Commit 20577a56 authored by Martin Reinecke's avatar Martin Reinecke
Browse files

Merge branch 'switch_to_ducc' of gitlab.mpcdf.mpg.de:ift/nifty into switch_to_ducc

parents 98b214c2 fce04fb8
Pipeline #77095 passed with stages
in 25 minutes and 38 seconds
...@@ -9,6 +9,21 @@ now uses the DUCC package (<https://gitlab.mpcdf.mpg.de/mtr/ducc)>, ...@@ -9,6 +9,21 @@ now uses the DUCC package (<https://gitlab.mpcdf.mpg.de/mtr/ducc)>,
which is their successor. which is their successor.
Naming of operator tests
------------------------
The implementation tests for nonlinear operators are now available in
`ift.extra.check_operator()` and for linear operators
`ift.extra.check_linear_operator()`.
MetricGaussianKL interface
--------------------------
Users do not instanciate `MetricGaussianKL` by its constructor anymore. Rather
`MetricGaussianKL.make()` shall be used.
Changes since NIFTy 5 Changes since NIFTy 5
===================== =====================
...@@ -70,6 +85,19 @@ print(met) ...@@ -70,6 +85,19 @@ print(met)
print(met.draw_sample()) print(met.draw_sample())
``` ```
New approach for sampling complex numbers
=========================================
When calling draw_sample_with_dtype with a complex dtype,
the variance is now used for the imaginary part and real part separately.
This is done in order to be consistent with the Hamiltonian.
Note that by this,
```
np.std(ift.from_random(domain, 'normal', dtype=np.complex128).val)
````
does not give 1, but sqrt(2) as a result.
MPI parallelisation over samples in MetricGaussianKL MPI parallelisation over samples in MetricGaussianKL
---------------------------------------------------- ----------------------------------------------------
...@@ -85,6 +113,7 @@ the generation of reproducible random numbers in the presence of MPI parallelism ...@@ -85,6 +113,7 @@ the generation of reproducible random numbers in the presence of MPI parallelism
and leads to cleaner code overall. Please see the documentation of and leads to cleaner code overall. Please see the documentation of
`nifty7.random` for details. `nifty7.random` for details.
Interface Change for from_random and OuterProduct Interface Change for from_random and OuterProduct
------------------------------------------------- -------------------------------------------------
......
...@@ -131,7 +131,7 @@ def main(): ...@@ -131,7 +131,7 @@ def main():
# Draw new samples to approximate the KL five times # Draw new samples to approximate the KL five times
for i in range(5): for i in range(5):
# Draw new samples and minimize KL # Draw new samples and minimize KL
KL = ift.MetricGaussianKL(mean, H, N_samples) KL = ift.MetricGaussianKL.make(mean, H, N_samples)
KL, convergence = minimizer(KL) KL, convergence = minimizer(KL)
mean = KL.position mean = KL.position
...@@ -144,7 +144,7 @@ def main(): ...@@ -144,7 +144,7 @@ def main():
name=filename.format("loop_{:02d}".format(i))) name=filename.format("loop_{:02d}".format(i)))
# Draw posterior samples # Draw posterior samples
KL = ift.MetricGaussianKL(mean, H, N_samples) KL = ift.MetricGaussianKL.make(mean, H, N_samples)
sc = ift.StatCalculator() sc = ift.StatCalculator()
for sample in KL.samples: for sample in KL.samples:
sc.add(signal(sample + KL.position)) sc.add(signal(sample + KL.position))
......
...@@ -152,10 +152,8 @@ ...@@ -152,10 +152,8 @@
"sigmas = [1.0, 0.5, 0.1]\n", "sigmas = [1.0, 0.5, 0.1]\n",
"\n", "\n",
"for i in range(3):\n", "for i in range(3):\n",
" op = ift.library.correlated_fields._LognormalMomentMatching(mean=mean,\n", " op = ift.LognormalTransform(mean=mean, sigma=sigmas[i],\n",
" sig=sigmas[i],\n", " key='foo', N_copies=0)\n",
" key='foo',\n",
" N_copies=0)\n",
" op_samples = np.array(\n", " op_samples = np.array(\n",
" [op(s).val for s in [ift.from_random(op.domain) for i in range(10000)]])\n", " [op(s).val for s in [ift.from_random(op.domain) for i in range(10000)]])\n",
"\n", "\n",
......
