Skip to content
GitLab
Commit 89c555e1 authored by Jakob Knollmueller's avatar Jakob Knollmueller
Browse files

added memorization of the residual in WienerFilterEnergy

parent 0fd5d047
Pipeline #14033 passed with stage
in 5 minutes
Hide whitespace changes
Inline Side-by-side
demos/critical_filtering.py
View file @ 89c555e1
......@@ -11,18 +11,16 @@ rank = comm.rank
np.random.seed(42)
def plot_parameters(m,t,t_true, t_real, t_d):
def plot_parameters(m,t,p, p_d):
x = log(t.domain[0].kindex)
m = fft.adjoint_times(m)
m_data = m.val.get_full_data().real
t_data = t.val.get_full_data().real
t_d_data = t_d.val.get_full_data().real
t_true_data = t_true.val.get_full_data().real
t_real_data = t_real.val.get_full_data().real
pl.plot([go.Heatmap(z=m_data)], filename='map.html')
pl.plot([go.Scatter(x=x,y=t_data), go.Scatter(x=x ,y=t_true_data),
go.Scatter(x=x, y=t_real_data), go.Scatter(x=x, y=t_d_data)], filename="t.html")
m = m.val.get_full_data().real
t = t.val.get_full_data().real
p = p.val.get_full_data().real
p_d = p_d.val.get_full_data().real
pl.plot([go.Heatmap(z=m)], filename='map.html')
pl.plot([go.Scatter(x=x,y=t), go.Scatter(x=x ,y=p), go.Scatter(x=x, y=p_d)], filename="t.html")
class AdjointFFTResponse(LinearOperator):
......@@ -63,11 +61,11 @@ if __name__ == "__main__":
h_space = fft.target[0]
# Setting up power space
p_space = PowerSpace(h_space, logarithmic=False,
p_space = PowerSpace(h_space, logarithmic=True,
distribution_strategy=distribution_strategy)#, nbin=5)
# Choosing the prior correlation structure and defining correlation operator
p_spec = (lambda k: (.05 / (k + 1) ** 3))
p_spec = (lambda k: (.5 / (k + 1) ** 3))
S = create_power_operator(h_space, power_spectrum=p_spec,
distribution_strategy=distribution_strategy)
......@@ -87,7 +85,7 @@ if __name__ == "__main__":
#Adding a harmonic transformation to the instrument
R = AdjointFFTResponse(fft, Instrument)
noise = .1
noise = 1.
N = DiagonalOperator(s_space, diagonal=noise, bare=True)
n = Field.from_random(domain=s_space,
random_type='normal',
......@@ -112,8 +110,8 @@ if __name__ == "__main__":
print (x, iteration)
minimizer1 = RelaxedNewton(convergence_tolerance=0,
convergence_level=1,
minimizer1 = RelaxedNewton(convergence_tolerance=10e-2,
convergence_level=2,
iteration_limit=3,
callback=convergence_measure)
minimizer2 = VL_BFGS(convergence_tolerance=0,
......@@ -125,8 +123,8 @@ if __name__ == "__main__":
flat_power = Field(p_space,val=10e-8)
m0 = flat_power.power_synthesize(real_signal=True)
# t0 = Field(p_space, val=log(1./(1+p_space.kindex)**2))
t0 = data_power- 1.
