Skip to content

GitLab

Commit d3ef2730 authored by Jakob Knollmueller's avatar Jakob Knollmueller
Browse files
Options

finally all errors gone? back to KL, samples now reasonable, performs excellent

parent 8bdaa000
Hide whitespace changes
Inline Side-by-side
Showing with 183 additions and 109 deletions
paper/1d_separation.py
View file @ d3ef2730
......@@ -4,132 +4,174 @@ from matplotlib import rc
rc('font',**{'family':'serif','serif':['Palatino']})
rc('text', usetex=True)
from matplotlib import pyplot as plt
from matplotlib.colors import LogNorm
from astropy.io import fits
import starblade as sb
from scipy.stats import invgamma
np.random.seed(42)
if __name__ == '__main__':
s_space = ift.RGSpace([128,128])
h_space = s_space.get_default_codomain()
FFT = ift.FFTOperator(h_space)
p_spec = lambda k: (1./(1+k)**4)
mod_p_spec = lambda k: p_spec(k*1024)*1024**4
binbounds = ift.PowerSpace.useful_binbounds(h_space,logarithmic=True)#, nbin=100)
binbounds = ift.PowerSpace.useful_binbounds(h_space,logarithmic=False)#, nbin=100)
p_space = ift.PowerSpace(h_space,binbounds=binbounds)
k_lengths=p_space.k_lengths
# p_spec = ift.Field(p_space,val=p_spec(p_space.k_lengths))
S = ift.create_power_operator(h_space, power_spectrum=p_spec)
sh = S.draw_sample()
s = FFT(sh)
# k_lengths = sh.domain[0].k_lengths
u = ift.Field(s_space, val = -12.)
# u = 3*(ift.Field.from_random('normal',s_space)-1)
# u.val[20,20] = 3
# u.val[15,96] = 4
# u.val[128,128] = 5
# u.val[65,33] = 6
# u.val[156,119] = 4.5
# u.val[16,125] = 4.5
# u.val[156,51] = 2
# u.val[235,62] = 3.5
# u.val[54,125] = 1.3
x=np.random.randint(0,s_space.shape[0],(s_space.shape[0]/2,1))
y=np.random.randint(0,s_space.shape[0],(s_space.shape[0]/2,1))
brightness = np.random.uniform(-2,5,(s_space.shape[0]/2,1))
u.val[x,y] = brightness
brightness =ift.log(ift.Field(s_space,val=invgamma(0.5).rvs(s_space.shape))/1000)
# u.val[200] = 1
# u.val[300] = 3
# u.val[500] = 4
# u.val[700] = 5
# u.val[900] = 2
# u.val[154] = 0.5
# u.val[421] = 0.25
# u.val[652] = 1
# u.val[1002] = 2.5
# R = ift.FFTSmoothingOperator(s_space,sigma=0.001)
# R(u)
q=1e-3
alpha = 1.5
brightness =ift.log(ift.Field(s_space,val=invgamma.rvs(alpha-1.,scale=q,size=s_space.shape)))
u=brightness
d = ift.exp(s) +ift.exp(u)
data = d.val
energy1 = sb.build_starblade(data,1.25, newton_iterations=5, cg_iterations=500, q=1e-30)#, manual_power_spectrum= mod_p_spec)
energy2 = sb.build_starblade(data,1.5, newton_iterations=5, cg_iterations=500, q=1e-30)#, manual_power_spectrum= mod_p_spec)
energy3 = sb.build_starblade(data,1.75, newton_iterations=5, cg_iterations=500, q=1e-30)#, manual_power_spectrum=mod_p_spec)
hdu = fits.PrimaryHDU(data)
hdul = fits.HDUList([hdu])
hdul.writeto('mock_data.fits',overwrite=True)
sextracted64 = fits.open('check64.fits')
sextracted64 = sextracted64[0].data
sextracted64 = sextracted64.clip(0.001)
sextracted8 = fits.open('check8.fits')
sextracted8 = sextracted8[0].data
sextracted8 = sextracted8.clip(0.001)
energy1 = sb.build_starblade(data,1.0, newton_iterations=200, cg_iterations=3,
