def test_fail():
# Some vv noisy images that result in errors in the fit to check the error reporting.
scale = 1.3
g1 = 0.33
g2 = -0.27
flux = 15
noise = 2.
seed = 1234
psf = galsim.Moffat(half_light_radius=1.0, beta=2.5, trunc=3.0)
psf = psf.dilate(scale).shear(g1=g1, g2=g2).withFlux(flux)
image = psf.drawImage(nx=64, ny=64, scale=0.3)
weight = image.copy()
weight.fill(1 / noise**2)
noisy_image = image.copy()
rng = galsim.BaseDeviate(seed)
noisy_image.addNoise(galsim.GaussianNoise(sigma=noise, rng=rng))
star1 = piff.Star(piff.StarData(image, image.true_center, weight), None)
star2 = piff.Star(piff.StarData(noisy_image, image.true_center, weight),
None)
model1 = piff.Moffat(fastfit=True, beta=2.5)
with np.testing.assert_raises(RuntimeError):
model1.initialize(star2)
with np.testing.assert_raises(RuntimeError):
model1.fit(star2)
star3 = model1.initialize(star1)
star3 = model1.fit(star3)
star3 = piff.Star(star2.data, star3.fit)
with np.testing.assert_raises(RuntimeError):
model1.fit(star3)
# This is contrived to hit the fit failure for the reference.
# I'm not sure what realistic use case would actually hit it, but at least it's
# theoretically possible to fail there.
model2 = piff.GSObjectModel(galsim.InterpolatedImage(noisy_image),
fastfit=True)
with np.testing.assert_raises(RuntimeError):
model2.initialize(star1)
model3 = piff.Moffat(fastfit=False,
beta=2.5,
scipy_kwargs={'max_nfev': 10})
with np.testing.assert_raises(RuntimeError):
model3.initialize(star2)
with np.testing.assert_raises(RuntimeError):
model3.fit(star2).fit
star3 = model3.initialize(star1)
star3 = model3.fit(star3)
star3 = piff.Star(star2.data, star3.fit)
with np.testing.assert_raises(RuntimeError):
model3.fit(star3)
def test_var():
"""Check that the variance estimate in params_var is sane.
"""
# Here is the true PSF
scale = 1.3
g1 = 0.23
g2 = -0.17
du = 0.1
dv = 0.4
flux = 500
wcs = galsim.JacobianWCS(0.26, 0.05, -0.08, -0.29)
noise = 0.2
gsobjs = [
galsim.Gaussian(sigma=1.0),
galsim.Kolmogorov(half_light_radius=1.0),
galsim.Moffat(half_light_radius=1.0, beta=3.0),
galsim.Moffat(half_light_radius=1.0, beta=2.5, trunc=3.0)
]
# Mix of centered = True/False,
# fastfit = True/False,
# include_pixel = True/False
models = [
piff.Gaussian(fastfit=False, include_pixel=False, centered=False),
piff.Kolmogorov(fastfit=True, include_pixel=True, centered=False),
piff.Moffat(fastfit=False, beta=4.8, include_pixel=True,
centered=True),
piff.Moffat(fastfit=True,
beta=2.5,
trunc=3.0,
include_pixel=False,
centered=True)
]
names = ['Gaussian', 'Kolmogorov', 'Moffat3', 'Moffat2.5']
for gsobj, model, name in zip(gsobjs, models, names):
print()
print("gsobj = ", gsobj)
print()
psf = gsobj.dilate(scale).shear(g1=g1, g2=g2).shift(du,
dv).withFlux(flux)
image = psf.drawImage(nx=64, ny=64, wcs=wcs, method='no_pixel')
weight = image.copy()
weight.fill(1 / noise**2)
# Save this one without noise.
image1 = image.copy()
image1.addNoise(galsim.GaussianNoise(sigma=noise))
# Make a StarData instance for this image
stardata = piff.StarData(image, image.true_center, weight)
star = piff.Star(stardata, None)
star = model.initialize(star)
fit = model.fit(star).fit
file_name = 'input/test_%s_var.npz' % name
print(file_name)
if not os.path.isfile(file_name):
num_runs = 1000
all_params = []
for i in range(num_runs):
image1 = image.copy()
image1.addNoise(galsim.GaussianNoise(sigma=noise))
sd = piff.StarData(image1, image1.true_center, weight)
s = piff.Star(sd, None)
try:
s = model.initialize(s)
s = model.fit(s)
except RuntimeError as e: # Occasionally hsm fails.
print('Caught ', e)
continue
print(s.fit.params)
all_params.append(s.fit.params)
var = np.var(all_params, axis=0)
np.savez(file_name, var=var)
var = np.load(file_name)['var']
print('params = ', fit.params)
print('empirical var = ', var)
print('piff estimate = ', fit.params_var)
print('ratio = ', fit.params_var / var)
print('max ratio = ', np.max(fit.params_var / var))
print('min ratio = ', np.min(fit.params_var / var))
print('mean ratio = ', np.mean(fit.params_var / var))
# Note: The fastfit=False estimates are better -- typically better than 10%
# The fastfit=True estimates are much rougher. Especially size. Need rtol=0.3.
np.testing.assert_allclose(fit.params_var, var, rtol=0.3)
def test_direct():
""" Simple test for directly instantiated Gaussian, Kolmogorov, and Moffat without going through
GSObjectModel explicitly.
"""
# Here is the true PSF
scale = 1.3
g1 = 0.23
g2 = -0.17
du = 0.1
dv = 0.4
gsobjs = [
galsim.Gaussian(sigma=1.0),
galsim.Kolmogorov(half_light_radius=1.0),
galsim.Moffat(half_light_radius=1.0, beta=3.0),
galsim.Moffat(half_light_radius=1.0, beta=2.5, trunc=3.0)
]
models = [
piff.Gaussian(fastfit=True, include_pixel=False),
piff.Kolmogorov(fastfit=True, include_pixel=False),
piff.Moffat(fastfit=True, beta=3.0, include_pixel=False),
piff.Moffat(fastfit=True, beta=2.5, trunc=3.0, include_pixel=False)
]
for gsobj, model in zip(gsobjs, models):
print()
print("gsobj = ", gsobj)
print()
psf = gsobj.dilate(scale).shear(g1=g1, g2=g2).shift(du, dv)
# Draw the PSF onto an image. Let's go ahead and give it a non-trivial WCS.
wcs = galsim.JacobianWCS(0.26, 0.05, -0.08, -0.29)
image = galsim.Image(64, 64, wcs=wcs)
# This is only going to come out right if we (unphysically) don't convolve by the pixel.
psf.drawImage(image, method='no_pixel')
# Make a StarData instance for this image
stardata = piff.StarData(image, image.true_center)
star = piff.Star(stardata, None)
star = model.initialize(star)
# First try fastfit.
print('Fast fit')
fit = model.fit(star).fit
print('True scale = ', scale, ', model scale = ', fit.params[0])
print('True g1 = ', g1, ', model g1 = ', fit.params[1])
print('True g2 = ', g2, ', model g2 = ', fit.params[2])
print('True du = ', du, ', model du = ', fit.center[0])
print('True dv = ', dv, ', model dv = ', fit.center[1])
# This test is fairly accurate, since we didn't add any noise and didn't convolve by
# the pixel, so the image is very accurately a sheared GSObject.
# These tests are more strict above. The truncated Moffat included here but not there
# doesn't work quite as well.
np.testing.assert_allclose(fit.params[0], scale, rtol=1e-4)
np.testing.assert_allclose(fit.params[1], g1, rtol=0, atol=1e-5)
np.testing.assert_allclose(fit.params[2], g2, rtol=0, atol=1e-5)
np.testing.assert_allclose(fit.center[0], du, rtol=0, atol=1e-5)
np.testing.assert_allclose(fit.center[1], dv, rtol=0, atol=1e-5)
# Also need to test ability to serialize
outfile = os.path.join('output', 'gsobject_direct_test.fits')
with fitsio.FITS(outfile, 'rw', clobber=True) as f:
model.write(f, 'psf_model')
with fitsio.FITS(outfile, 'r') as f:
roundtrip_model = piff.GSObjectModel.read(f, 'psf_model')
assert model.__dict__ == roundtrip_model.__dict__
# repeat with fastfit=False
models = [
piff.Gaussian(fastfit=False, include_pixel=False),
piff.Kolmogorov(fastfit=False, include_pixel=False),
piff.Moffat(fastfit=False, beta=3.0, include_pixel=False),
piff.Moffat(fastfit=False, beta=2.5, trunc=3.0, include_pixel=False)
]
for gsobj, model in zip(gsobjs, models):
print()
print("gsobj = ", gsobj)
print()
psf = gsobj.dilate(scale).shear(g1=g1, g2=g2).shift(du, dv)
# Draw the PSF onto an image. Let's go ahead and give it a non-trivial WCS.
wcs = galsim.JacobianWCS(0.26, 0.05, -0.08, -0.29)
image = galsim.Image(64, 64, wcs=wcs)
# This is only going to come out right if we (unphysically) don't convolve by the pixel.
psf.drawImage(image, method='no_pixel')
# Make a StarData instance for this image
stardata = piff.StarData(image, image.true_center)
star = piff.Star(stardata, None)
star = model.initialize(star)
print('Slow fit')
fit = model.fit(star).fit
print('True scale = ', scale, ', model scale = ', fit.params[0])
print('True g1 = ', g1, ', model g1 = ', fit.params[1])
print('True g2 = ', g2, ', model g2 = ', fit.params[2])
print('True du = ', du, ', model du = ', fit.center[0])
print('True dv = ', dv, ', model dv = ', fit.center[1])
# This test is fairly accurate, since we didn't add any noise and didn't convolve by
# the pixel, so the image is very accurately a sheared GSObject.
np.testing.assert_allclose(fit.params[0], scale, rtol=1e-5)
np.testing.assert_allclose(fit.params[1], g1, rtol=0, atol=1e-5)
np.testing.assert_allclose(fit.params[2], g2, rtol=0, atol=1e-5)
np.testing.assert_allclose(fit.center[0], du, rtol=0, atol=1e-5)
np.testing.assert_allclose(fit.center[1], dv, rtol=0, atol=1e-5)
# Also need to test ability to serialize
outfile = os.path.join('output', 'gsobject_direct_test.fits')
with fitsio.FITS(outfile, 'rw', clobber=True) as f:
model.write(f, 'psf_model')
with fitsio.FITS(outfile, 'r') as f:
roundtrip_model = piff.GSObjectModel.read(f, 'psf_model')
assert model.__dict__ == roundtrip_model.__dict__