...@@ -131,7 +131,7 @@ def main(): ...@@ -131,7 +131,7 @@ def main():
for i in range(10): for i in range(10):
# Draw new samples and minimize KL # Draw new samples and minimize KL
KL = ift.MetricGaussianKL(mean, H, N_samples) KL = ift.MetricGaussianKL.make(mean, H, N_samples)
KL, convergence = minimizer(KL) KL, convergence = minimizer(KL)
mean = KL.position mean = KL.position
...@@ -157,7 +157,7 @@ def main(): ...@@ -157,7 +157,7 @@ def main():
name=filename.format("loop_{:02d}".format(i))) name=filename.format("loop_{:02d}".format(i)))
# Done, draw posterior samples # Done, draw posterior samples
KL = ift.MetricGaussianKL(mean, H, N_samples) KL = ift.MetricGaussianKL.make(mean, H, N_samples)
sc = ift.StatCalculator() sc = ift.StatCalculator()
scA1 = ift.StatCalculator() scA1 = ift.StatCalculator()
scA2 = ift.StatCalculator() scA2 = ift.StatCalculator()
......
...@@ -34,6 +34,7 @@ from matplotlib.colors import LogNorm ...@@ -34,6 +34,7 @@ from matplotlib.colors import LogNorm
import nifty7 as ift import nifty7 as ift
def main(): def main():
dom = ift.UnstructuredDomain(1) dom = ift.UnstructuredDomain(1)
scale = 10 scale = 10
...@@ -90,7 +91,7 @@ def main(): ...@@ -90,7 +91,7 @@ def main():
plt.figure(figsize=[12, 8]) plt.figure(figsize=[12, 8])
for ii in range(15): for ii in range(15):
if ii % 3 == 0: if ii % 3 == 0:
mgkl = ift.MetricGaussianKL(pos, ham, 40) mgkl = ift.MetricGaussianKL.make(pos, ham, 40)
plt.cla() plt.cla()
plt.imshow(z.T, origin='lower', norm=LogNorm(), vmin=1e-3, plt.imshow(z.T, origin='lower', norm=LogNorm(), vmin=1e-3,
......
...@@ -97,7 +97,7 @@ def main(): ...@@ -97,7 +97,7 @@ def main():
p_space = ift.UnstructuredDomain(N_params) p_space = ift.UnstructuredDomain(N_params)
params = ift.full(p_space, 0.) params = ift.full(p_space, 0.)
R = PolynomialResponse(p_space, x) R = PolynomialResponse(p_space, x)
ift.extra.consistency_check(R) ift.extra.check_linear_operator(R)
d_space = R.target d_space = R.target
d = ift.makeField(d_space, y) d = ift.makeField(d_space, y)
......
...@@ -52,6 +52,7 @@ from .operators.energy_operators import ( ...@@ -52,6 +52,7 @@ from .operators.energy_operators import (
BernoulliEnergy, StandardHamiltonian, AveragedEnergy, QuadraticFormOperator, BernoulliEnergy, StandardHamiltonian, AveragedEnergy, QuadraticFormOperator,
Squared2NormOperator, StudentTEnergy, VariableCovarianceGaussianEnergy) Squared2NormOperator, StudentTEnergy, VariableCovarianceGaussianEnergy)
from .operators.convolution_operators import FuncConvolutionOperator from .operators.convolution_operators import FuncConvolutionOperator
from .operators.normal_operators import NormalTransform, LognormalTransform
from .probing import probe_with_posterior_samples, probe_diagonal, \ from .probing import probe_with_posterior_samples, probe_diagonal, \
StatCalculator, approximation2endo StatCalculator, approximation2endo
......
...@@ -15,21 +15,115 @@ ...@@ -15,21 +15,115 @@
# #
# NIFTy is being developed at the Max-Planck-Institut fuer Astrophysik. # NIFTy is being developed at the Max-Planck-Institut fuer Astrophysik.
from itertools import combinations
import numpy as np import numpy as np
from numpy.testing import assert_ from numpy.testing import assert_
from . import random
from .domain_tuple import DomainTuple from .domain_tuple import DomainTuple
from .field import Field from .field import Field
from .linearization import Linearization from .linearization import Linearization
from .multi_domain import MultiDomain from .multi_domain import MultiDomain
from .multi_field import MultiField from .multi_field import MultiField
from .operators.energy_operators import EnergyOperator
from .operators.linear_operator import LinearOperator from .operators.linear_operator import LinearOperator
from .operators.operator import Operator
from .sugar import from_random from .sugar import from_random
__all__ = ["consistency_check", "check_jacobian_consistency", __all__ = ["check_linear_operator", "check_operator",
"assert_allclose"] "assert_allclose"]
def check_linear_operator(op, domain_dtype=np.float64, target_dtype=np.float64,
atol=1e-12, rtol=1e-12, only_r_linear=False):
"""
Checks an operator for algebraic consistency of its capabilities.
Checks whether times(), adjoint_times(), inverse_times() and
adjoint_inverse_times() (if in capability list) is implemented
consistently. Additionally, it checks whether the operator is linear.
Parameters
----------
op : LinearOperator
Operator which shall be checked.
domain_dtype : dtype
The data type of the random vectors in the operator's domain. Default
is `np.float64`.
target_dtype : dtype
The data type of the random vectors in the operator's target. Default
is `np.float64`.
atol : float
Absolute tolerance for the check. If rtol is specified,
then satisfying any tolerance will let the check pass.
Default: 0.
rtol : float
Relative tolerance for the check. If atol is specified,
then satisfying any tolerance will let the check pass.
Default: 0.
only_r_linear: bool
set to True if the operator is only R-linear, not C-linear.
This will relax the adjointness test accordingly.
"""
if not isinstance(op, LinearOperator):
raise TypeError('This test tests only linear operators.')
_domain_check_linear(op, domain_dtype)
_domain_check_linear(op.adjoint, target_dtype)
_domain_check_linear(op.inverse, target_dtype)
_domain_check_linear(op.adjoint.inverse, domain_dtype)
_check_linearity(op, domain_dtype, atol, rtol)
_check_linearity(op.adjoint, target_dtype, atol, rtol)
_check_linearity(op.inverse, target_dtype, atol, rtol)
_check_linearity(op.adjoint.inverse, domain_dtype, atol, rtol)
_full_implementation(op, domain_dtype, target_dtype, atol, rtol,
only_r_linear)
_full_implementation(op.adjoint, target_dtype, domain_dtype, atol, rtol,
only_r_linear)
_full_implementation(op.inverse, target_dtype, domain_dtype, atol, rtol,
only_r_linear)
_full_implementation(op.adjoint.inverse, domain_dtype, target_dtype, atol,
rtol, only_r_linear)
def check_operator(op, loc, tol=1e-12, ntries=100, perf_check=True,
only_r_differentiable=True, metric_sampling=True):
"""
Performs various checks of the implementation of linear and nonlinear
operators.
Computes the Jacobian with finite differences and compares it to the
implemented Jacobian.
Parameters
----------
op : Operator
Operator which shall be checked.
loc : Field or MultiField
An Field or MultiField instance which has the same domain
as op. The location at which the gradient is checked
tol : float
Tolerance for the check.
perf_check : Boolean
Do performance check. May be disabled for very unimportant operators.
only_r_differentiable : Boolean
Jacobians of C-differentiable operators need to be C-linear.
Default: True
metric_sampling: Boolean
If op is an EnergyOperator, metric_sampling determines whether the
test shall try to sample from the metric or not.
"""
if not isinstance(op, Operator):
raise TypeError('This test tests only linear operators.')
_domain_check_nonlinear(op, loc)
_performance_check(op, loc, bool(perf_check))
_linearization_value_consistency(op, loc)
_jac_vs_finite_differences(op, loc, np.sqrt(tol), ntries,
only_r_differentiable)
_check_nontrivial_constant(op, loc, tol, ntries, only_r_differentiable,
metric_sampling)
def assert_allclose(f1, f2, atol, rtol): def assert_allclose(f1, f2, atol, rtol):
if isinstance(f1, Field): if isinstance(f1, Field):
return np.testing.assert_allclose(f1.val, f2.val, atol=atol, rtol=rtol) return np.testing.assert_allclose(f1.val, f2.val, atol=atol, rtol=rtol)
...@@ -37,6 +131,27 @@ def assert_allclose(f1, f2, atol, rtol): ...@@ -37,6 +131,27 @@ def assert_allclose(f1, f2, atol, rtol):
assert_allclose(val, f2[key], atol=atol, rtol=rtol) assert_allclose(val, f2[key], atol=atol, rtol=rtol)
def assert_equal(f1, f2):
if isinstance(f1, Field):
return np.testing.assert_equal(f1.val, f2.val)
for key, val in f1.items():
assert_equal(val, f2[key])
def _nozero(fld):
if isinstance(fld, Field):
return np.testing.assert_((fld != 0).s_all())
for val in fld.values():
_nozero(val)
def _allzero(fld):
if isinstance(fld, Field):
return np.testing.assert_((fld == 0.).s_all())
for val in fld.values():
_allzero(val)
def _adjoint_implementation(op, domain_dtype, target_dtype, atol, rtol, def _adjoint_implementation(op, domain_dtype, target_dtype, atol, rtol,
only_r_linear): only_r_linear):
needed_cap = op.TIMES | op.ADJOINT_TIMES needed_cap = op.TIMES | op.ADJOINT_TIMES
...@@ -83,7 +198,8 @@ def _check_linearity(op, domain_dtype, atol, rtol): ...@@ -83,7 +198,8 @@ def _check_linearity(op, domain_dtype, atol, rtol):
assert_allclose(val1, val2, atol=atol, rtol=rtol) assert_allclose(val1, val2, atol=atol, rtol=rtol)
def _actual_domain_check_linear(op, domain_dtype=None, inp=None): def _domain_check_linear(op, domain_dtype=None, inp=None):
_domain_check(op)
needed_cap = op.TIMES needed_cap = op.TIMES
if (op.capability & needed_cap) != needed_cap: if (op.capability & needed_cap) != needed_cap:
return return
...@@ -95,8 +211,9 @@ def _actual_domain_check_linear(op, domain_dtype=None, inp=None): ...@@ -95,8 +211,9 @@ def _actual_domain_check_linear(op, domain_dtype=None, inp=None):
assert_(op(inp).domain is op.target) assert_(op(inp).domain is op.target)
def _actual_domain_check_nonlinear(op, loc): def _domain_check_nonlinear(op, loc):
assert isinstance(loc, (Field, MultiField)) _domain_check(op)
assert_(isinstance(loc, (Field, MultiField)))
assert_(loc.domain is op.domain) assert_(loc.domain is op.domain)
for wm in [False, True]: for wm in [False, True]:
lin = Linearization.make_var(loc, wm) lin = Linearization.make_var(loc, wm)
...@@ -111,8 +228,8 @@ def _actual_domain_check_nonlinear(op, loc): ...@@ -111,8 +228,8 @@ def _actual_domain_check_nonlinear(op, loc):
assert_(reslin.jac.domain is reslin.domain) assert_(reslin.jac.domain is reslin.domain)
assert_(reslin.jac.target is reslin.target) assert_(reslin.jac.target is reslin.target)
assert_(lin.want_metric == reslin.want_metric) assert_(lin.want_metric == reslin.want_metric)
_actual_domain_check_linear(reslin.jac, inp=loc) _domain_check_linear(reslin.jac, inp=loc)
_actual_domain_check_linear(reslin.jac.adjoint, inp=reslin.jac(loc)) _domain_check_linear(reslin.jac.adjoint, inp=reslin.jac(loc))
if reslin.metric is not None: if reslin.metric is not None:
assert_(reslin.metric.domain is reslin.metric.target) assert_(reslin.metric.domain is reslin.metric.target)
assert_(reslin.metric.domain is op.domain) assert_(reslin.metric.domain is op.domain)
...@@ -164,58 +281,6 @@ def _performance_check(op, pos, raise_on_fail): ...@@ -164,58 +281,6 @@ def _performance_check(op, pos, raise_on_fail):
raise RuntimeError(s) raise RuntimeError(s)
def consistency_check(op, domain_dtype=np.float64, target_dtype=np.float64,
atol=0, rtol=1e-7, only_r_linear=False):
"""
Checks an operator for algebraic consistency of its capabilities.
Checks whether times(), adjoint_times(), inverse_times() and
adjoint_inverse_times() (if in capability list) is implemented
consistently. Additionally, it checks whether the operator is linear.
Parameters
----------
op : LinearOperator
Operator which shall be checked.
domain_dtype : dtype
The data type of the random vectors in the operator's domain. Default
is `np.float64`.
target_dtype : dtype
The data type of the random vectors in the operator's target. Default
is `np.float64`.
atol : float
Absolute tolerance for the check. If rtol is specified,
then satisfying any tolerance will let the check pass.
Default: 0.
rtol : float
Relative tolerance for the check. If atol is specified,
then satisfying any tolerance will let the check pass.
Default: 0.
only_r_linear: bool
set to True if the operator is only R-linear, not C-linear.
This will relax the adjointness test accordingly.
"""
if not isinstance(op, LinearOperator):
raise TypeError('This test tests only linear operators.')
_domain_check(op)
_actual_domain_check_linear(op, domain_dtype)
_actual_domain_check_linear(op.adjoint, target_dtype)
_actual_domain_check_linear(op.inverse, target_dtype)
_actual_domain_check_linear(op.adjoint.inverse, domain_dtype)
_check_linearity(op, domain_dtype, atol, rtol)
_check_linearity(op.adjoint, target_dtype, atol, rtol)
_check_linearity(op.inverse, target_dtype, atol, rtol)
_check_linearity(op.adjoint.inverse, domain_dtype, atol, rtol)
_full_implementation(op, domain_dtype, target_dtype, atol, rtol,
only_r_linear)
_full_implementation(op.adjoint, target_dtype, domain_dtype, atol, rtol,
only_r_linear)
_full_implementation(op.inverse, target_dtype, domain_dtype, atol, rtol,
only_r_linear)
_full_implementation(op.adjoint.inverse, domain_dtype, target_dtype, atol,
rtol, only_r_linear)
def _get_acceptable_location(op, loc, lin): def _get_acceptable_location(op, loc, lin):
if not np.isfinite(lin.val.s_sum()): if not np.isfinite(lin.val.s_sum()):
raise ValueError('Initial value must be finite') raise ValueError('Initial value must be finite')
...@@ -248,34 +313,51 @@ def _linearization_value_consistency(op, loc): ...@@ -248,34 +313,51 @@ def _linearization_value_consistency(op, loc):
assert_allclose(fld0, fld1, 0, 1e-7) assert_allclose(fld0, fld1, 0, 1e-7)
def check_jacobian_consistency(op, loc, tol=1e-8, ntries=100, perf_check=True, def _check_nontrivial_constant(op, loc, tol, ntries, only_r_differentiable,
only_r_differentiable=True): metric_sampling):
""" if isinstance(op.domain, DomainTuple):
Checks the Jacobian of an operator against its finite difference return
approximation. keys = op.domain.keys()
combis = []
Computes the Jacobian with finite differences and compares it to the if len(keys) > 4:
implemented Jacobian. from .logger import logger
logger.warning('Operator domain has more than 4 keys.')
Parameters logger.warning('Check derivatives only with one constant key at a time.')
---------- combis = [[kk] for kk in keys]
op : Operator else:
Operator which shall be checked. for ll in range(1, len(keys)):
loc : Field or MultiField combis.extend(list(combinations(keys, ll)))
An Field or MultiField instance which has the same domain for cstkeys in combis:
as op. The location at which the gradient is checked varkeys = set(keys) - set(cstkeys)
tol : float cstloc = loc.extract_by_keys(cstkeys)
Tolerance for the check. varloc = loc.extract_by_keys(varkeys)
perf_check : Boolean
Do performance check. May be disabled for very unimportant operators. val0 = op(loc)
only_r_differentiable : Boolean _, op0 = op.simplify_for_constant_input(cstloc)
Jacobians of C-differentiable operators need to be C-linear. assert op0.domain is varloc.domain
Default: True val1 = op0(varloc)
""" assert_equal(val0, val1)
_domain_check(op)
_actual_domain_check_nonlinear(op, loc) lin = Linearization.make_partial_var(loc, cstkeys, want_metric=True)
_performance_check(op, loc, bool(perf_check)) lin0 = Linearization.make_var(varloc, want_metric=True)
_linearization_value_consistency(op, loc) oplin0 = op0(lin0)
oplin = op(lin)
assert oplin.jac.target is oplin0.jac.target
rndinp = from_random(oplin.jac.target)
assert_allclose(oplin.jac.adjoint(rndinp).extract(varloc.domain),
oplin0.jac.adjoint(rndinp), 1e-13, 1e-13)
foo = oplin.jac.adjoint(rndinp).extract(cstloc.domain)
assert_equal(foo, 0*foo)
if isinstance(op, EnergyOperator) and metric_sampling:
oplin.metric.draw_sample()
# _jac_vs_finite_differences(op0, varloc, np.sqrt(tol), ntries,
# only_r_differentiable)
def _jac_vs_finite_differences(op, loc, tol, ntries, only_r_differentiable):
for _ in range(ntries): for _ in range(ntries):
lin = op(Linearization.make_var(loc)) lin = op(Linearization.make_var(loc))
loc2, lin2 = _get_acceptable_location(op, loc, lin) loc2, lin2 = _get_acceptable_location(op, loc, lin)
...@@ -300,8 +382,7 @@ def check_jacobian_consistency(op, loc, tol=1e-8, ntries=100, perf_check=True, ...@@ -300,8 +382,7 @@ def check_jacobian_consistency(op, loc, tol=1e-8, ntries=100, perf_check=True,
print(hist) print(hist)
raise ValueError("gradient and value seem inconsistent") raise ValueError("gradient and value seem inconsistent")
loc = locnext loc = locnext
check_linear_operator(linmid.jac, domain_dtype=loc.dtype,
ddtype = loc.values()[0].dtype if isinstance(loc, MultiField) else loc.dtype target_dtype=dirder.dtype,
tdtype = dirder.values()[0].dtype if isinstance(dirder, MultiField) else dirder.dtype only_r_linear=only_r_differentiable,
consistency_check(linmid.jac, domain_dtype=ddtype, target_dtype=tdtype, atol=tol**2, rtol=tol**2)
only_r_linear=only_r_differentiable)
...@@ -136,6 +136,8 @@ class Field(Operator): ...@@ -136,6 +136,8 @@ class Field(Operator):
The domain of the output random Field. The domain of the output random Field.
dtype : type dtype : type
The datatype of the output random Field. The datatype of the output random Field.
If the datatype is complex, each real and imaginary part
have variance 1
Returns Returns
------- -------
......
...@@ -37,48 +37,11 @@ from ..operators.harmonic_operators import HarmonicTransformOperator ...@@ -37,48 +37,11 @@ from ..operators.harmonic_operators import HarmonicTransformOperator
from ..operators.linear_operator import LinearOperator from ..operators.linear_operator import LinearOperator
from ..operators.operator import Operator from ..operators.operator import Operator
from ..operators.simple_linear_operators import ducktape from ..operators.simple_linear_operators import ducktape
from ..operators.normal_operators import NormalTransform, LognormalTransform
from ..probing import StatCalculator from ..probing import StatCalculator
from ..sugar import full, makeDomain, makeField, makeOp from ..sugar import full, makeDomain, makeField, makeOp
def _reshaper(x, N):
x = np.asfarray(x)
if x.shape in [(), (1,)]:
return np.full(N, x) if N != 0 else x.reshape(())
elif x.shape == (N,):
return x
else:
raise TypeError("Shape of parameters cannot be interpreted")
def _lognormal_moments(mean, sig, N=0):
if N == 0:
mean, sig = np.asfarray(mean), np.asfarray(sig)
else:
mean, sig = (_reshaper(param, N) for param in (mean, sig))
if not np.all(mean > 0):
raise ValueError("mean must be greater 0; got {!r}".format(mean))
if not np.all(sig > 0):
raise ValueError("sig must be greater 0; got {!r}".format(sig))
logsig = np.sqrt(np.log1p((sig/mean)**2))
logmean = np.log(mean) - logsig**2/2