q=q)#, manual_power_spectrum= p_spec)
energy2 = sb.build_starblade(data,1.5, newton_iterations=200, cg_iterations=3,
q=q)#, manual_power_spectrum= p_spec)
energy3 = sb.build_starblade(data,3., newton_iterations=200, cg_iterations=3,
q=q)#, manual_power_spectrum= p_spec)
# ift.extra.check_value_gradient_consistency(energy1, tol=1e-3)
for i in range(10):
energy1 = sb.starblade_iteration(energy1, samples=0)
energy2 = sb.starblade_iteration(energy2, samples=0)
energy3 = sb.starblade_iteration(energy3, samples=0)
plt.imsave("2d_data.png",np.log(data))
energy1 = sb.starblade_iteration(energy1, samples=5)
energy2 = sb.starblade_iteration(energy2, samples=5)
energy3 = sb.starblade_iteration(energy3, samples=5)
print "error energy1:", np.sqrt(((ift.exp(energy1.s).val - ift.exp(s).val) ** 2).mean())
print "error energy2:", np.sqrt(((ift.exp(energy2.s).val - ift.exp(s).val) ** 2).mean())
print "error energy3:", np.sqrt(((ift.exp(energy3.s).val - ift.exp(s).val) ** 2).mean())
samples = []
n=30
for i in range(n):
samples.append(energy1.curvature.inverse.draw_sample())
m = 0
v = 0
s_s=0
v_s = 0
pos_tanh = ift.library.nonlinearities.PositiveTanh()
p=0
for sample in samples:
a_s = pos_tanh(sample)
m += a_s
v += a_s**2
s_s += ift.log(d*(1-a_s))
v_s += ift.log(d*(1-a_s))**2
p += ift.power_analyze(FFT.adjoint(s))
m /= n
v /= n
v -= m**2
s_s /= n
v_s /= n
v_s -= s_s**2
p /= n
lim_low = 1e-1
lim_high = 1e4
size = 15
plt.gray()
plt.figure()
plt.imshow(data,norm=LogNorm())
cbar = plt.colorbar()
cbar.set_label('intensity', size=size)
plt.axis('off')
plt.title('logarithmic mock data',size=size)
plt.savefig('2d_data.pdf')
# plt.figure(figsize=(8, 8))
# plt.plot(data, 'k-')
f, (ax0, ax1,ax2) = plt.subplots(3, sharex=True, sharey=True)
f, (ax0, ax1,ax2,ax3,ax4) = plt.subplots(5, sharex=True, sharey=True,figsize=(8, 10))
plt.suptitle('diffuse components', size=size)
ax0.plot(ift.exp(energy1.s).val, 'k-',alpha=0.1)
for i in range(energy1.s.val.shape[0]):
ax0.plot(ift.exp(energy1.s).val[i], 'k-',alpha=(0.15/(energy1.s.val.shape[0])*i))
ax0.yaxis.set_label_position("right")
ax0.set_ylabel(r'$\alpha = 1.25$', size=size)
ax0.set_ylim(1e-1,1e3)
ax0.set_ylabel(r'$\alpha = 1.0$', size=size)
ax0.set_ylim(lim_low ,lim_high)
ax0.set_yscale("log")
ax1.plot(ift.exp(energy2.s).val, 'k-',alpha=0.1)
for i in range(energy1.s.val.shape[0]):
ax1.plot(ift.exp(energy2.s).val[i], 'k-', alpha=(0.15 / (energy1.s.val.shape[0]) * i))
ax1.yaxis.set_label_position("right")
ax1.set_ylabel(r'$\alpha = 1.5$', size=size)
ax2.plot(ift.exp(energy3.s).val, 'k-',alpha=0.1)
for i in range(energy1.s.val.shape[0]):
ax2.plot(ift.exp(energy3.s).val[i], 'k-', alpha=(0.15 / (energy1.s.val.shape[0]) * i))
ax2.yaxis.set_label_position("right")
ax2.set_ylabel(r'$\alpha = 1.75$', size=size)
ax2.set_ylabel(r'$\alpha = 3.0$', size=size)
for i in range(energy1.s.val.shape[0]):
ax3.plot(sextracted64[i], 'k-', alpha=(0.15 / (energy1.s.val.shape[0]) * i))
ax3.yaxis.set_label_position("right")
