Commit b10937f2 authored by Philipp Arras's avatar Philipp Arras
Browse files

Merge branch 'NIFTy_5' into domain_tuple_ops

parents c25acc23 3b338acb
......@@ -73,14 +73,15 @@ if __name__ == '__main__':
# Minimize the Hamiltonian
H = ift.Hamiltonian(likelihood, ic_sampling)
H = ift.EnergyAdapter(position, H)
H = ift.EnergyAdapter(position, H, want_metric=True)
# minimizer = ift.L_BFGS(ic_newton)
H, convergence = minimizer(H)
reconstruction = sky(H.position)
ift.plot(reconstruction, title='reconstruction')
ift.plot(GR.adjoint_times(data), title='data')
ift.plot(sky(mock_position), title='truth')
ift.plot_finish(nx=3, xsize=16, ysize=5, title="results",
name="bernoulli.png")
plot = ift.Plot()
plot.add(reconstruction, title='reconstruction')
plot.add(GR.adjoint_times(data), title='data')
plot.add(sky(mock_position), title='truth')
plot.output(nx=3, xsize=16, ysize=5, title="results",
name="bernoulli.png")
......@@ -103,18 +103,17 @@ if __name__ == '__main__':
# PLOTTING
rg = isinstance(position_space, ift.RGSpace)
plot = ift.Plot()
if rg and len(position_space.shape) == 1:
ift.plot([HT(MOCK_SIGNAL), GR.adjoint(data), HT(m)],
plot.add([HT(MOCK_SIGNAL), GR.adjoint(data), HT(m)],
label=['Mock signal', 'Data', 'Reconstruction'],
alpha=[1, .3, 1])
ift.plot(mask_to_nan(mask, HT(m-MOCK_SIGNAL)), title='Residuals')
ift.plot_finish(nx=2, ny=1, xsize=10, ysize=4,
title="getting_started_1")
plot.add(mask_to_nan(mask, HT(m-MOCK_SIGNAL)), title='Residuals')
plot.output(nx=2, ny=1, xsize=10, ysize=4, title="getting_started_1")
else:
ift.plot(HT(MOCK_SIGNAL), title='Mock Signal')
ift.plot(mask_to_nan(mask, (GR(Mask)).adjoint(data)),
plot.add(HT(MOCK_SIGNAL), title='Mock Signal')
plot.add(mask_to_nan(mask, (GR(Mask)).adjoint(data)),
title='Data')
ift.plot(HT(m), title='Reconstruction')
ift.plot(mask_to_nan(mask, HT(m-MOCK_SIGNAL)), title='Residuals')
ift.plot_finish(nx=2, ny=2, xsize=10, ysize=10,
title="getting_started_1")
plot.add(HT(m), title='Reconstruction')
plot.add(mask_to_nan(mask, HT(m-MOCK_SIGNAL)), title='Residuals')
plot.output(nx=2, ny=2, xsize=10, ysize=10, title="getting_started_1")
......@@ -93,14 +93,15 @@ if __name__ == '__main__':
# Minimize the Hamiltonian
H = ift.Hamiltonian(likelihood)
H = ift.EnergyAdapter(position, H)
H = ift.EnergyAdapter(position, H, want_metric=True)
H, convergence = minimizer(H)
# Plot results
signal = sky(mock_position)
reconst = sky(H.position)
ift.plot(signal, title='Signal')
ift.plot(GR.adjoint(data), title='Data')
ift.plot(reconst, title='Reconstruction')
ift.plot(reconst - signal, title='Residuals')
ift.plot_finish(name='getting_started_2.png', xsize=16, ysize=16)
plot = ift.Plot()
plot.add(signal, title='Signal')
plot.add(GR.adjoint(data), title='Data')
plot.add(reconst, title='Reconstruction')
plot.add(reconst - signal, title='Residuals')
plot.output(name='getting_started_2.png', xsize=16, ysize=16)
......@@ -41,11 +41,12 @@ if __name__ == '__main__':
power_space = A.target[0]
power_distributor = ift.PowerDistributor(harmonic_space, power_space)
dummy = ift.Field.from_random('normal', harmonic_space)
domain = ift.MultiDomain.union(
(A.domain, ift.MultiDomain.make({'xi': harmonic_space})))
domain = ift.MultiDomain.union((A.domain,
ift.MultiDomain.make({
'xi': harmonic_space
})))
correlated_field = ht(
power_distributor(A)*ift.FieldAdapter(domain, "xi"))
correlated_field = ht(power_distributor(A)*ift.FieldAdapter(domain, "xi"))
# alternatively to the block above one can do:
# correlated_field = ift.CorrelatedField(position_space, A)
......@@ -54,8 +55,7 @@ if __name__ == '__main__':
# Building the Line of Sight response
LOS_starts, LOS_ends = get_random_LOS(100)
R = ift.LOSResponse(position_space, starts=LOS_starts,
ends=LOS_ends)
R = ift.LOSResponse(position_space, starts=LOS_starts, ends=LOS_ends)
# build signal response model and model likelihood
signal_response = R(signal)
# specify noise
......@@ -68,8 +68,7 @@ if __name__ == '__main__':
data = signal_response(MOCK_POSITION) + N.draw_sample()
# set up model likelihood
likelihood = ift.GaussianEnergy(
mean=data, covariance=N)(signal_response)
likelihood = ift.GaussianEnergy(mean=data, covariance=N)(signal_response)
# set up minimization and inversion schemes
ic_sampling = ift.GradientNormController(iteration_limit=100)
......@@ -83,34 +82,33 @@ if __name__ == '__main__':
INITIAL_POSITION = ift.from_random('normal', domain)
position = INITIAL_POSITION
ift.plot(signal(MOCK_POSITION), title='ground truth')
ift.plot(R.adjoint_times(data), title='data')
ift.plot([A(MOCK_POSITION)], title='power')
ift.plot_finish(nx=3, xsize=16, ysize=5, title="setup", name="setup.png")
plot = ift.Plot()
plot.add(signal(MOCK_POSITION), title='Ground Truth')
plot.add(R.adjoint_times(data), title='Data')
plot.add([A(MOCK_POSITION)], title='Power Spectrum')
plot.output(ny=1, nx=3, xsize=24, ysize=6, name="setup.png")
# number of samples used to estimate the KL
N_samples = 20
for i in range(2):
metric = H(ift.Linearization.make_var(position)).metric
samples = [metric.draw_sample(from_inverse=True)
for _ in range(N_samples)]
KL = ift.SampledKullbachLeiblerDivergence(H, samples)
KL = ift.EnergyAdapter(position, KL)
KL = ift.KL_Energy(position, H, N_samples, want_metric=True)
KL, convergence = minimizer(KL)
position = KL.position
ift.plot(signal(position), title="reconstruction")
ift.plot([A(position), A(MOCK_POSITION)], title="power")
ift.plot_finish(nx=2, xsize=12, ysize=6, title="loop", name="loop.png")
plot = ift.Plot()
plot.add(signal(KL.position), title="reconstruction")
plot.add([A(KL.position), A(MOCK_POSITION)], title="power")
plot.output(ny=1, ysize=6, xsize=16, name="loop.png")
plot = ift.Plot()
sc = ift.StatCalculator()
for sample in samples:
sc.add(signal(sample+position))
ift.plot(sc.mean, title="mean")
ift.plot(ift.sqrt(sc.var), title="std deviation")
powers = [A(s+position) for s in samples]
ift.plot([A(position), A(MOCK_POSITION)]+powers, title="power")
ift.plot_finish(nx=3, xsize=16, ysize=5, title="results",
name="results.png")
for sample in KL.samples:
sc.add(signal(sample+KL.position))
plot.add(sc.mean, title="Posterior Mean")
plot.add(ift.sqrt(sc.var), title="Posterior Standard Deviation")
powers = [A(s+KL.position) for s in KL.samples]
plot.add(
[A(KL.position), A(MOCK_POSITION)]+powers,
title="Sampled Posterior Power Spectrum")
plot.output(ny=1, nx=3, xsize=24, ysize=6, name="results.png")
......@@ -20,21 +20,24 @@ def plot_test():
# Start various plotting tests
ift.plot(field_rg1_1, title='Single plot')
ift.plot_finish()
ift.plot(field_rg2, title='2d rg')
ift.plot([field_rg1_1, field_rg1_2], title='list 1d rg', label=['1', '2'])
ift.plot(field_rg1_2, title='1d rg, xmin, ymin', xmin=0.5, ymin=0.,
plot = ift.Plot()
plot.add(field_rg1_1, title='Single plot')
plot.output()
plot = ift.Plot()
plot.add(field_rg2, title='2d rg')
plot.add([field_rg1_1, field_rg1_2], title='list 1d rg', label=['1', '2'])
plot.add(field_rg1_2, title='1d rg, xmin, ymin', xmin=0.5, ymin=0.,
xlabel='xmin=0.5', ylabel='ymin=0')
ift.plot_finish(title='Three plots')
ift.plot(field_hp, title='HP planck-color', colormap='Planck-like')
ift.plot(field_rg1_2, title='1d rg')
ift.plot(field_ps)
ift.plot(field_gl, title='GL')
ift.plot(field_rg2, title='2d rg')
ift.plot_finish(nx=2, ny=3, title='Five plots')
plot.output(title='Three plots')
plot = ift.Plot()
plot.add(field_hp, title='HP planck-color', colormap='Planck-like')
plot.add(field_rg1_2, title='1d rg')
plot.add(field_ps)
plot.add(field_gl, title='GL')
plot.add(field_rg2, title='2d rg')
plot.output(nx=2, ny=3, title='Five plots')
if __name__ == '__main__':
......
......@@ -86,15 +86,16 @@ N = ift.DiagonalOperator(ift.from_global_data(d_space, var))
IC = ift.GradientNormController(tol_abs_gradnorm=1e-8)
likelihood = ift.GaussianEnergy(d, N)(R)
H = ift.Hamiltonian(likelihood, IC)
H = ift.EnergyAdapter(params, H, IC)
Ham = ift.Hamiltonian(likelihood, IC)
H = ift.EnergyAdapter(params, Ham, want_metric=True)
# Minimize
minimizer = ift.NewtonCG(IC)
H, _ = minimizer(H)
# Draw posterior samples
samples = [H.metric.draw_sample(from_inverse=True) + H.position
metric = Ham(ift.Linearization.make_var(H.position, want_metric=True)).metric
samples = [metric.draw_sample(from_inverse=True) + H.position
for _ in range(N_samples)]
# Plotting
......
......@@ -22,7 +22,8 @@ from .operators.operator import Operator
from .operators.central_zero_padder import CentralZeroPadder
from .operators.diagonal_operator import DiagonalOperator
from .operators.distributors import DOFDistributor, PowerDistributor
from .operators.domain_tuple_operators import DomainDistributor, DomainTupleFieldInserter
from .operators.domain_tuple_operators import DomainTupleFieldInserter
from .operators.contraction_operator import ContractionOperator
from .operators.endomorphic_operator import EndomorphicOperator
from .operators.exp_transform import ExpTransform
from .operators.harmonic_operators import (
......@@ -67,9 +68,10 @@ from .minimization.energy import Energy
from .minimization.quadratic_energy import QuadraticEnergy
from .minimization.line_energy import LineEnergy
from .minimization.energy_adapter import EnergyAdapter
from .minimization.kl_energy import KL_Energy
from .sugar import *
from .plotting.plot import plot, plot_finish
from .plotting.plot import Plot
from .library.amplitude_model import AmplitudeModel
from .library.inverse_gamma_model import InverseGammaModel
......
......@@ -41,7 +41,7 @@ def _get_acceptable_location(op, loc, lin):
for i in range(50):
try:
loc2 = loc+dir
lin2 = op(Linearization.make_var(loc2))
lin2 = op(Linearization.make_var(loc2, lin.want_metric))
if np.isfinite(lin2.val.sum()) and abs(lin2.val.sum()) < 1e20:
break
except FloatingPointError:
......@@ -54,14 +54,14 @@ def _get_acceptable_location(op, loc, lin):
def _check_consistency(op, loc, tol, ntries, do_metric):
for _ in range(ntries):
lin = op(Linearization.make_var(loc))
lin = op(Linearization.make_var(loc, do_metric))
loc2, lin2 = _get_acceptable_location(op, loc, lin)
dir = loc2-loc
locnext = loc2
dirnorm = dir.norm()
for i in range(50):
locmid = loc + 0.5*dir
linmid = op(Linearization.make_var(locmid))
linmid = op(Linearization.make_var(locmid, do_metric))
dirder = linmid.jac(dir)
numgrad = (lin2.val-lin.val)
xtol = tol * dirder.norm() / np.sqrt(dirder.size)
......
......@@ -21,8 +21,8 @@ from __future__ import absolute_import, division, print_function
from ..compat import *
from ..domain_tuple import DomainTuple
from ..multi_domain import MultiDomain
from ..operators.contraction_operator import ContractionOperator
from ..operators.distributors import PowerDistributor
from ..operators.domain_tuple_operators import DomainDistributor
from ..operators.harmonic_operators import HarmonicTransformOperator
from ..operators.simple_linear_operators import FieldAdapter
from ..sugar import exp
......@@ -65,8 +65,8 @@ def MfCorrelatedField(s_space_spatial, s_space_energy, amplitude_model_spatial,
pd_energy = PowerDistributor(pd_spatial.domain, p_space_energy, 1)
pd = pd_spatial(pd_energy)
dom_distr_spatial = DomainDistributor(pd.domain, 0)
dom_distr_energy = DomainDistributor(pd.domain, 1)
dom_distr_spatial = ContractionOperator(pd.domain, 0).adjoint
dom_distr_energy = ContractionOperator(pd.domain, 1).adjoint
a_spatial = dom_distr_spatial(amplitude_model_spatial)
a_energy = dom_distr_energy(amplitude_model_energy)
......
......@@ -53,7 +53,7 @@ class InverseGammaModel(Operator):
outer = 1/outer_inv
jac = makeOp(Field.from_local_data(self._domain, inner*outer))
jac = jac(x.jac)
return Linearization(points, jac)
return x.new(points, jac)
@staticmethod
def IG(field, alpha, q):
......
......@@ -9,13 +9,17 @@ from .sugar import makeOp
class Linearization(object):
def __init__(self, val, jac, metric=None):
def __init__(self, val, jac, metric=None, want_metric=False):
self._val = val
self._jac = jac
if self._val.domain != self._jac.target:
raise ValueError("domain mismatch")
self._want_metric = want_metric
self._metric = metric
def new(self, val, jac, metric=None):
return Linearization(val, jac, metric, self._want_metric)
@property
def domain(self):
return self._jac.domain
......@@ -37,6 +41,10 @@ class Linearization(object):
"""Only available if target is a scalar"""
return self._jac.adjoint_times(Field.scalar(1.))
@property
def want_metric(self):
return self._want_metric
@property
def metric(self):
"""Only available if target is a scalar"""
......@@ -44,35 +52,34 @@ class Linearization(object):
def __getitem__(self, name):
from .operators.simple_linear_operators import FieldAdapter
return Linearization(self._val[name], FieldAdapter(self.domain, name))
return self.new(self._val[name], FieldAdapter(self.domain, name))
def __neg__(self):
return Linearization(
-self._val, -self._jac,
None if self._metric is None else -self._metric)
return self.new(-self._val, -self._jac,
None if self._metric is None else -self._metric)
def conjugate(self):
return Linearization(
return self.new(
self._val.conjugate(), self._jac.conjugate(),
None if self._metric is None else self._metric.conjugate())
@property
def real(self):
return Linearization(self._val.real, self._jac.real)
return self.new(self._val.real, self._jac.real)
def _myadd(self, other, neg):
if isinstance(other, Linearization):
met = None
if self._metric is not None and other._metric is not None:
met = self._metric._myadd(other._metric, neg)
return Linearization(
return self.new(
self._val.flexible_addsub(other._val, neg),
self._jac._myadd(other._jac, neg), met)
if isinstance(other, (int, float, complex, Field, MultiField)):
if neg:
return Linearization(self._val-other, self._jac, self._metric)
return self.new(self._val-other, self._jac, self._metric)
else:
return Linearization(self._val+other, self._jac, self._metric)
return self.new(self._val+other, self._jac, self._metric)
def __add__(self, other):
return self._myadd(other, False)
......@@ -91,7 +98,7 @@ class Linearization(object):
if isinstance(other, Linearization):
if self.target != other.target:
raise ValueError("domain mismatch")
return Linearization(
return self.new(
self._val*other._val,
(makeOp(other._val)(self._jac))._myadd(
makeOp(self._val)(other._jac), False))
......@@ -99,11 +106,11 @@ class Linearization(object):
if other == 1:
return self
met = None if self._metric is None else self._metric.scale(other)
return Linearization(self._val*other, self._jac.scale(other), met)
return self.new(self._val*other, self._jac.scale(other), met)
if isinstance(other, (Field, MultiField)):
if self.target != other.domain:
raise ValueError("domain mismatch")
return Linearization(self._val*other, makeOp(other)(self._jac))
return self.new(self._val*other, makeOp(other)(self._jac))
def __rmul__(self, other):
return self.__mul__(other)
......@@ -111,46 +118,48 @@ class Linearization(object):
def vdot(self, other):
from .operators.simple_linear_operators import VdotOperator
if isinstance(other, (Field, MultiField)):
return Linearization(
return self.new(
Field.scalar(self._val.vdot(other)),
VdotOperator(other)(self._jac))
return Linearization(
return self.new(
Field.scalar(self._val.vdot(other._val)),
VdotOperator(self._val)(other._jac) +
VdotOperator(other._val)(self._jac))
def sum(self):
from .operators.simple_linear_operators import SumReductionOperator
return Linearization(
return self.new(
Field.scalar(self._val.sum()),
SumReductionOperator(self._jac.target)(self._jac))
def exp(self):
tmp = self._val.exp()
return Linearization(tmp, makeOp(tmp)(self._jac))
return self.new(tmp, makeOp(tmp)(self._jac))
def log(self):
tmp = self._val.log()
return Linearization(tmp, makeOp(1./self._val)(self._jac))
return self.new(tmp, makeOp(1./self._val)(self._jac))
def tanh(self):
tmp = self._val.tanh()
return Linearization(tmp, makeOp(1.-tmp**2)(self._jac))
return self.new(tmp, makeOp(1.-tmp**2)(self._jac))
def positive_tanh(self):
tmp = self._val.tanh()
tmp2 = 0.5*(1.+tmp)
return Linearization(tmp2, makeOp(0.5*(1.-tmp**2))(self._jac))
return self.new(tmp2, makeOp(0.5*(1.-tmp**2))(self._jac))
def add_metric(self, metric):
return Linearization(self._val, self._jac, metric)
return self.new(self._val, self._jac, metric)
@staticmethod
def make_var(field):
def make_var(field, want_metric=False):
from .operators.scaling_operator import ScalingOperator
return Linearization(field, ScalingOperator(1., field.domain))
return Linearization(field, ScalingOperator(1., field.domain),
want_metric=want_metric)
@staticmethod
def make_const(field):
def make_const(field, want_metric=False):
from .operators.simple_linear_operators import NullOperator
return Linearization(field, NullOperator(field.domain, field.domain))
return Linearization(field, NullOperator(field.domain, field.domain),
want_metric=want_metric)
......@@ -75,7 +75,7 @@ class ConjugateGradient(Minimizer):
return energy, controller.CONVERGED
while True:
q = energy.metric(d)
q = energy.apply_metric(d)
ddotq = d.vdot(q).real
if ddotq == 0.:
logger.error("Error: ConjugateGradient: ddotq==0.")
......
......@@ -180,7 +180,7 @@ class NewtonCG(DescentMinimizer):
while True:
if abs(ri).sum() <= termcond:
return xsupi
Ap = energy.metric(psupi)
Ap = energy.apply_metric(psupi)
# check curvature
curv = psupi.vdot(Ap)
if 0 <= curv <= 3*float64eps:
......
......@@ -109,6 +109,20 @@ class Energy(NiftyMetaBase()):
"""
raise NotImplementedError
def apply_metric(self, x):
"""
Parameters
----------
x: Field/MultiField
Argument for the metric operator
Returns
-------
Field/MultiField:
Output of the metric operator
"""
raise NotImplementedError
def longest_step(self, dir):
"""Returns the longest allowed step size along `dir`
......
......@@ -8,58 +8,38 @@ from ..operators.scaling_operator import ScalingOperator
class EnergyAdapter(Energy):
def __init__(self, position, op, controller=None, preconditioner=None,
constants=[]):
def __init__(self, position, op, constants=[], want_metric=False):
super(EnergyAdapter, self).__init__(position)
self._op = op
self._val = self._grad = self._metric = None
self._controller = controller
self._preconditioner = preconditioner
self._constants = constants
def at(self, position):
return EnergyAdapter(position, self._op, self._controller,
self._preconditioner, self._constants)
def _fill_all(self):
self._want_metric = want_metric
if len(self._constants) == 0:
tmp = self._op(Linearization.make_var(self._position))
tmp = self._op(Linearization.make_var(self._position, want_metric))
else:
ops = [ScalingOperator(0. if key in self._constants else 1., dom)
for key, dom in self._position.domain.items()]
bdop = BlockDiagonalOperator(self._position.domain, tuple(ops))
tmp = self._op(Linearization(self._position, bdop))
tmp = self._op(Linearization(self._position, bdop,
want_metric=want_metric))
self._val = tmp.val.local_data[()]
self._grad = tmp.gradient
if self._controller is not None:
from ..operators.linear_operator import LinearOperator
from ..operators.inversion_enabler import InversionEnabler
self._metric = tmp._metric
if self._preconditioner is None:
precond = None
elif isinstance(self._preconditioner, LinearOperator):
precond = self._preconditioner
elif isinstance(self._preconditioner, Energy):
precond = self._preconditioner.at(self._position).metric
self._metric = InversionEnabler(tmp._metric, self._controller,
precond)
else:
self._metric = tmp._metric
def at(self, position):
return EnergyAdapter(position, self._op, self._constants,
self._want_metric)
@property
def value(self):
if self._val is None:
self._val = self._op(self._position).local_data[()]
return self._val
@property
def gradient(self):
if self._grad is None:
self._fill_all()
return self._grad
@property
def metric(self):
if self._metric is None:
self._fill_all()
return self._metric
def apply_metric(self, x):
return self._metric(x)
from __future__ import absolute_import, division, print_function
from ..compat import *
from .energy import Energy