValue())
plt.suptitle('W%i psf3 (opt2): %g = %s, resid %.3f' %
(band, np.sum(mypsf.amp), ','.join(
['%.3f' % a for a in mypsf.amp]), np.sum(pix - mod)))
ps.savefig()
# Write
mog = mypsf.getMixtureOfGaussians()
T = fits_table()
T.amp = mog.amp
T.mean = mog.mean
T.var = mog.var
T.writeto('psf4.fits', append=(band != 1))
def _allwise_psf_models():
global plotslice
# AllWISE PSF models
for band in [1, 2, 3, 4]:
print()
print('W%i' % band)
print()
pix = reduce(np.add, [
fitsio.read('wise-w%i-psf-wpro-09x09-%02ix%02i.fits' %
(band, 1 + (i / 9), 1 + (i % 9))) for i in range(9 * 9)
])
print('Pix', pix.shape)
h, w = pix.shape
print('Pix sum', pix.sum())
print('Pix max', pix.max())
pix /= pix.sum()
psf = GaussianMixturePSF.fromStamp(pix)
print('Fit PSF', psf)
psfmodel = psf.getPointSourcePatch(0., 0., radius=h / 2)
mod = psfmodel.patch
print('Orig mod sum', mod.sum())
S = h / 2
plotss = 30
plotslice = slice(S - plotss,
-(S - plotss)), slice(S - plotss, -(S - plotss))
_plot_psf(pix, mod, psf)
plt.suptitle('W%i: from stamp' % band)
ps.savefig()
# Try concentric gaussian PSF
sh, sw = pix.shape
sigmas = []
for k in range(psf.mog.K):
v = psf.mog.var[k, :, :]
sigmas.append(np.sqrt(np.sqrt(v[0, 0] * v[1, 1])))
print('Initializing concentric Gaussian model with sigmas', sigmas)
print('And weights', psf.mog.amp)
# (Initial) W1:
# sig [7.1729391870564569, 24.098952864976805, 108.78869923786333]
# weights [ 0.80355109 0.15602835 0.04195982]
# sigmas [ 6.928, 17.091, 45.745 ], weights [ 0.760, 0.154, 0.073 ]
# W2:
# sig [8.2356371659198189, 25.741694812001221, 110.17431488810806]
# weights [ 0.80636191 0.1563742 0.03926182]
# sigmas [ 7.177, 10.636, 30.247 ], weights [ 0.500, 0.341, 0.134 ]
# W3:
# sig [6.860124763919889, 19.160849966251824, 134.20812055907825]
# weights [ 0.28227047 0.55080156 0.16706357]
# sigmas [ 6.850, 18.922, 109.121 ], weights [ 0.276, 0.554, 0.148 ]
# W4:
# [4.6423676603462001, 5.31542132669962, 18.375512539373585]
# weights [ 0.10764341 0.26915295 0.62320374]
# (Opt)
# sigmas [ 6.293, 6.314, 20.932 ], weights [ 0.129, 0.309, 0.598 ]
weights = psf.mog.amp
# HACK
if band == 4:
sigmas = [2., 6., 25.]
weights = [0.5, 0.25, 0.25]
else:
sigmas = [8., 24., 100.]
weights = [0.5, 0.25, 0.25]
gpsf = NCircularGaussianPSF(sigmas, weights)
gpsf.radius = sh / 2
mypsf = gpsf
tim = Image(data=pix, invvar=1e6 * np.ones_like(pix), psf=mypsf)
cx, cy = sw / 2, sh / 2
src = PointSource(PixPos(cx, cy), Flux(1.0))
tractor = Tractor([tim], [src])
tim.freezeAllBut('psf')
# tractor.freezeParam('catalog')
src.freezeAllBut('pos')
print('Optimizing Params:')
tractor.printThawedParams()
for i in range(100):
dlnp, X, alpha = tractor.optimize(damp=0.1)
print('dlnp', dlnp)
print('PSF', mypsf)
if dlnp < 0.001:
break
print('PSF3(opt): flux', src.brightness)
print('PSF amps:', np.sum(mypsf.amp))
print('PSF amps * Source brightness:',
src.brightness.getValue() * np.sum(mypsf.amp))
print('pix sum:', pix.sum())
print('Optimize source:', src)
print('Optimized PSF:', mypsf)
mod = tractor.getModelImage(0)
print('Mod sum:', mod.sum())
print('Mod min:', mod.min())
print('Mod max:', mod.max())
_plot_psf(pix, mod, mypsf, flux=src.brightness.getValue())
plt.suptitle('W%i psf3 (opt): %g = %s, resid %.3f' %
(band, np.sum(mypsf.amp), ','.join(
['%.3f' % a for a in mypsf.amp]), np.sum(pix - mod)))
ps.savefig()
# These postage stamps are subsampled at 8x the 2.75"/pix,
# except for the W4 one, which is at 4x that.
scale = 8
if band == 4:
scale = 4
T = fits_table()
T.amp = mypsf.mog.amp
T.mean = mypsf.mog.mean / scale
T.var = mypsf.mog.var / scale**2
T.writeto('psf-allwise-con3-w%i.fits' % band)
T.writeto('psf-allwise-con3.fits', append=(band != 1))
mypsf.weights.setParams(
np.array(mypsf.weights.getParams()) /
sum(mypsf.weights.getParams()))
print('Normalized PSF weights:', mypsf)
mod = tractor.getModelImage(0)
print('Mod sum:', mod.sum())
_plot_psf(pix, mod, mypsf, flux=src.brightness.getValue())
plt.suptitle('W%i psf3 (opt): %g = %s, resid %.3f' %
(band, np.sum(mypsf.amp), ','.join(
['%.3f' % a for a in mypsf.amp]), np.sum(pix - mod)))
ps.savefig()
psf2 = GaussianMixturePSF(mypsf.mog.amp[:2] / mypsf.mog.amp[:2].sum(),
mypsf.mog.mean[:2, :], mypsf.mog.var[:2, :, :])
psf2.radius = sh / 2
tim.psf = psf2
mod = tractor.getModelImage(0)
_plot_psf(lpix, mod, psf2, flux=src.brightness.getValue())
plt.suptitle('psf2 (init)')
ps.savefig()
src.freezeAllBut('brightness')
tractor.freezeParam('catalog')
for i in range(100):
print('Optimizing PSF:')
tractor.printThawedParams()
dlnp, X, alpha = tractor.optimize(damp=1.)
tractor.freezeParam('images')
tractor.thawParam('catalog')
print('Optimizing flux:')
tractor.printThawedParams()
tractor.optimize_forced_photometry()
tractor.thawParam('images')
tractor.freezeParam('catalog')
mod = tractor.getModelImage(0)
_plot_psf(lpix, mod, psf2, flux=src.brightness.getValue())
def _meisner_psf_models():
global plotslice
# Meisner's PSF models
#for band in [4]:
for band in [1,2,3,4]:
print
print 'W%i' % band
print
#pix = fitsio.read('wise-psf-avg-pix.fits', ext=band-1)
pix = fitsio.read('wise-psf-avg-pix-bright.fits', ext=band-1)
fit = fits_table('wise-psf-avg.fits', hdu=band)
#fit = fits_table('psf-allwise-con3.fits', hdu=band)
scale = 1.
print 'Pix shape', pix.shape
h,w = pix.shape
xx,yy = np.meshgrid(np.arange(w), np.arange(h))
cx,cy = np.sum(xx*pix)/np.sum(pix), np.sum(yy*pix)/np.sum(pix)
print 'Centroid:', cx,cy
#S = 100
S = h/2
slc = slice(h/2-S, h/2+S+1), slice(w/2-S, w/2+S+1)
plotss = 30
plotslice = slice(S-plotss, -(S-plotss)), slice(S-plotss, -(S-plotss))
opix = pix
print 'Original pixels sum:', opix.sum()
pix /= pix.sum()
pix = pix[slc]
print 'Sliced pix sum', pix.sum()
psf = GaussianMixturePSF(fit.amp, fit.mean * scale, fit.var * scale**2)
psfmodel = psf.getPointSourcePatch(0., 0., radius=h/2)
mod = psfmodel.patch
print 'Orig mod sum', mod.sum()
mod = mod[slc]
print 'Sliced mod sum', mod.sum()
print 'Amps:', np.sum(psf.mog.amp)
_plot_psf(pix, mod, psf)
plt.suptitle('W%i: Orig' % band)
ps.savefig()
# Lanczos sub-sample
if band == 4:
lpix = _lanczos_subsample(opix, 2)
pix = lpix
h,w = pix.shape
#S = 140
slc = slice(h/2-S, h/2+S+1), slice(w/2-S, w/2+S+1)
print 'Resampled pix sum', pix.sum()
pix = pix[slc]
print 'sliced pix sum', pix.sum()
psf = GaussianMixturePSF(fit.amp, fit.mean * scale, fit.var * scale**2)
psfmodel = psf.getPointSourcePatch(0., 0., radius=h/2)
mod = psfmodel.patch
print 'Scaled mod sum', mod.sum()
mod = mod[slc]
print 'Sliced mod sum', mod.sum()
_plot_psf(pix, mod, psf)
plt.suptitle('W%i: Scaled' % band)
ps.savefig()
plotslice = slice(S-plotss,-(S-plotss)),slice(S-plotss,-(S-plotss))
psfx = GaussianMixturePSF.fromStamp(pix, P0=(fit.amp, fit.mean*scale, fit.var*scale**2))
psfmodel = psfx.getPointSourcePatch(0., 0., radius=h/2)
mod = psfmodel.patch
print 'From stamp: mod sum', mod.sum()
mod = mod[slc]
print 'Sliced mod sum:', mod.sum()
_plot_psf(pix, mod, psfx)
plt.suptitle('W%i: fromStamp: %g = %s, res %.3f' %
(band, np.sum(psfx.mog.amp), ','.join(['%.3f'%a for a in psfx.mog.amp]),
np.sum(pix - mod)))
ps.savefig()
print 'Stamp-Fit PSF params:', psfx
# class MyGaussianMixturePSF(GaussianMixturePSF):
# def getLogPrior(self):
# if np.any(self.mog.amp < 0.):
# return -np.inf
# for k in range(self.mog.K):
# if np.linalg.det(self.mog.var[k]) <= 0:
# return -np.inf
# return 0
#
# @property
# def amp(self):
# return self.mog.amp
#
# mypsf = MyGaussianMixturePSF(psfx.mog.amp, psfx.mog.mean, psfx.mog.var)
# mypsf.radius = sh/2
sh,sw = pix.shape
# Initialize from original fit params
psfx = psf
# Try concentric gaussian PSF
sigmas = []
for k in range(psfx.mog.K):
v = psfx.mog.var[k,:,:]
sigmas.append(np.sqrt(np.sqrt(np.abs(v[0,0] * v[1,1]))))
print 'Initializing concentric Gaussian PSF with sigmas', sigmas
gpsf = NCircularGaussianPSF(sigmas, psfx.mog.amp)
gpsf.radius = sh/2
mypsf = gpsf
tim = Image(data=pix, invvar=1e6 * np.ones_like(pix), psf=mypsf)
tim.modelMinval = 1e-16
# xx,yy = np.meshgrid(np.arange(sw), np.arange(sh))
# cx,cy = np.sum(xx*pix)/np.sum(pix), np.sum(yy*pix)/np.sum(pix)
# print 'Centroid:', cx,cy
# print 'Pix midpoint:', sw/2, sh/2
cx,cy = sw/2, sh/2
src = PointSource(PixPos(cx, cy), Flux(1.0))
tractor = Tractor([tim], [src])
tim.freezeAllBut('psf')
#tractor.freezeParam('catalog')
src.freezeAllBut('pos')
# src.freezeAllBut('brightness')
# tractor.freezeParam('images')
# tractor.optimize_forced_photometry()
# tractor.thawParam('images')
# print 'Source flux after forced-photom fit:', src.getBrightness()
print 'Optimizing Params:'
tractor.printThawedParams()
for i in range(200):
#dlnp,X,alpha = tractor.optimize(damp=0.1)
dlnp,X,alpha = tractor.optimize(damp=1)
print i,'dlnp %.3g' % dlnp, 'psf', gpsf
if dlnp < 1e-6:
break
tractor.freezeParam('catalog')
gpsf.sigmas.stepsize = len(gpsf.sigmas) * [1e-6]
gpsf.weights.stepsize = len(gpsf.sigmas) * [1e-6]
for i in range(200):
#dlnp,X,alpha = tractor.optimize(damp=0.1)
dlnp,X,alpha = tractor.optimize(damp=1)
print i,'dlnp %.3g' % dlnp, 'psf', gpsf
if dlnp < 1e-6:
break
print 'PSF3(opt): flux', src.brightness
print 'PSF amps:', np.sum(mypsf.amp)
print 'PSF amps * Source brightness:', src.brightness.getValue() * np.sum(mypsf.amp)
print 'pix sum:', pix.sum()
print 'Optimize source:', src
print 'Optimized PSF:', mypsf
mod = tractor.getModelImage(0)
print 'Mod sum:', mod.sum()
_plot_psf(pix, mod, mypsf, flux=src.brightness.getValue())
plt.suptitle('W%i psf3 (opt): %g = %s, resid %.3f' %
(band, np.sum(mypsf.amp), ','.join(['%.3f'%a for a in mypsf.amp]), np.sum(pix-mod)))
ps.savefig()
# Write before normalizing!
T = fits_table()
T.amp = mypsf.mog.amp
T.mean = mypsf.mog.mean
T.var = mypsf.mog.var
T.writeto('psf3-w%i.fits' % band)
T.writeto('psf3.fits', append=(band != 1))
mypsf.weights.setParams(np.array(mypsf.weights.getParams()) /
sum(mypsf.weights.getParams()))
print 'Normalized PSF weights:', mypsf
mod = tractor.getModelImage(0)
print 'Mod sum:', mod.sum()
_plot_psf(pix, mod, mypsf, flux=src.brightness.getValue())
plt.suptitle('W%i psf3 (opt): %g = %s, resid %.3f' %
(band, np.sum(mypsf.amp), ','.join(['%.3f'%a for a in mypsf.amp]), np.sum(pix-mod)))
ps.savefig()
class MyGaussianPSF(NCircularGaussianPSF):
'''
A PSF model with strictly positive weights that sum to
unity.
'''
def __init__(self, sigmas, weights):
ww = np.array(weights)
ww = np.log(ww)
super(MyGaussianPSF, self).__init__(sigmas, ww)
@staticmethod
def getNamedParams():
return dict(sigmas=0, logweights=1)
def __str__(self):
return ('MyGaussianPSF: sigmas [ ' +
', '.join(['%.3f'%s for s in self.mysigmas]) +
' ], weights [ ' +
', '.join(['%.3f'%w for w in self.myweights]) +
' ]')
@property
def myweights(self):
ww = np.exp(self.logweights.getAllParams())
ww /= ww.sum()
return ww
@property
def weights(self):
ww = np.exp(self.logweights.getParams())
wsum = np.sum(np.exp(self.logweights.getAllParams()))
return ww / wsum
if band == 4:
# HACK
mypsf = MyGaussianPSF([1.7, 6.4, 15.0], [0.333, 0.666, 0.1])
else:
mypsf = MyGaussianPSF(gpsf.sigmas, gpsf.amp)
mypsf.radius = sh/2
tim.psf = mypsf
print 'Optimizing Params:'
tractor.printThawedParams()
for i in range(200):
#dlnp,X,alpha = tractor.optimize(damp=0.1)
dlnp,X,alpha = tractor.optimize(damp=1)
print i,'dlnp %.3g' % dlnp, 'psf', mypsf
if dlnp < 1e-6:
break
mypsf.sigmas.stepsize = len(mypsf.sigmas) * [1e-6]
mypsf.logweights.stepsize = len(mypsf.sigmas) * [1e-6]
for i in range(200):
#dlnp,X,alpha = tractor.optimize(damp=0.1)
dlnp,X,alpha = tractor.optimize(damp=1)
print i,'dlnp %.3g' % dlnp, 'psf', mypsf
if dlnp < 1e-6:
break
print 'PSF amps:', np.sum(mypsf.amp)
print 'pix sum:', pix.sum()
print 'Optimize source:', src
print 'Optimized PSF:', mypsf
mod = tractor.getModelImage(0)
print 'Mod sum:', mod.sum()
_plot_psf(pix, mod, mypsf, flux=src.brightness.getValue())
plt.suptitle('W%i psf3 (opt2): %g = %s, resid %.3f' %
(band, np.sum(mypsf.amp), ','.join(['%.3f'%a for a in mypsf.amp]), np.sum(pix-mod)))
ps.savefig()
# Write
mog = mypsf.getMixtureOfGaussians()
T = fits_table()
T.amp = mog.amp
T.mean = mog.mean
T.var = mog.var
T.writeto('psf4.fits', append=(band != 1))
def _allwise_psf_models():
global plotslice
# AllWISE PSF models
for band in [1,2,3,4]:
print
print 'W%i' % band
print
pix = reduce(np.add,
[fitsio.read('wise-w%i-psf-wpro-09x09-%02ix%02i.fits' %
(band, 1+(i/9), 1+(i%9)))
for i in range(9*9)])
print 'Pix', pix.shape
h,w = pix.shape
print 'Pix sum', pix.sum()
print 'Pix max', pix.max()
pix /= pix.sum()
psf = GaussianMixturePSF.fromStamp(pix)
print 'Fit PSF', psf
psfmodel = psf.getPointSourcePatch(0., 0., radius=h/2)
mod = psfmodel.patch
print 'Orig mod sum', mod.sum()
S = h/2
plotss = 30
plotslice = slice(S-plotss, -(S-plotss)), slice(S-plotss, -(S-plotss))
_plot_psf(pix, mod, psf)
plt.suptitle('W%i: from stamp' % band)
ps.savefig()
# Try concentric gaussian PSF
sh,sw = pix.shape
sigmas = []
for k in range(psf.mog.K):
v = psf.mog.var[k,:,:]
sigmas.append(np.sqrt(np.sqrt(v[0,0] * v[1,1])))
print 'Initializing concentric Gaussian model with sigmas', sigmas
print 'And weights', psf.mog.amp
# (Initial) W1:
# sig [7.1729391870564569, 24.098952864976805, 108.78869923786333]
# weights [ 0.80355109 0.15602835 0.04195982]
# sigmas [ 6.928, 17.091, 45.745 ], weights [ 0.760, 0.154, 0.073 ]
# W2:
# sig [8.2356371659198189, 25.741694812001221, 110.17431488810806]
# weights [ 0.80636191 0.1563742 0.03926182]
# sigmas [ 7.177, 10.636, 30.247 ], weights [ 0.500, 0.341, 0.134 ]
# W3:
# sig [6.860124763919889, 19.160849966251824, 134.20812055907825]
# weights [ 0.28227047 0.55080156 0.16706357]
# sigmas [ 6.850, 18.922, 109.121 ], weights [ 0.276, 0.554, 0.148 ]
# W4:
# [4.6423676603462001, 5.31542132669962, 18.375512539373585]
# weights [ 0.10764341 0.26915295 0.62320374]
# (Opt)
# sigmas [ 6.293, 6.314, 20.932 ], weights [ 0.129, 0.309, 0.598 ]
weights = psf.mog.amp
# HACK
if band == 4:
sigmas = [ 2., 6., 25. ]
weights = [ 0.5, 0.25, 0.25 ]
else:
sigmas = [ 8., 24., 100. ]
weights = [ 0.5, 0.25, 0.25 ]
gpsf = NCircularGaussianPSF(sigmas, weights)
gpsf.radius = sh/2
mypsf = gpsf
tim = Image(data=pix, invvar=1e6 * np.ones_like(pix), psf=mypsf)
cx,cy = sw/2, sh/2
src = PointSource(PixPos(cx, cy), Flux(1.0))
tractor = Tractor([tim], [src])
tim.freezeAllBut('psf')
#tractor.freezeParam('catalog')
src.freezeAllBut('pos')
print 'Optimizing Params:'
tractor.printThawedParams()
for i in range(100):
dlnp,X,alpha = tractor.optimize(damp=0.1)
print 'dlnp', dlnp
print 'PSF', mypsf
if dlnp < 0.001:
break
print 'PSF3(opt): flux', src.brightness
print 'PSF amps:', np.sum(mypsf.amp)
print 'PSF amps * Source brightness:', src.brightness.getValue() * np.sum(mypsf.amp)
print 'pix sum:', pix.sum()
print 'Optimize source:', src
print 'Optimized PSF:', mypsf
mod = tractor.getModelImage(0)
print 'Mod sum:', mod.sum()
print 'Mod min:', mod.min()
print 'Mod max:', mod.max()
_plot_psf(pix, mod, mypsf, flux=src.brightness.getValue())
plt.suptitle('W%i psf3 (opt): %g = %s, resid %.3f' %
(band, np.sum(mypsf.amp), ','.join(['%.3f'%a for a in mypsf.amp]), np.sum(pix-mod)))
ps.savefig()
# These postage stamps are subsampled at 8x the 2.75"/pix,
# except for the W4 one, which is at 4x that.
scale = 8
if band == 4:
scale = 4
T = fits_table()
T.amp = mypsf.mog.amp
T.mean = mypsf.mog.mean / scale
T.var = mypsf.mog.var / scale**2
T.writeto('psf-allwise-con3-w%i.fits' % band)
T.writeto('psf-allwise-con3.fits', append=(band != 1))
mypsf.weights.setParams(np.array(mypsf.weights.getParams()) /
sum(mypsf.weights.getParams()))
print 'Normalized PSF weights:', mypsf
mod = tractor.getModelImage(0)
print 'Mod sum:', mod.sum()
_plot_psf(pix, mod, mypsf, flux=src.brightness.getValue())
plt.suptitle('W%i psf3 (opt): %g = %s, resid %.3f' %
(band, np.sum(mypsf.amp), ','.join(['%.3f'%a for a in mypsf.amp]), np.sum(pix-mod)))
ps.savefig()
psf2 = GaussianMixturePSF(mypsf.mog.amp[:2]/mypsf.mog.amp[:2].sum(),
mypsf.mog.mean[:2,:], mypsf.mog.var[:2,:,:])
psf2.radius = sh/2
tim.psf = psf2
mod = tractor.getModelImage(0)
_plot_psf(lpix, mod, psf2, flux=src.brightness.getValue())
plt.suptitle('psf2 (init)')
ps.savefig()
src.freezeAllBut('brightness')
tractor.freezeParam('catalog')
for i in range(100):
print 'Optimizing PSF:'
tractor.printThawedParams()
dlnp,X,alpha = tractor.optimize(damp=1.)
tractor.freezeParam('images')
tractor.thawParam('catalog')
print 'Optimizing flux:'
tractor.printThawedParams()
tractor.optimize_forced_photometry()
tractor.thawParam('images')
tractor.freezeParam('catalog')
mod = tractor.getModelImage(0)
_plot_psf(lpix, mod, psf2, flux=src.brightness.getValue())