ax3.set_ylabel('sextractor'+ '\n'+ r'default', size=size)
for i in range(energy1.s.val.shape[0]):
ax4.plot((sextracted8)[i].clip(0.0001), 'k-', alpha=(0.15 / (energy1.s.val.shape[0]) * i))
ax4.yaxis.set_label_position("right")
ax4.set_ylabel(r'sextractor'+'\n' + r'\textsf{BACK\_SIZE}'+r'$=8$', size=size)
plt.savefig('1d_diffuse.pdf')
plt.figure()
f, (ax0, ax1,ax2) = plt.subplots(3, sharex=True, sharey=True)
f, (ax0, ax1,ax2,ax3,ax4) = plt.subplots(5, sharex=True, sharey=True,figsize=(8, 10))
plt.suptitle('point-like components', size=size)
ax0.plot(ift.exp(energy1.u).val, 'k-',alpha=0.1)
for i in range(energy1.s.val.shape[0]):
ax0.plot(ift.exp(energy1.u).val[i], 'k-', alpha=(0.15 / (energy1.s.val.shape[0]) * i))
ax0.yaxis.set_label_position("right")
ax0.set_ylabel(r'$\alpha = 1.25$', size=size)
ax0.set_ylim(1e-1,1e3)
ax0.set_ylabel(r'$\alpha = 1.0$', size=size)
ax0.set_ylim(lim_low ,lim_high)
ax0.set_yscale("log")
ax1.plot(ift.exp(energy2.u).val, 'k-',alpha=0.1)
for i in range(energy1.s.val.shape[0]):
ax1.plot(ift.exp(energy2.u).val[i], 'k-', alpha=(0.15 / (energy1.s.val.shape[0]) * i))
ax1.yaxis.set_label_position("right")
ax1.set_ylabel(r'$\alpha = 1.5$', size=size)
ax2.plot(ift.exp(energy3.u).val, 'k-',alpha=0.1)
for i in range(energy1.s.val.shape[0]):
ax2.plot(ift.exp(energy3.u).val[i], 'k-', alpha=(0.15 / (energy1.s.val.shape[0]) * i))
ax2.yaxis.set_label_position("right")
ax2.set_ylabel(r'$\alpha = 1.75$', size=size)
ax2.set_ylabel(r'$\alpha = 3.0$', size=size)
for i in range(energy1.s.val.shape[0]):
ax3.plot((data-sextracted64)[i], 'k-', alpha=(0.15 / (energy1.s.val.shape[0]) * i))
ax3.yaxis.set_label_position("right")
ax3.set_ylabel(r'sextractor'+ '\n'+ r'default', size=size)
for i in range(energy1.s.val.shape[0]):
ax4.plot((data-sextracted8)[i].clip(0.0001), 'k-', alpha=(0.15 / (energy1.s.val.shape[0]) * i))
ax4.yaxis.set_label_position("right")
ax4.set_ylabel(r'sextractor'+'\n' + r'\textsf{BACK\_SIZE}'+r'$=8$', size=size)
ax0.set_yscale("log")
ax0.set_ylim(1e-1,1e3)
ax0.set_ylim(lim_low ,lim_high)
# plt.ylim(1e-0)
plt.savefig('1d_points.pdf')
plt.figure()
f, (ax0, ax1,ax2) = plt.subplots(3, sharex=True, sharey=True)
plt.suptitle('data and true components', size=size)
ax0.plot(data, 'k-',alpha=0.1)
for i in range(energy1.s.val.shape[0]):
ax0.plot(data[i], 'k-', alpha=(0.15 / (energy1.s.val.shape[0]) * i))
ax0.set_yscale("log")
ax0.set_ylim(1e-1,1e3)
ax0.set_ylim(lim_low ,lim_high)
ax0.yaxis.set_label_position("right")
ax0.set_ylabel(r'data', size=size)
ax1.plot(ift.exp(s).val, 'k-',alpha=0.1)
for i in range(energy1.s.val.shape[0]):
ax1.plot(ift.exp(s).val[i], 'k-', alpha=(0.15 / (energy1.s.val.shape[0]) * i))
ax1.yaxis.set_label_position("right")
ax1.set_ylabel(r'diffuse', size=size)
ax2.plot(ift.exp(u).val, 'k-',alpha=0.1)
for i in range(energy1.s.val.shape[0]):
ax2.plot(ift.exp(u).val[i], 'k-', alpha=(0.15 / (energy1.s.val.shape[0]) * i))
ax2.yaxis.set_label_position("right")
ax2.set_ylabel(r'point-like', size=size)
......@@ -137,25 +179,25 @@ if __name__ == '__main__':
plt.savefig('1d_data.pdf')
plt.figure()
binbounds = ift.PowerSpace.useful_binbounds(energy2.FFT.domain[0],logarithmic=True)#, nbin=100)
power1 = ift.power_analyze(energy2.FFT.inverse((energy1.s)),binbounds=binbounds)
power2 = ift.power_analyze(energy2.FFT.inverse((energy2.s)),binbounds=binbounds)
power3 = ift.power_analyze(energy2.FFT.inverse((energy3.s)),binbounds=binbounds)
binbounds = ift.PowerSpace.useful_binbounds(energy2.FFT.domain[0],logarithmic=False)#, nbin=100)
power1 = ift.power_analyze(energy2.FFT.adjoint((energy1.s)),binbounds=binbounds)
power2 = ift.power_analyze(energy2.FFT.adjoint((energy2.s)),binbounds=binbounds)
power3 = ift.power_analyze(energy2.FFT.adjoint((energy3.s)),binbounds=binbounds)
pp1 = energy1.power_spectrum
pp2 = energy2.power_spectrum
pp3 = energy3.power_spectrum
real_power = ift.power_analyze(sh)
power_d = ift.power_analyze(energy2.FFT.inverse(ift.log(energy2.d)),binbounds=binbounds)
power_d = ift.power_analyze(energy2.FFT.adjoint(ift.log(energy2.d)),binbounds=binbounds)
power_s = ift.power_analyze(energy2.FFT.adjoint(s),binbounds=binbounds)
# power_u = ift.power_analyze(energy2.FFT.inverse(ift.exp(energy2.u)),binbounds=binbounds)
k_lengths=power1.domain[0].k_lengths
plt.plot(k_lengths, p_spec(k_lengths*1024.)*1024**4, 'k-', label='theoretical')
plt.plot(k_lengths, power_d.val, 'k:', label='data')
plt.plot(k_lengths, power1.val, 'k-', label=(r'$\alpha = 1.25$'), alpha=0.6)
# plt.plot(k_lengths, p_spec(k_lengths*1024.)*1024, 'k-', label='theoretical')
plt.plot(k_lengths, power_d.val, 'k+-', label='data')
plt.plot(k_lengths, power1.val, 'k-', label=(r'$\alpha = 1.$'), alpha=0.15)
plt.plot(k_lengths, power2.val, 'k-', label=(r'$\alpha = 1.5$'),alpha=0.3)
plt.plot(k_lengths, power3.val, 'k-', label=(r'$\alpha = 1.75$'), alpha=0.15)
plt.plot(k_lengths, pp1.val, 'r-', label=(r'$\alpha = 1.25$'), alpha=0.6)
plt.plot(k_lengths, pp2.val, 'r-', label=(r'$\alpha = 1.5$'),alpha=0.3)
plt.plot(k_lengths, pp3.val, 'r-', label=(r'$\alpha = 1.75$'), alpha=0.15)
plt.plot(k_lengths, power3.val, 'k-', label=(r'$\alpha = 3.0$'), alpha=0.6)
plt.plot(k_lengths, power_s.val, 'k:', label=('signal'))
# plt.plot(k_lengths, power_u.val, 'k:',label='point-like')
plt.legend()
......@@ -166,4 +208,9 @@ if __name__ == '__main__':
plt.xlabel('harmonic mode',size=15)
plt.savefig('1d_power.pdf')
plt.close('all')
print "error energy1:", np.sqrt(((ift.exp(energy1.s).val-ift.exp(s).val)**2).mean())
print "error energy2:", np.sqrt(((ift.exp(energy2.s).val-ift.exp(s).val)**2).mean())
print "error energy3:", np.sqrt(((ift.exp(energy3.s).val-ift.exp(s).val)**2).mean())
print "error back_size8:", np.sqrt(((sextracted8-ift.exp(s).val)**2).mean())
print "error back_size64:", np.sqrt(((sextracted64-ift.exp(s).val)**2).mean())
starblade/starblade_energy.py
View file @ d3ef2730
......@@ -16,9 +16,10 @@
#
# Starblade is being developed at the Max-Planck-Institut fuer Astrophysik
from nifty4 import Energy, Field, log, exp, DiagonalOperator, create_power_operator
