extra.py 7.26 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program.  If not, see <http://www.gnu.org/licenses/>.
#
14
# Copyright(C) 2013-2019 Max-Planck-Society
15
#
16
# NIFTy is being developed at the Max-Planck-Institut fuer Astrophysik.
17 18

import numpy as np
Philipp Arras's avatar
Philipp Arras committed
19

20
from .domain_tuple import DomainTuple
Martin Reinecke's avatar
fix  
Martin Reinecke committed
21 22
from .field import Field
from .linearization import Linearization
23
from .multi_domain import MultiDomain
24
from .operators.linear_operator import LinearOperator
Martin Reinecke's avatar
fix  
Martin Reinecke committed
25
from .sugar import from_random
26

Martin Reinecke's avatar
Martin Reinecke committed
27
__all__ = ["consistency_check", "check_jacobian_consistency"]
28

Philipp Arras's avatar
Philipp Arras committed
29

Martin Reinecke's avatar
Martin Reinecke committed
30 31 32 33 34 35 36 37
def _assert_allclose(f1, f2, atol, rtol):
    if isinstance(f1, Field):
        return np.testing.assert_allclose(f1.local_data, f2.local_data,
                                          atol=atol, rtol=rtol)
    for key, val in f1.items():
        _assert_allclose(val, f2[key], atol=atol, rtol=rtol)


38 39
def _adjoint_implementation(op, domain_dtype, target_dtype, atol, rtol,
                            only_r_linear):
Martin Reinecke's avatar
Martin Reinecke committed
40 41 42 43 44 45 46
    needed_cap = op.TIMES | op.ADJOINT_TIMES
    if (op.capability & needed_cap) != needed_cap:
        return
    f1 = from_random("normal", op.domain, dtype=domain_dtype)
    f2 = from_random("normal", op.target, dtype=target_dtype)
    res1 = f1.vdot(op.adjoint_times(f2))
    res2 = op.times(f1).vdot(f2)
47 48
    if only_r_linear:
        res1, res2 = res1.real, res2.real
Martin Reinecke's avatar
Martin Reinecke committed
49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64
    np.testing.assert_allclose(res1, res2, atol=atol, rtol=rtol)


def _inverse_implementation(op, domain_dtype, target_dtype, atol, rtol):
    needed_cap = op.TIMES | op.INVERSE_TIMES
    if (op.capability & needed_cap) != needed_cap:
        return
    foo = from_random("normal", op.target, dtype=target_dtype)
    res = op(op.inverse_times(foo))
    _assert_allclose(res, foo, atol=atol, rtol=rtol)

    foo = from_random("normal", op.domain, dtype=domain_dtype)
    res = op.inverse_times(op(foo))
    _assert_allclose(res, foo, atol=atol, rtol=rtol)


65 66 67 68
def _full_implementation(op, domain_dtype, target_dtype, atol, rtol,
                         only_r_linear):
    _adjoint_implementation(op, domain_dtype, target_dtype, atol, rtol,
                            only_r_linear)
Martin Reinecke's avatar
Martin Reinecke committed
69 70 71
    _inverse_implementation(op, domain_dtype, target_dtype, atol, rtol)


72 73 74
def _check_linearity(op, domain_dtype, atol, rtol):
    fld1 = from_random("normal", op.domain, dtype=domain_dtype)
    fld2 = from_random("normal", op.domain, dtype=domain_dtype)
Martin Reinecke's avatar
Martin Reinecke committed
75
    alpha = np.random.random()  # FIXME: this can break badly with MPI!
76 77 78 79 80
    val1 = op(alpha*fld1+fld2)
    val2 = alpha*op(fld1)+op(fld2)
    _assert_allclose(val1, val2, atol=atol, rtol=rtol)


81 82 83
def _domain_check(op):
    for dd in [op.domain, op.target]:
        if not isinstance(dd, (DomainTuple, MultiDomain)):
Martin Reinecke's avatar
Martin Reinecke committed
84 85 86
            raise TypeError(
                'The domain and the target of an operator need to',
                'be instances of either DomainTuple or MultiDomain.')
87 88


Martin Reinecke's avatar
Martin Reinecke committed
89
def consistency_check(op, domain_dtype=np.float64, target_dtype=np.float64,
90
                      atol=0, rtol=1e-7, only_r_linear=False):
Reimar H Leike's avatar
Reimar H Leike committed
91 92 93 94
    """
    Checks an operator for algebraic consistency of its capabilities.

    Checks whether times(), adjoint_times(), inverse_times() and
Philipp Arras's avatar
Philipp Arras committed
95
    adjoint_inverse_times() (if in capability list) is implemented
Reimar H Leike's avatar
Reimar H Leike committed
96
    consistently. Additionally, it checks whether the operator is linear.
Philipp Arras's avatar
Philipp Arras committed
97 98 99 100 101

    Parameters
    ----------
    op : LinearOperator
        Operator which shall be checked.
Reimar H Leike's avatar
Reimar H Leike committed
102
    domain_dtype : dtype
Philipp Arras's avatar
Philipp Arras committed
103 104
        The data type of the random vectors in the operator's domain. Default
        is `np.float64`.
Reimar H Leike's avatar
Reimar H Leike committed
105
    target_dtype : dtype
Philipp Arras's avatar
Philipp Arras committed
106 107 108
        The data type of the random vectors in the operator's target. Default
        is `np.float64`.
    atol : float
Martin Reinecke's avatar
Martin Reinecke committed
109 110
        Absolute tolerance for the check. If rtol is specified,
        then satisfying any tolerance will let the check pass.
Reimar H Leike's avatar
Reimar H Leike committed
111
        Default: 0.
Philipp Arras's avatar
Philipp Arras committed
112
    rtol : float
Martin Reinecke's avatar
Martin Reinecke committed
113 114
        Relative tolerance for the check. If atol is specified,
        then satisfying any tolerance will let the check pass.
Reimar H Leike's avatar
Reimar H Leike committed
115
        Default: 0.
116 117 118
    only_r_linear: bool
        set to True if the operator is only R-linear, not C-linear.
        This will relax the adjointness test accordingly.
Philipp Arras's avatar
Philipp Arras committed
119
    """
120 121
    if not isinstance(op, LinearOperator):
        raise TypeError('This test tests only linear operators.')
122
    _domain_check(op)
123
    _check_linearity(op, domain_dtype, atol, rtol)
124 125 126 127 128 129
    _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)
Martin Reinecke's avatar
Martin Reinecke committed
130
    _full_implementation(op.adjoint.inverse, domain_dtype, target_dtype, atol,
131
                         rtol, only_r_linear)
Martin Reinecke's avatar
Martin Reinecke committed
132 133


Martin Reinecke's avatar
Martin Reinecke committed
134
def _get_acceptable_location(op, loc, lin):
Martin Reinecke's avatar
Martin Reinecke committed
135
    if not np.isfinite(lin.val.sum()):
Martin Reinecke's avatar
Martin Reinecke committed
136 137 138 139
        raise ValueError('Initial value must be finite')
    dir = from_random("normal", loc.domain)
    dirder = lin.jac(dir)
    if dirder.norm() == 0:
Martin Reinecke's avatar
Martin Reinecke committed
140
        dir = dir * (lin.val.norm()*1e-5)
Martin Reinecke's avatar
Martin Reinecke committed
141
    else:
Martin Reinecke's avatar
Martin Reinecke committed
142
        dir = dir * (lin.val.norm()*1e-5/dirder.norm())
Martin Reinecke's avatar
Martin Reinecke committed
143 144 145 146
    # Find a step length that leads to a "reasonable" location
    for i in range(50):
        try:
            loc2 = loc+dir
147
            lin2 = op(Linearization.make_var(loc2, lin.want_metric))
Martin Reinecke's avatar
Martin Reinecke committed
148 149 150 151 152 153 154 155 156
            if np.isfinite(lin2.val.sum()) and abs(lin2.val.sum()) < 1e20:
                break
        except FloatingPointError:
            pass
        dir = dir*0.5
    else:
        raise ValueError("could not find a reasonable initial step")
    return loc2, lin2

Martin Reinecke's avatar
Martin Reinecke committed
157

Martin Reinecke's avatar
Martin Reinecke committed
158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175
def check_jacobian_consistency(op, loc, tol=1e-8, ntries=100):
    """
    Checks the Jacobian of an operator against its finite difference
    approximation.

    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.
    """
176
    _domain_check(op)
Martin Reinecke's avatar
Martin Reinecke committed
177
    for _ in range(ntries):
178
        lin = op(Linearization.make_var(loc))
Martin Reinecke's avatar
Martin Reinecke committed
179
        loc2, lin2 = _get_acceptable_location(op, loc, lin)
Martin Reinecke's avatar
Martin Reinecke committed
180
        dir = loc2-loc
Martin Reinecke's avatar
Martin Reinecke committed
181 182 183 184
        locnext = loc2
        dirnorm = dir.norm()
        for i in range(50):
            locmid = loc + 0.5*dir
185
            linmid = op(Linearization.make_var(locmid))
Martin Reinecke's avatar
Martin Reinecke committed
186 187
            dirder = linmid.jac(dir)
            numgrad = (lin2.val-lin.val)
Martin Reinecke's avatar
Martin Reinecke committed
188
            xtol = tol * dirder.norm() / np.sqrt(dirder.size)
Martin Reinecke's avatar
Martin Reinecke committed
189
            if (abs(numgrad-dirder) <= xtol).all():
Martin Reinecke's avatar
Martin Reinecke committed
190 191 192
                break
            dir = dir*0.5
            dirnorm *= 0.5
Martin Reinecke's avatar
Martin Reinecke committed
193
            loc2, lin2 = locmid, linmid
Martin Reinecke's avatar
Martin Reinecke committed
194 195 196
        else:
            raise ValueError("gradient and value seem inconsistent")
        loc = locnext