t0 = Field(p_space, val=log(1./(1+p_space.kindex)**2))
for i in range(500):
......@@ -142,42 +140,15 @@ if __name__ == "__main__":
m0 = map_energy.position
D0 = map_energy.curvature
# Initializing the power energy with updated parameters
power_energy = CriticalPowerEnergy(position=t0, m=m0, D=D0, sigma=100., samples=5)
power_energy = CriticalPowerEnergy(position=t0, m=m0, D=D0, sigma=100., samples=3)
(power_energy, convergence) = minimizer1(power_energy)
# Setting new power spectrum
t0.val = power_energy.position.val.real
# Plotting current estimate
plot_parameters(m0,t0,log(sp),realized_power, data_power)
# Transforming fields to position space for plotting
ss = fft.adjoint_times(sh)
m = fft.adjoint_times(map_energy.position)
plot_parameters(m0,t0,log(sp), data_power)
# Plotting
d_data = d.val.get_full_data().real
if rank == 0:
pl.plot([go.Heatmap(z=d_data)], filename='data.html')
tt_data = power_energy.position.val.get_full_data().real
t_data = log(sp**2).val.get_full_data().real
if rank == 0:
pl.plot([go.Scatter(y=t_data),go.Scatter(y=tt_data)], filename="t.html")
ss_data = ss.val.get_full_data().real
if rank == 0:
pl.plot([go.Heatmap(z=ss_data)], filename='ss.html')
sh_data = sh.val.get_full_data().real
if rank == 0:
pl.plot([go.Heatmap(z=sh_data)], filename='sh.html')
m_data = m.val.get_full_data().real
if rank == 0:
pl.plot([go.Heatmap(z=m_data)], filename='map.html')
demos/laplace_testing.py deleted 100644 → 0
View file @ 0fd5d047
from nifty import *
import numpy as np
from nifty import Field,\
EndomorphicOperator,\
PowerSpace
import plotly.offline as pl
import plotly.graph_objs as go
import numpy as np
from nifty import Field, \
EndomorphicOperator, \
PowerSpace
class TestEnergy(Energy):
def __init__(self, position, Op):
super(TestEnergy, self).__init__(position)
self.Op = Op
def at(self, position):
return self.__class__(position=position, Op=self.Op)
@property
def value(self):
return 0.5 * self.position.dot(self.Op(self.position))
@property
def gradient(self):
return self.Op(self.position)
@property
def curvature(self):
curv = CurvOp(self.Op)
return curv
class CurvOp(InvertibleOperatorMixin, EndomorphicOperator):
def __init__(self, Op,inverter=None, preconditioner=None):
self.Op = Op
self._domain = Op.domain
super(CurvOp, self).__init__(inverter=inverter,
preconditioner=preconditioner)
def _times(self, x, spaces):
return self.Op(x)
if __name__ == "__main__":
distribution_strategy = 'not'
# Set up position space
s_space = RGSpace([128,128])
# s_space = HPSpace(32)
# Define harmonic transformation and associated harmonic space
fft = FFTOperator(s_space)
h_space = fft.target[0]
# Setting up power space
p_space = PowerSpace(h_space, logarithmic=False,
distribution_strategy=distribution_strategy, nbin=200)
# Choosing the prior correlation structure and defining correlation operator
pow_spec = (lambda k: (.05 / (k + 1) ** 2))
# t = Field(p_space, val=pow_spec)
t= Field.from_random("normal", domain=p_space)
lap = LaplaceOperator(p_space, logarithmic=True)
T = SmoothnessOperator(p_space,sigma=1., logarithmic=True)
test_energy = TestEnergy(t,T)
def convergence_measure(a_energy, iteration): # returns current energy
x = a_energy.value
print (x, iteration)
minimizer1 = VL_BFGS(convergence_tolerance=0,
iteration_limit=10,
callback=convergence_measure,
max_history_length=10)
def explicify(op, domain):
space = domain
d = space.dim
res = np.zeros((d, d))
for i in range(d):
x = np.zeros(d)
x[i] = 1.
f = Field(space, val=x)
res[:, i] = op(f).val
return res
A = explicify(lap.times, p_space)
B = explicify(lap.adjoint_times, p_space)
test_energy,convergence = minimizer1(test_energy)
data = test_energy.position.val.get_full_data()
pl.plot([go.Scatter(x=(p_space.kindex)[1:], y=data[1:])], filename="t.html")
tt = Field.from_random("normal", domain=t.domain)
print "adjointness"
print t.dot(lap(tt))
print tt.dot(lap.adjoint_times(t))
print "log kindex"
aa = Field(p_space, val=p_space.kindex.copy())
aa.val[0] = 1
print lap(log(aa)**2).val
print "######################"
print test_energy.position.val
\ No newline at end of file
demos/train_example.py deleted 100644 → 0
View file @ 0fd5d047
from nifty import *
import plotly.offline as pl
import plotly.graph_objs as go
from mpi4py import MPI
comm = MPI.COMM_WORLD
rank = comm.rank
np.random.seed(62)
class NonlinearResponse(LinearOperator):
def __init__(self, FFT, Instrument, function, derivative, default_spaces=None):
super(NonlinearResponse, self).__init__(default_spaces)
self._domain = FFT.target
self._target = Instrument.target
self.function = function
self.derivative = derivative
self.I = Instrument
self.FFT = FFT
def _times(self, x, spaces=None):
return self.I(self.function(self.FFT.adjoint_times(x)))
def _adjoint_times(self, x, spaces=None):
return self.FFT(self.function(self.I.adjoint_times(x)))
def derived_times(self, x, position):
position_derivative = self.derivative(self.FFT.adjoint_times(position))
return self.I(position_derivative * self.FFT.adjoint_times(x))
def derived_adjoint_times(self, x, position):
position_derivative = self.derivative(self.FFT.adjoint_times(position))
return self.FFT(position_derivative * self.I.adjoint_times(x))
@property
def domain(self):
return self._domain
@property
def target(self):
return self._target
@property
def unitary(self):
return False
def plot_parameters(m,t, t_real):
m = fft.adjoint_times(m)
x = log(t.domain[0].kindex[1:])
m_data = m.val.get_full_data().real
t_data = t.val.get_full_data().real
t_real_data = t_real.val.get_full_data().real
pl.plot([go.Scatter(y=m_data)], filename='map.html')
pl.plot([go.Scatter(x = x,y=t_data),
go.Scatter(x= x, y=t_real_data[1:])], filename="t.html")
class AdjointFFTResponse(LinearOperator):
def __init__(self, FFT, R, default_spaces=None):
super(AdjointFFTResponse, self).__init__(default_spaces)
self._domain = FFT.target
self._target = R.target
self.R = R
self.FFT = FFT
def _times(self, x, spaces=None):
return self.R(self.FFT.adjoint_times(x))
def _adjoint_times(self, x, spaces=None):
return self.FFT(self.R.adjoint_times(x))
@property
def domain(self):
return self._domain
@property
def target(self):
return self._target
@property
def unitary(self):
return False
if __name__ == "__main__":
distribution_strategy = 'not'
full_data = np.genfromtxt("train_data.csv", delimiter = ',')
d = full_data.T[2]
d[0] = 0.
d -= d.mean()
d[0] = 0.
# Set up position space
s_space = RGSpace([len(d)])
# s_space = HPSpace(32)
d = Field(s_space, val=d)
# Define harmonic transformation and associated harmonic space
fft = FFTOperator(s_space)
h_space = fft.target[0]
p_space = PowerSpace(h_space, logarithmic=False,
distribution_strategy=distribution_strategy)#, nbin=50)
# Choosing the measurement instrument
# Instrument = SmoothingOperator(s_space, sigma=0.01)
Instrument = DiagonalOperator(s_space, diagonal=1.)
# Instrument._diagonal.val[200:400, 200:400] = 0
#Adding a harmonic transformation to the instrument
R = AdjointFFTResponse(fft, Instrument)
noise = .1
N = DiagonalOperator(s_space, diagonal=noise, bare=True)
d_data = d.val.get_full_data().real
if rank == 0:
pl.plot([go.Scatter(y=d_data)], filename='data.html')
# Choosing the minimization strategy
def convergence_measure(a_energy, iteration): # returns current energy
x = a_energy.value
print (x, iteration)
# minimizer1 = SteepestDescent(convergence_tolerance=0,
# iteration_limit=50,
# callback=convergence_measure)
minimizer1 = RelaxedNewton(convergence_tolerance=0,
convergence_level=1,
iteration_limit=2,
callback=convergence_measure)
minimizer2 = RelaxedNewton(convergence_tolerance=0,
convergence_level=1,
iteration_limit=30,
callback=convergence_measure)
minimizer3 = VL_BFGS(convergence_tolerance=0,
iteration_limit=50,
callback=convergence_measure,
max_history_length=10)
# Setting starting position
flat_power = Field(p_space,val=10e-8)
m0 = flat_power.power_synthesize(real_signal=True)
t0 = Field(p_space, val=log(1./(1+p_space.kindex)**2))
t0 = Field(p_space,val=-13.)
# t0 = log(sp.copy()**2)
S0 = create_power_operator(h_space, power_spectrum=exp(t0),
distribution_strategy=distribution_strategy)
data_power = log(fft(d).power_analyze(logarithmic=p_space.config["logarithmic"],
nbin=p_space.config["nbin"]))
for i in range(500):
S0 = create_power_operator(h_space, power_spectrum=exp(t0),
distribution_strategy=distribution_strategy)
# Initializing the nonlinear Wiener Filter energy
map_energy = WienerFilterEnergy(position=m0, d=d, R=R, N=N, S=S0)
# Minimization with chosen minimizer
map_energy = map_energy.analytic_solution()
# Updating parameters for correlation structure reconstruction
m0 = map_energy.position
D0 = map_energy.curvature
# Initializing the power energy with updated parameters
power_energy = CriticalPowerEnergy(position=t0, m=m0, D=D0, sigma=100000., samples=20)
(power_energy, convergence) = minimizer3(power_energy)
# Setting new power spectrum
t0.val = power_energy.position.val.real
# t0.val[-1] = t0.val[-2]
# Plotting current estimate
plot_parameters(m0, t0, data_power)
# Transforming fields to position space for plotting
ss = fft.adjoint_times(sh)
m = fft.adjoint_times(map_energy.position)
# Plotting
d_data = d.val.get_full_data().real
if rank == 0:
pl.plot([go.Heatmap(z=d_data)], filename='data.html')
tt_data = power_energy.position.val.get_full_data().real
t_data = log(sp**2).val.get_full_data().real
if rank == 0:
pl.plot([go.Scatter(y=t_data),go.Scatter(y=tt_data)], filename="t.html")
ss_data = ss.val.get_full_data().real
if rank == 0:
pl.plot([go.Heatmap(z=ss_data)], filename='ss.html')
sh_data = sh.val.get_full_data().real
if rank == 0:
pl.plot([go.Heatmap(z=sh_data)], filename='sh.html')
m_data = m.val.get_full_data().real
if rank == 0:
pl.plot([go.Heatmap(z=m_data)], filename='map.html')
f_m_data = function(m).val.get_full_data().real
if rank == 0:
pl.plot([go.Heatmap(z=f_m_data)], filename='f_map.html')
f_ss_data = function(ss).val.get_full_data().real
if rank == 0:
pl.plot([go.Heatmap(z=f_ss_data)], filename='f_ss.html')
demos/wiener_filter_advanced.py
View file @ 89c555e1
......@@ -42,7 +42,7 @@ if __name__ == "__main__":
distribution_strategy = 'not'
# Set up position space
s_space = RGSpace([512,512])
s_space = RGSpace([128,128])
# s_space = HPSpace(32)
# Define harmonic transformation and associated harmonic space
......@@ -108,9 +108,29 @@ if __name__ == "__main__":
# Initializing the Wiener Filter energy
energy = WienerFilterEnergy(position=m0, d=d, R=R, N=N, S=S)
# Solving the problem analytically
solution = energy.analytic_solution()
sample_variance = Field(sh.power_analyze(logarithmic=False).domain,val=0. + 0j)
sample_mean = Field(sh.domain,val=0. + 0j)
n_samples = 200
for i in range(n_samples):
sample = sugar.generate_posterior_sample(solution.position,solution.curvature)
sample_variance += sample.power_analyze(logarithmic=False)
sample_mean += sample
sample_variance = sample_variance/n_samples - (sample_mean/n_samples).power_analyze(logarithmic=False)
D = HarmonicPropagatorOperator(S=S, N=N, R=Instrument)
class DiagonalProber(DiagonalProberMixin, Prober):
pass
diagProber = DiagonalProber(domain=sh.domain)
diagProber(D)
diag = diagProber.diagonal
pr_diag = diag.power_analyze(logarithmic=False)**0.5
# Solving the problem with chosen minimization strategy
(energy, convergence) = minimizer(energy)
......
nifty/library/energy_library/wiener_filter_energy.py
View file @ 89c555e1
from nifty.energies.energy import Energy
from nifty.energies.memoization import memo
from nifty.library.operator_library import WienerFilterCurvature
class WienerFilterEnergy(Energy):
......@@ -33,16 +34,17 @@ class WienerFilterEnergy(Energy):
@property
def value(self):
residual = self._residual()
energy = 0.5 * self.position.dot(self.S.inverse_times(self.position))
energy += 0.5 * (self.d - self.R(self.position)).dot(
self.N.inverse_times(self.d - self.R(self.position)))
energy += 0.5 * (residual).dot(self.N.inverse_times(residual))
return energy.real
@property
def gradient(self):
residual = self._residual()
gradient = self.S.inverse_times(self.position)
gradient -= self.R.adjoint_times(
self.N.inverse_times(self.d - self.R(self.position)))
self.N.inverse_times(residual))
return gradient
@property
......@@ -50,6 +52,12 @@ class WienerFilterEnergy(Energy):
curvature = WienerFilterCurvature(R=self.R, N=self.N, S=self.S)
return curvature
@memo
def _residual(self):
residual = self.d - self.R(self.position)
return residual
def analytic_solution(self):
D_inverse = self.curvature
j = self.R.adjoint_times(self.N.inverse_times(self.d))
......
Supports Markdown
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