from nifty4 import Energy, Field, log, exp, DiagonalOperator,\
create_power_operator, SandwichOperator, ScalingOperator, InversionEnabler
from nifty4.library import WienerFilterCurvature
from nifty4.library.nonlinearities import PositiveTanh
from nifty4.library.nonlinearities import PositiveTanh, Tanh
class StarbladeEnergy(Energy):
......@@ -64,12 +65,16 @@ class StarbladeEnergy(Energy):
self.update_power = parameters['update_power']
self.newton_iterations = parameters['newton_iterations']
pos_tanh = PositiveTanh()
self.S = self.FFT * self.correlation * self.FFT.adjoint
tanh = Tanh()
self.S = SandwichOperator.make(self.FFT.adjoint, self.correlation)
# self.S = self.FFT * self.correlation * self.FFT.adjoint
self.a = pos_tanh(self.position)
self.a_p = pos_tanh.derivative(self.position)
self.a_pp = -tanh(position)*tanh.derivative(self.position)
self.u = log(self.d * self.a)
self.u_p = self.a_p/self.a
self.u_a = -log(self.a)
self.u_ap = - self.a_p/self.a
one_m_a = 1 - self.a
self.s = log(self.d * one_m_a)
self.s_p = - self.a_p / one_m_a
......@@ -80,24 +85,40 @@ class StarbladeEnergy(Energy):
@property
def value(self):
point = 0
diffuse = 0
det = 0
diffuse = 0.5 * self.s.vdot(self.S.inverse(self.s))
point = (self.alpha-1).vdot(self.u) + self.q.vdot(exp(-self.u))
det = self.s.integrate()
det = - self.s.sum()
det += - self.u_a.sum()
det += -log(self.a_p).sum()
det += 0.5 / self.var_x * self.position.vdot(self.position)
return diffuse + point + det
return diffuse + point + det
@property
def gradient(self):
point = 0
diffuse = 0
det = 0
diffuse = self.S.inverse(self.s) * self.s_p
point = (self.alpha - 1) * self.u_p - self.q * exp(-self.u) * self.u_p
det = self.position / self.var_x
det += self.s_p
return diffuse + point + det
det += - self.s_p
det += - self.u_ap
det += -1./self.a_p * self.a_pp
return +diffuse + point +det
@property
def curvature(self):
point = self.q * exp(-self.u) * self.u_p ** 2
R = self.FFT.inverse * self.s_p
N = self.correlation
S = DiagonalOperator(1/(point + 1/self.var_x))
return WienerFilterCurvature(R=R, N=N, S=S, inverter=self.inverter)
# R = self.FFT.inverse * self.s_p
# N = self.correlation
N_inv = DiagonalOperator(point + 1/self.var_x )#+ 2*self.a_p))
R = ScalingOperator(1., point.domain)
S_p = DiagonalOperator(self.s_p)
my_S_inv = SandwichOperator.make(self.FFT.adjoint.inverse.adjoint * S_p, self.correlation.inverse)
curv = InversionEnabler(N_inv + my_S_inv, self.inverter)
return curv
# return WienerFilterCurvature(R=R, N=N, S=S, inverter=self.inverter)
starblade/starblade_kl.py
View file @ d3ef2730
......@@ -16,7 +16,7 @@
#
# Starblade is being developed at the Max-Planck-Institut fuer Astrophysik
from nifty4 import Energy, Field, DiagonalOperator, InversionEnabler
from nifty4 import Energy, Field, DiagonalOperator, InversionEnabler, full
from starblade_energy import StarbladeEnergy
class StarbladeKL(Energy):
......@@ -68,7 +68,7 @@ class StarbladeKL(Energy):
@property
def gradient(self):
gradient = Field.zeros(self.position.domain)
gradient = full(self.position.domain,0.)
