Commit ece9c551 authored by Martin Reinecke's avatar Martin Reinecke
Browse files

Merge branch 'matrix_product_operator_mf' into 'NIFTy_6'

Extend MatrixProductOperator to be able to operate on subdomains of fields

See merge request !433
parents 193a276f 9e4b327a
Pipeline #71894 passed with stages
in 25 minutes and 23 seconds
......@@ -16,12 +16,14 @@
# NIFTy is being developed at the Max-Planck-Institut fuer Astrophysik.
from ..domain_tuple import DomainTuple
from ..multi_domain import MultiDomain
from ..domains.unstructured_domain import UnstructuredDomain
from ..field import Field
from ..multi_domain import MultiDomain
from ..multi_field import MultiField
from .endomorphic_operator import EndomorphicOperator
from .linear_operator import LinearOperator
from .endomorphic_operator import EndomorphicOperator
from .. import utilities
import numpy as np
class VdotOperator(LinearOperator):
......@@ -352,29 +354,113 @@ class PartialExtractor(LinearOperator):
class MatrixProductOperator(EndomorphicOperator):
"""Endomorphic matrix multiplication with input field.
This operator supports scipy.sparse matrices and numpy arrays
as the matrix to be applied.
For numpy array matrices, can apply the matrix over a subspace
of the input.
If the input arrays have more than one dimension, for
scipy.sparse matrices the `flatten` keyword argument must be
set to true. This means that the input field will be flattened
before applying the matrix and reshaped to its original shape
afterwards.
Matrices are tested regarding their compatibility with the
called for application method.
Flattening and subspace application are mutually exclusive.
Parameters
----------
domain: :class:`Domain` or :class:`DomainTuple`
Domain of the operator.
If :class:`DomainTuple` it is assumed to have only one entry.
matrix: scipy.sparse matrix or numpy array
Matrix of shape `(domain.shape, domain.shape)`. Needs to support
`dot()` and `transpose()` in the style of numpy arrays.
Quadratic matrix of shape `(domain.shape, domain.shape)`
(if `not flatten`) that supports `matrix.transpose()`.
If it is not a numpy array, needs to be applicable to the val
array of input fields by `matrix.dot()`.
spaces: int or tuple of int, optional
The subdomain(s) of "domain" which the operator acts on.
If None, it acts on all elements.
Only possible for numpy array matrices.
If `len(domain) > 1` and `flatten=False`, this parameter is
mandatory.
flatten: boolean, optional
Whether the input value array should be flattened before
applying the matrix and reshaped to its original shape
afterwards.
Needed for scipy.sparse matrices if `len(domain) > 1`.
"""
def __init__(self, domain, matrix):
self._domain = DomainTuple.make(domain)
shp = self._domain.shape
if len(shp) > 1:
raise TypeError('Only 1D-domain supported yet.')
if matrix.shape != (*shp, *shp):
raise ValueError
def __init__(self, domain, matrix, spaces=None, flatten=False):
self._capability = self.TIMES | self.ADJOINT_TIMES
self._domain = DomainTuple.make(domain)
mat_dim = len(matrix.shape)
if mat_dim % 2 != 0 or \
matrix.shape != (matrix.shape[:mat_dim//2] + matrix.shape[:mat_dim//2]):
raise ValueError("Matrix must be quadratic.")
appl_dim = mat_dim // 2 # matrix application space dimension
# take shortcut for trivial case
if spaces is not None:
if len(self._domain.shape) == 1 and spaces == (0, ):
spaces = None
if spaces is None:
self._spaces = None
self._active_axes = utilities.my_sum(self._domain.axes)
appl_space_shape = self._domain.shape
if flatten:
appl_space_shape = (utilities.my_product(appl_space_shape), )
else:
if flatten:
raise ValueError(
"Cannot flatten input AND apply to a subspace")
