Merge branch 'nifty6_einsum' into 'NIFTy_6'

Introduce a numpy.einsum wrapper with derivatives

See merge request !460

nifty6/__init__.py

@@ -25,6 +25,7 @@ from .operators.adder import Adder
 from .operators.diagonal_operator import DiagonalOperator
 from .operators.distributors import DOFDistributor, PowerDistributor
 from .operators.domain_tuple_field_inserter import DomainTupleFieldInserter
+from .operators.einsum import LinearEinsum, MultiLinearEinsum
 from .operators.contraction_operator import ContractionOperator
 from .operators.linear_interpolation import LinearInterpolator
 from .operators.endomorphic_operator import EndomorphicOperator

nifty6/operators/einsum.py

+# 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
+# Authors: Gordian Edenhofer, Philipp Frank
+#
+# NIFTy is being developed at the Max-Planck-Institut fuer Astrophysik.
+
+import numpy as np
+import string
+from ..domain_tuple import DomainTuple
+from ..linearization import Linearization
+from ..field import Field
+from ..multi_domain import MultiDomain
+from ..multi_field import MultiField
+from .operator import Operator
+from .linear_operator import LinearOperator
+
+
+class MultiLinearEinsum(Operator):
+    """Multi-linear Einsum operator with corresponding derivates
+
+    Parameters
+    ----------
+    domain : MultiDomain or dict{name: DomainTuple}
+        The operator's input domain.
+    subscripts : str
+        The subscripts which is passed to einsum.
+    key_order: tuple of str, optional
+        The order of the keys in the multi-field. If not specified, defaults to
+        the order of the keys in the multi-field.
+    static_mf: MultiField or dict{name: Field}, optional
+        A dictionary like type from which Fields are to be taken if the key from
+        `key_order` is not part of the `domain`. Fields in this object are
+        supposed to be static as they will not appear as FieldAdapter in the
+        Linearization.
+    optimize: bool, String or List, optional
+        Parameter passed on to einsum_path.
+
+    Notes
+    -----
+    By convention :class:`MultiLinearEinsum` only performs operations with
+    lower indices. Therefore no complex conjugation is performed on complex
+    inputs. To achieve operations with upper/lower indices use
+    :class:`PartialConjugate` before applying this operator.
+    """
+    def __init__(self, domain, subscripts,
+                 key_order=None, static_mf=None, optimize='optimal'):
+        self._domain = MultiDomain.make(domain)
+        if key_order is None:
+            self._key_order = tuple(self._domain.keys())
+        else:
+            self._key_order = key_order
+        if static_mf is not None and key_order is None:
+            ve = "`key_order` mus be specified if additional fields are munged"
+            raise ValueError(ve)
+        self._stat_mf = static_mf
+        iss, oss, *rest = subscripts.split("->")
+        iss_spl = iss.split(",")
+        len_consist = len(self._key_order) == len(iss_spl)
+        sscr_consist = all(o in iss for o in oss)
+        if rest or not sscr_consist or "," in oss or not len_consist:
+            raise ValueError(f"invalid subscripts specified; got {subscripts}")
+        ve = f"invalid order of keys {self._key_order} for subscripts {subscripts}"
+        shapes, numpy_subscripts, subscriptmap = {}, '', {}
+        alphabet = list(string.ascii_lowercase)[::-1]
+        for k, ss in zip(self._key_order, iss_spl):
+            dom = self._domain[k] if k in self._domain.keys(
+            ) else self._stat_mf[k].domain
+            if len(dom) != len(ss):
+                raise ValueError(ve)
+            for i, a in enumerate(list(ss)):
+                if a not in subscriptmap.keys():
+                    subscriptmap[a] = [alphabet.pop() for _ in
+                                       range(len(dom[i].shape))]
+                numpy_subscripts += ''.join(subscriptmap[a])
+            numpy_subscripts += ','
+            shapes[k] = dom.shape
+        numpy_subscripts = numpy_subscripts[:-1] + '->'
+        dom_sscr = dict(zip(self._key_order, iss_spl))
+        tgt = []
+        for o in oss:
+            k_hit = tuple(k for k, sscr in dom_sscr.items() if o in sscr)[0]
+            dom_k_idx = dom_sscr[k_hit].index(o)
+            if k_hit in self._domain.keys():
+                tgt += [self._domain[k_hit][dom_k_idx]]
+            else:
+                if k_hit not in self._stat_mf.keys():
+                    ve = f"{k_hit} is not in domain nor in static_mf"
+                    raise ValueError(ve)
+                tgt += [self._stat_mf[k_hit].domain[dom_k_idx]]
+            numpy_subscripts += ''.join(subscriptmap[o])
+        self._target = DomainTuple.make(tgt)
+
+        numpy_iss, numpy_oss, *_ = numpy_subscripts.split("->")
+        numpy_iss_spl = numpy_iss.split(",")
+
+        self._sscr_endswith = dict()
+        self._linpaths = dict()
+        for k, (i, ss) in zip(self._key_order, enumerate(numpy_iss_spl)):
+            left_ss_spl = (*numpy_iss_spl[:i], *numpy_iss_spl[i + 1:], ss)
+            linpath = '->'.join((','.join(left_ss_spl), numpy_oss))
+
+            plc = tuple(np.broadcast_to(np.nan, shapes[q]) for q in shapes if q != k)
+            plc += (np.broadcast_to(np.nan, shapes[k]),)
+            self._sscr_endswith[k] = linpath
+            self._linpaths[k] = np.einsum_path(linpath, *plc, optimize=optimize)[0]
+        if isinstance(optimize, list):
+            path = optimize
+        else:
+            plc = (np.broadcast_to(np.nan, shapes[k]) for k in shapes)
+            path = np.einsum_path(numpy_subscripts, *plc, optimize=optimize)[0]
+        self._sscr = numpy_subscripts
+        self._ein_kw = {"optimize": path}
+
+    def apply(self, x):
+        self._check_input(x)
+        if isinstance(x, Linearization):
+            val = x.val.val
+        else:
+            val = x.val
+        v = (
+            val[k] if k in val else self._stat_mf[k].val
+            for k in self._key_order
+        )
+        res = np.einsum(self._sscr, *v, **self._ein_kw)
+
+        if isinstance(x, Linearization):
+            jac = None
+            for wrt in self.domain.keys():
+                plc = {
+                    k: x.val[k] if k in x.val else self._stat_mf[k]
+                    for k in self._key_order if k != wrt
+                }
+                mf_wo_k = MultiField.from_dict(plc)
+                ss = self._sscr_endswith[wrt]
+                # Use the fact that the insertion order in a dictionary is the
+                # ordering of keys as to pass on `key_order`
+                jac_k = LinearEinsum(
+                    self.domain[wrt],
+                    mf_wo_k,
+                    ss,
+                    key_order=tuple(plc.keys()),
+                    optimize=self._linpaths[wrt],
+                    _target=self._target,
+                    _calling_as_lin=True
+                ).ducktape(wrt)
+                jac = jac + jac_k if jac is not None else jac_k
+            return x.new(Field.from_raw(self.target, res), jac)
+        return Field.from_raw(self.target, res)
+
+
+class LinearEinsum(LinearOperator):
+    """Linear Einsum operator with exactly one freely varying field
+
+    Parameters
+    ----------
+    domain : Domain, DomainTuple or tuple of Domain
+        The operator's input domain.
+    mf : MultiField
+        The first part of the left-hand side of the einsum.
+    subscripts : str
+        The subscripts which is passed to einsum. Everything before the very
+        last scripts before the '->' is treated as part of the fixed mulfi-
+        field while the last scripts are taken to correspond to the freely
+        varying field.
+    key_order: tuple of str, optional
+        The order of the keys in the multi-field. If not specified, defaults to
+        the order of the keys in the multi-field.
+    optimize: bool, String or List, optional
+        Parameter passed on to einsum_path.
+
+    Notes
+    -----
+    By convention :class:`LinearEinsum` only performs operations with
+    lower indices. Therefore no complex conjugation is performed on complex
+    inputs or mf. To achieve operations with upper/lower indices use
+    :class:`PartialConjugate` before applying this operator.