for energy in self.energy_list:
gradient += energy.gradient
gradient /= len(self.energy_list)
......@@ -76,7 +76,7 @@ class StarbladeKL(Energy):
@property
def curvature(self):
curvature = DiagonalOperator(Field.zeros(self.position.domain))
curvature = DiagonalOperator(full(self.position.domain, 0.))
for energy in self.energy_list:
curvature += energy.curvature
curvature *= Field(self.position.domain,val=1./len(self.energy_list))
......
starblade/sugar.py
View file @ d3ef2730
......@@ -43,16 +43,17 @@ def build_starblade(data, alpha=1.5, q=1e-40, cg_iterations=100, newton_iteratio
If it is not specified, the algorithm will try to infer it via critical filtering.
"""
s_space = ift.RGSpace(data.shape, distances=len(data.shape) * [1])
s_space = ift.RGSpace(data.shape)#, distances=len(data.shape) * [1])
h_space = s_space.get_default_codomain()
data = ift.Field(s_space,val=data)
FFT = ift.FFTOperator(h_space, target=s_space)
binbounds = ift.PowerSpace.useful_binbounds(h_space, logarithmic = True)
binbounds = ift.PowerSpace.useful_binbounds(h_space, logarithmic = False)
p_space = ift.PowerSpace(h_space, binbounds=binbounds)
if manual_power_spectrum is None:
initial_spectrum = ift.power_analyze(FFT.inverse_times(ift.log(data)),
initial_spectrum = ift.power_analyze(FFT.adjoint(ift.log(data)),
binbounds=p_space.binbounds)
initial_spectrum /= (p_space.k_lengths+1.)**4
initial_spectrum /= 100*(p_space.k_lengths+1.)**4
# initial_spectrum = ift.Field(p_space,val=1e-3)
update_power = True
......@@ -89,19 +90,24 @@ def starblade_iteration(starblade, samples=3):
tol_abs_gradnorm=1e-8,
iteration_limit=starblade.newton_iterations)
minimizer = ift.RelaxedNewton(controller=controller)
# if len(sample_list)>0:
# energy = StarbladeKL(starblade.position, samples=sample_list, parameters=starblade.parameters)
# else:
# minimizer = ift.VL_BFGS(controller=controller)
energy = starblade
sample_list = []
for i in range(samples):
sample = energy.curvature.inverse.draw_sample()
sample_list.append(sample)
if len(sample_list)>0:
energy = StarbladeKL(starblade.position, samples=sample_list, parameters=starblade.parameters)
else:
energy = starblade
energy, convergence = minimizer(energy)
energy = StarbladeEnergy(energy.position, parameters=energy.parameters)
sample_list = []
for i in range(samples):
sample = energy.curvature.inverse.draw_sample()
sample_list.append(sample)
if len(sample_list) == 0:
sample_list.append(energy.position)
# energy = StarbladeKL(energy.position, samples=sample_list, parameters=energy.parameters)
new_position = energy.position
new_parameters = energy.parameters
......@@ -171,11 +177,11 @@ def update_power(energy):
if isinstance(energy, StarbladeKL):
power = 0.
for en in energy.energy_list:
power = ift.power_analyze(energy.parameters['FFT'].inverse_times(en.s),
power += ift.power_analyze(energy.parameters['FFT'].inverse(en.s),
binbounds=en.parameters['power_spectrum'].domain[0].binbounds)
# power /= len(energy.energy_list)
power /= len(energy.energy_list)
else:
power = ift.power_analyze(energy.FFT.inverse_times(energy.s),
power = ift.power_analyze(energy.FFT.inverse(energy.s),
binbounds=energy.parameters['power_spectrum'].domain[0].binbounds)
return power
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment