metric_gaussian_kl_mpi.py 5.02 KB
Newer Older
Martin Reinecke's avatar
Martin Reinecke committed
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42
# 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/>.
#
# Copyright(C) 2013-2019 Max-Planck-Society
#
# NIFTy is being developed at the Max-Planck-Institut fuer Astrophysik.

from .. import utilities
from ..linearization import Linearization
from ..operators.energy_operators import StandardHamiltonian
from .energy import Energy
from mpi4py import MPI
import numpy as np
from ..field import Field
from ..multi_field import MultiField


_comm = MPI.COMM_WORLD
ntask = _comm.Get_size()
rank = _comm.Get_rank()
master = (rank == 0)


def _shareRange(nwork, nshares, myshare):
    nbase = nwork//nshares
    additional = nwork % nshares
    lo = myshare*nbase + min(myshare, additional)
    hi = lo + nbase + int(myshare < additional)
    return lo, hi


def np_allreduce_sum(arr):
Martin Reinecke's avatar
Martin Reinecke committed
43
    arr = np.array(arr)
Martin Reinecke's avatar
Martin Reinecke committed
44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95
    res = np.empty_like(arr)
    _comm.Allreduce(arr, res, MPI.SUM)
    return res


def allreduce_sum_field(fld):
    if isinstance(fld, Field):
        return Field.from_local_data(fld.domain,
                                     np_allreduce_sum(fld.local_data))
    res = tuple(
        Field.from_local_data(f.domain, np_allreduce_sum(f.local_data))
        for f in fld.values())
    return MultiField(fld.domain, res)


class MetricGaussianKL_MPI(Energy):
    def __init__(self, mean, hamiltonian, n_samples, constants=[],
                 point_estimates=[], mirror_samples=False,
                 _samples=None):
        super(MetricGaussianKL_MPI, self).__init__(mean)

        if not isinstance(hamiltonian, StandardHamiltonian):
            raise TypeError
        if hamiltonian.domain is not mean.domain:
            raise ValueError
        if not isinstance(n_samples, int):
            raise TypeError
        self._constants = list(constants)
        self._point_estimates = list(point_estimates)
        if not isinstance(mirror_samples, bool):
            raise TypeError

        self._hamiltonian = hamiltonian

        if _samples is None:
            lo, hi = _shareRange(n_samples, ntask, rank)
            met = hamiltonian(Linearization.make_partial_var(
                mean, point_estimates, True)).metric
            _samples = []
            for i in range(lo, hi):
                np.random.seed(i)
                _samples.append(met.draw_sample(from_inverse=True))
            if mirror_samples:
                _samples += [-s for s in _samples]
                n_samples *= 2
            _samples = tuple(_samples)
        self._samples = _samples
        self._n_samples = n_samples
        self._lin = Linearization.make_partial_var(mean, constants)
        v, g = None, None
        if len(self._samples) == 0:  # hack if there are too many MPI tasks
            tmp = self._hamiltonian(self._lin)
Martin Reinecke's avatar
Martin Reinecke committed
96
            v = 0. * tmp.val.local_data
Martin Reinecke's avatar
Martin Reinecke committed
97 98 99 100 101
            g = 0. * tmp.gradient
        else:
            for s in self._samples:
                tmp = self._hamiltonian(self._lin+s)
                if v is None:
Martin Reinecke's avatar
Martin Reinecke committed
102
                    v = tmp.val.local_data.copy()
Martin Reinecke's avatar
Martin Reinecke committed
103 104
                    g = tmp.gradient
                else:
Martin Reinecke's avatar
Martin Reinecke committed
105
                    v += tmp.val.local_data
Martin Reinecke's avatar
Martin Reinecke committed
106
                    g = g + tmp.gradient
Martin Reinecke's avatar
Martin Reinecke committed
107
        self._val = np_allreduce_sum(v)[()] / self._n_samples
Martin Reinecke's avatar
Martin Reinecke committed
108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149
        self._grad = allreduce_sum_field(g) / self._n_samples
        self._metric = None

    def at(self, position):
        return MetricGaussianKL_MPI(
            position, self._hamiltonian, self._n_samples, self._constants,
            self._point_estimates, _samples=self._samples)

    @property
    def value(self):
        return self._val

    @property
    def gradient(self):
        return self._grad

    def _get_metric(self):
        lin = self._lin.with_want_metric()
        if self._metric is None:
            if len(self._samples) == 0:  # hack if there are too many MPI tasks
                self._metric = self._hamiltonian(lin).metric.scale(0.)
            else:
                mymap = map(lambda v: self._hamiltonian(lin+v).metric,
                            self._samples)
                self._metric = utilities.my_sum(mymap)
                self._metric = self._metric.scale(1./self._n_samples)

    def apply_metric(self, x):
        self._get_metric()
        return allreduce_sum_field(self._metric(x))

    @property
    def metric(self):
        if ntask > 1:
            raise ValueError("not supported when MPI is active")
        return self._metric

    @property
    def samples(self):
        res = _comm.allgather(self._samples)
        res = [item for sublist in res for item in sublist]
        return res