+    """
+    def __init__(self, domain, mf, subscripts, key_order=None, optimize='optimal',
+                 _target=None, _calling_as_lin=False):
+        self._domain = DomainTuple.make(domain)
+        if _calling_as_lin:
+            self._init_wo_preproc(mf, subscripts, key_order, optimize, _target)
+        else:
+            self._mf = mf
+            if key_order is None:
+                _key_order = tuple(self._mf.domain.keys())
+            else:
+                _key_order = key_order
+            self._ein_kw = {"optimize": optimize}
+            iss, oss, *rest = subscripts.split("->")
+            iss_spl = iss.split(",")
+            sscr_consist = all(o in iss for o in oss)
+            len_consist = len(_key_order) == len(iss_spl[:-1])
+            if rest or not sscr_consist or "," in oss or not len_consist:
+                raise ValueError(f"invalid subscripts specified; got {subscripts}")
+            ve = f"invalid order of keys {_key_order} for subscripts {subscripts}"
+            shapes, numpy_subscripts, subscriptmap = (), '', {}
+            alphabet = list(string.ascii_lowercase)
+            for k, ss in zip(_key_order, iss_spl[:-1]):
+                dom = self._mf[k].domain
+                if len(dom) != len(ss):
+                    raise ValueError(ve)
+                for i, a in enumerate(list(ss)):
+                    if a not in subscriptmap.keys():
+                        subscriptmap[a] = [alphabet.pop() for _ in
+                                           range(len(dom[i].shape))]
+                    numpy_subscripts += ''.join(subscriptmap[a])
+                numpy_subscripts += ','
+                shapes += (dom.shape,)
+            if len(self._domain) != len(iss_spl[-1]):
+                raise ValueError(ve)
+            for i, a in enumerate(list(iss_spl[-1])):
+                if a not in subscriptmap.keys():
+                    subscriptmap[a] = [alphabet.pop() for _ in
+                                       range(len(self._domain[i].shape))]
+                numpy_subscripts += ''.join(subscriptmap[a])
+            shapes += (self._domain.shape,)
+            numpy_subscripts += '->'
+
+            dom_sscr = dict(zip(_key_order, iss_spl[:-1]))
+            dom_sscr[id(self)] = iss_spl[-1]
+            tgt = []
+            for o in oss:
+                k_hit = tuple(k for k, sscr in dom_sscr.items() if o in sscr)[0]
+                dom_k_idx = dom_sscr[k_hit].index(o)
+                if k_hit in _key_order:
+                    tgt += [self._mf.domain[k_hit][dom_k_idx]]
+                else:
+                    assert k_hit == id(self)
+                    tgt += [self._domain[dom_k_idx]]
+                numpy_subscripts += "".join(subscriptmap[o])
+            _target = DomainTuple.make(tgt)
+            self._sscr = numpy_subscripts
+
+            if isinstance(optimize, list):
+                path = optimize
+            else:
+                plc = (np.broadcast_to(np.nan, shp) for shp in shapes)
+                path = np.einsum_path(numpy_subscripts, *plc, optimize=optimize)[0]
+            self._init_wo_preproc(mf, numpy_subscripts, _key_order, path, _target)
+
+
+    def _init_wo_preproc(self, mf, subscripts, keyorder, optimize, target):
+        self._ein_kw = {"optimize": optimize}
+        self._mf = mf
+        self._sscr = subscripts
+        self._key_order = keyorder
+        self._target = target
+        iss, oss, *_ = subscripts.split("->")
+        iss_spl = iss.split(",")
+
+        adj_iss = ",".join((",".join(iss_spl[:-1]), oss))
+        self._adj_sscr = "->".join((adj_iss, iss_spl[-1]))
+        self._capability = self.TIMES | self.ADJOINT_TIMES
+
+    def apply(self, x, mode):
+        self._check_input(x, mode)
+        if mode == self.TIMES:
+            dom, ss, mf = self.target, self._sscr, self._mf
+        else:
+            dom, ss, mf = self.domain, self._adj_sscr, self._mf.conjugate()
+        res = np.einsum(
+            ss, *(mf[k].val for k in self._key_order), x.val,
+            **self._ein_kw
+        )
+        return Field.from_raw(dom, res)

nifty6/operators/partial_conjugate.py

+# 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
+# Authors: Philipp Frank
+#
+# NIFTy is being developed at the Max-Planck-Institut fuer Astrophysik.
+
+from .endomorphic_operator import EndomorphicOperator
+from ..multi_domain import MultiDomain
+from ..multi_field import MultiField
+
+class PartialConjugate(EndomorphicOperator):
+    """Perform partial conjugation of a :class:`MultiField`
+
+    Parameters
+    ----------
+    domain : MultiDomain
+        The operator's input domain and output target
+    conjugation_keys : iterable of string
+        The keys of the :class:`MultiField` for which complex conjugation
+        should be performed.
+    """
+    def __init__(self, domain, conjugation_keys):
+        if not isinstance(domain, MultiDomain):
+            raise ValueError("MultiDomain expected!")
+        indom = (key in domain.keys() for key in conjugation_keys)
+        if sum(indom) != len(conjugation_keys):
+            raise ValueError("conjugation_keys not in domain!")
+        self._domain = domain
+        self._conjugation_keys = conjugation_keys
+        self._capabilities = self._all_ops
+
+    def apply(self, x, mode):
+        self._check_input(x, mode)
+        x = x.to_dict()
+        for k in self._conjugation_keys:
+            x[k] = x[k].conjugate()
+        return MultiField.from_dict(x, self._domain)

test/test_operators/test_einsum.py

+# 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
+# Authors: Gordian Edenhofer, Philipp Frank
+#
+# NIFTy is being developed at the Max-Planck-Institut fuer Astrophysik.
+
+import pytest
+from numpy.testing import assert_allclose
+import numpy as np
+from nifty6.extra import check_jacobian_consistency, consistency_check
+
+import nifty6 as ift
+from ..common import list2fixture, setup_function, teardown_function
+
+spaces = (ift.UnstructuredDomain(4),
+          ift.RGSpace((3,2)),
+          ift.LMSpace(5),
+          ift.GLSpace(4))
+
+space1 = list2fixture(spaces)
+space2 = list2fixture(spaces)
+dtype = list2fixture([np.float64, np.complex128])
+
+
+def test_linear_einsum_outer(space1, space2, dtype, n_invocations=10):
+    mf_dom = ift.MultiDomain.make(
+        {
+            "dom01": space1,
+            "dom02":
+                ift.DomainTuple.make(
+                    (space1, space2)
+                )
+        }
+    )
+    mf = ift.from_random("normal", mf_dom, dtype=dtype)
+    ss = "i,ij,j->ij"
+    key_order = ("dom01", "dom02")
+    le = ift.LinearEinsum(space2, mf, ss, key_order=key_order)
+    assert consistency_check(le, domain_dtype=dtype,target_dtype=dtype) is None
+
+    le_ift = ift.DiagonalOperator(
+        mf["dom01"], domain=mf_dom["dom02"], spaces=0
+    ) @ ift.DiagonalOperator(mf["dom02"]) @ ift.OuterProduct(
+        ift.full(mf_dom["dom01"], 1.), ift.DomainTuple.make(mf_dom["dom02"][1])
+    )
+
+    for _ in range(n_invocations):
+        r = ift.from_random("normal", le.domain, dtype=dtype)
+        assert_allclose(le(r).val, le_ift(r).val)
+        r_adj = ift.from_random("normal", le.target, dtype=dtype)
+        assert_allclose(le.adjoint(r_adj).val, le_ift.adjoint(r_adj).val)
+
+
+def test_linear_einsum_contraction(space1, space2, dtype, n_invocations=10):
+    mf_dom = ift.MultiDomain.make(
+        {
+            "dom01": space1,
+            "dom02":
+                ift.DomainTuple.make(
+                    (space1, space2)
+                )
+        }
+    )
+    mf = ift.from_random("normal", mf_dom, dtype=dtype)
+    ss = "i,ij,j->i"
+    key_order = ("dom01", "dom02")
+    le = ift.LinearEinsum(space2, mf, ss, key_order=key_order)
+    assert consistency_check(le, domain_dtype=dtype,target_dtype=dtype) is None
+
+    le_ift = ift.ContractionOperator(mf_dom["dom02"], 1) @ ift.DiagonalOperator(
+        mf["dom01"], domain=mf_dom["dom02"], spaces=0
+    ) @ ift.DiagonalOperator(mf["dom02"]) @ ift.OuterProduct(
+        ift.full(mf_dom["dom01"], 1.), ift.DomainTuple.make(mf_dom["dom02"][1])
+    )
+
+    for _ in range(n_invocations):
+        r = ift.from_random("normal", le.domain, dtype=dtype)
+        assert_allclose(le(r).val, le_ift(r).val)
+        r_adj = ift.from_random("normal", le.target, dtype=dtype)
+        assert_allclose(le.adjoint(r_adj).val, le_ift.adjoint(r_adj).val)
+
+
+def test_multi_linear_einsum_outer(
+    space1, space2, dtype, n_invocations=10, ntries=100
+):
+    mf_dom = ift.MultiDomain.make(
+        {
+            "dom01": space1,
+            "dom02":ift.DomainTuple.make((space1, space2)),
+            "dom03": space2
+        }
+    )
+    ss = "i,ij,j->ij"
+    key_order = ("dom01", "dom02", "dom03")
+    mle = ift.MultiLinearEinsum(mf_dom, ss, key_order=key_order)
+    check_jacobian_consistency(
+        mle, ift.from_random("normal", mle.domain, dtype=dtype), ntries=ntries
+    )
+
+    outer_i = ift.OuterProduct(
+        ift.full(mf_dom["dom03"], 1.), ift.DomainTuple.make(mf_dom["dom02"][0])
+    )
+    outer_j = ift.OuterProduct(
+        ift.full(mf_dom["dom01"], 1.), ift.DomainTuple.make(mf_dom["dom02"][1])
+    )
+    # SwitchSpacesOperator is equivalent to LinearEinsum with "ij->ji"
+    mle_ift = ift.SwitchSpacesOperator(
+        outer_i.target, 1
+    ) @ outer_i @ ift.FieldAdapter(mf_dom["dom01"], "dom01") * ift.FieldAdapter(
+        mf_dom["dom02"], "dom02"
+    ) * (outer_j @ ift.FieldAdapter(mf_dom["dom03"], "dom03"))
+
+    for _ in range(n_invocations):
+        rl = ift.Linearization.make_var(
+                ift.from_random("normal", mle.domain, dtype=dtype))
+        mle_rl, mle_ift_rl = mle(rl), mle_ift(rl)
+        assert_allclose(mle_rl.val.val, mle_ift_rl.val.val)
+        assert_allclose(mle_rl.jac(rl.val).val, mle_ift_rl.jac(rl.val).val)
+
+        rj_adj = ift.from_random("normal", mle_rl.jac.target, dtype=dtype)
+        mle_j_val = mle_rl.jac.adjoint(rj_adj).val
+        mle_ift_j_val = mle_ift_rl.jac.adjoint(rj_adj).val
+        for k in mle_ift.domain.keys():
+            assert_allclose(mle_j_val[k], mle_ift_j_val[k])