if not isinstance(matrix, np.ndarray):
raise ValueError(
"Application to subspaces only supported for numpy array matrices."
)
self._spaces = utilities.parse_spaces(spaces, len(self._domain))
appl_space_shape = []
active_axes = []
for space_idx in spaces:
appl_space_shape += self._domain[space_idx].shape
active_axes += self._domain.axes[space_idx]
appl_space_shape = tuple(appl_space_shape)
self._active_axes = tuple(active_axes)
self._mat_last_n = tuple([-appl_dim + i for i in range(appl_dim)])
self._mat_first_n = np.arange(appl_dim)
# Test if the matrix and the array it will be applied to fit
if matrix.shape[:appl_dim] != appl_space_shape:
raise ValueError(
"Matrix and domain shapes are incompatible under the requested "
+ "application scheme.\n" +
f"Matrix appl shape: {matrix.shape[:appl_dim]}, " +
f"appl_space_shape: {appl_space_shape}.")
self._mat = matrix
self._mat_tr = matrix.transpose().conjugate()
self._flatten = flatten
def apply(self, x, mode):
self._check_input(x, mode)
res = x.val
f = self._mat.dot if mode == self.TIMES else self._mat_tr.dot
res = f(res)
times = (mode == self.TIMES)
m = self._mat if times else self._mat_tr
if self._spaces is None:
if not self._flatten:
res = m.dot(x.val)
else:
res = m.dot(x.val.flatten()).reshape(self._domain.shape)
return Field(self._domain, res)
mat_axes = self._mat_last_n if times else np.flip(self._mat_last_n)
move_axes = self._mat_first_n if times else np.flip(self._mat_first_n)
res = np.tensordot(m, x.val, axes=(mat_axes, self._active_axes))
res = np.moveaxis(res, move_axes, self._active_axes)
return Field(self._domain, res)
......@@ -277,19 +277,40 @@ def testSpecialSum(sp):
op = ift.library.correlated_fields._SpecialSum(sp)
ift.extra.consistency_check(op)
@pmp('sp', [ift.RGSpace(10)])
@pmp('seed', [12, 3])
def testMatrixProductOperator(sp, seed):
def metatestMatrixProductOperator(sp, mat_shape, seed, **kwargs):
ift.random.push_sseq_from_seed(seed)
mat = ift.random.current_rng().standard_normal((*sp.shape, *sp.shape))
op = ift.MatrixProductOperator(sp, mat)
mat = ift.random.current_rng().standard_normal(mat_shape)
op = ift.MatrixProductOperator(sp, mat, **kwargs)
ift.extra.consistency_check(op)
mat = mat + 1j*ift.random.current_rng().standard_normal((*sp.shape, *sp.shape))
op = ift.MatrixProductOperator(sp, mat)
mat = mat + 1j*ift.random.current_rng().standard_normal(mat_shape)
op = ift.MatrixProductOperator(sp, mat, **kwargs)
ift.extra.consistency_check(op)
ift.random.pop_sseq()
@pmp('sp', [ift.RGSpace(10)])
@pmp('spaces', [None, (0,)])
@pmp('seed', [12, 3])
def testMatrixProductOperator_1d(sp, spaces, seed):
mat_shape = sp.shape * 2
metatestMatrixProductOperator(sp, mat_shape, seed, spaces=spaces)
@pmp('sp', [ift.DomainTuple.make((ift.RGSpace((2)), ift.RGSpace((10))))])
@pmp('spaces', [(0,), (1,), (0, 1)])
@pmp('seed', [12, 3])
def testMatrixProductOperator_2d_spaces(sp, spaces, seed):
appl_shape = []
for sp_idx in spaces:
appl_shape += sp[sp_idx].shape
appl_shape = tuple(appl_shape)
mat_shape = appl_shape * 2
metatestMatrixProductOperator(sp, mat_shape, seed, spaces=spaces)
@pmp('sp', [ift.RGSpace((2, 10))])
@pmp('seed', [12, 3])
def testMatrixProductOperator_2d_flatten(sp, seed):
appl_shape = (ift.utilities.my_product(sp.shape),)
mat_shape = appl_shape * 2
metatestMatrixProductOperator(sp, mat_shape, seed, flatten=True)
@pmp('seed', [12, 3])
def testPartialExtractor(seed):
......@@ -303,7 +324,6 @@ def testPartialExtractor(seed):
ift.extra.consistency_check(op)
ift.random.pop_sseq()
@pmp('seed', [12, 3])
def testSlowFieldAdapter(seed):
dom = {'a': ift.RGSpace(1), 'b': ift.RGSpace(2)}
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment