1# Copyright 2016 The TensorFlow Authors. All Rights Reserved. 2# 3# Licensed under the Apache License, Version 2.0 (the "License"); 4# you may not use this file except in compliance with the License. 5# You may obtain a copy of the License at 6# 7# http://www.apache.org/licenses/LICENSE-2.0 8# 9# Unless required by applicable law or agreed to in writing, software 10# distributed under the License is distributed on an "AS IS" BASIS, 11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12# See the License for the specific language governing permissions and 13# limitations under the License. 14# ============================================================================== 15"""`LinearOperator` that wraps a [batch] matrix.""" 16 17from __future__ import absolute_import 18from __future__ import division 19from __future__ import print_function 20 21from tensorflow.python.framework import dtypes 22from tensorflow.python.framework import ops 23from tensorflow.python.ops import array_ops 24from tensorflow.python.ops import math_ops 25from tensorflow.python.ops.linalg import linear_operator 26from tensorflow.python.ops.linalg import linear_operator_util 27from tensorflow.python.util.tf_export import tf_export 28 29__all__ = ["LinearOperatorFullMatrix"] 30 31 32@tf_export("linalg.LinearOperatorFullMatrix") 33class LinearOperatorFullMatrix(linear_operator.LinearOperator): 34 """`LinearOperator` that wraps a [batch] matrix. 35 36 This operator wraps a [batch] matrix `A` (which is a `Tensor`) with shape 37 `[B1,...,Bb, M, N]` for some `b >= 0`. The first `b` indices index a 38 batch member. For every batch index `(i1,...,ib)`, `A[i1,...,ib, : :]` is 39 an `M x N` matrix. 40 41 ```python 42 # Create a 2 x 2 linear operator. 43 matrix = [[1., 2.], [3., 4.]] 44 operator = LinearOperatorFullMatrix(matrix) 45 46 operator.to_dense() 47 ==> [[1., 2.] 48 [3., 4.]] 49 50 operator.shape 51 ==> [2, 2] 52 53 operator.log_abs_determinant() 54 ==> scalar Tensor 55 56 x = ... Shape [2, 4] Tensor 57 operator.matmul(x) 58 ==> Shape [2, 4] Tensor 59 60 # Create a [2, 3] batch of 4 x 4 linear operators. 61 matrix = tf.random.normal(shape=[2, 3, 4, 4]) 62 operator = LinearOperatorFullMatrix(matrix) 63 ``` 64 65 #### Shape compatibility 66 67 This operator acts on [batch] matrix with compatible shape. 68 `x` is a batch matrix with compatible shape for `matmul` and `solve` if 69 70 ``` 71 operator.shape = [B1,...,Bb] + [M, N], with b >= 0 72 x.shape = [B1,...,Bb] + [N, R], with R >= 0. 73 ``` 74 75 #### Performance 76 77 `LinearOperatorFullMatrix` has exactly the same performance as would be 78 achieved by using standard `TensorFlow` matrix ops. Intelligent choices are 79 made based on the following initialization hints. 80 81 * If `dtype` is real, and `is_self_adjoint` and `is_positive_definite`, a 82 Cholesky factorization is used for the determinant and solve. 83 84 In all cases, suppose `operator` is a `LinearOperatorFullMatrix` of shape 85 `[M, N]`, and `x.shape = [N, R]`. Then 86 87 * `operator.matmul(x)` is `O(M * N * R)`. 88 * If `M=N`, `operator.solve(x)` is `O(N^3 * R)`. 89 * If `M=N`, `operator.determinant()` is `O(N^3)`. 90 91 If instead `operator` and `x` have shape `[B1,...,Bb, M, N]` and 92 `[B1,...,Bb, N, R]`, every operation increases in complexity by `B1*...*Bb`. 93 94 #### Matrix property hints 95 96 This `LinearOperator` is initialized with boolean flags of the form `is_X`, 97 for `X = non_singular, self_adjoint, positive_definite, square`. 98 These have the following meaning: 99 100 * If `is_X == True`, callers should expect the operator to have the 101 property `X`. This is a promise that should be fulfilled, but is *not* a 102 runtime assert. For example, finite floating point precision may result 103 in these promises being violated. 104 * If `is_X == False`, callers should expect the operator to not have `X`. 105 * If `is_X == None` (the default), callers should have no expectation either 106 way. 107 """ 108 109 def __init__(self, 110 matrix, 111 is_non_singular=None, 112 is_self_adjoint=None, 113 is_positive_definite=None, 114 is_square=None, 115 name="LinearOperatorFullMatrix"): 116 r"""Initialize a `LinearOperatorFullMatrix`. 117 118 Args: 119 matrix: Shape `[B1,...,Bb, M, N]` with `b >= 0`, `M, N >= 0`. 120 Allowed dtypes: `float16`, `float32`, `float64`, `complex64`, 121 `complex128`. 122 is_non_singular: Expect that this operator is non-singular. 123 is_self_adjoint: Expect that this operator is equal to its hermitian 124 transpose. 125 is_positive_definite: Expect that this operator is positive definite, 126 meaning the quadratic form `x^H A x` has positive real part for all 127 nonzero `x`. Note that we do not require the operator to be 128 self-adjoint to be positive-definite. See: 129 https://en.wikipedia.org/wiki/Positive-definite_matrix#Extension_for_non-symmetric_matrices 130 is_square: Expect that this operator acts like square [batch] matrices. 131 name: A name for this `LinearOperator`. 132 133 Raises: 134 TypeError: If `diag.dtype` is not an allowed type. 135 """ 136 parameters = dict( 137 matrix=matrix, 138 is_non_singular=is_non_singular, 139 is_self_adjoint=is_self_adjoint, 140 is_positive_definite=is_positive_definite, 141 is_square=is_square, 142 name=name 143 ) 144 145 with ops.name_scope(name, values=[matrix]): 146 self._matrix = linear_operator_util.convert_nonref_to_tensor( 147 matrix, name="matrix") 148 self._check_matrix(self._matrix) 149 150 super(LinearOperatorFullMatrix, self).__init__( 151 dtype=self._matrix.dtype, 152 is_non_singular=is_non_singular, 153 is_self_adjoint=is_self_adjoint, 154 is_positive_definite=is_positive_definite, 155 is_square=is_square, 156 parameters=parameters, 157 name=name) 158 # TODO(b/143910018) Remove graph_parents in V3. 159 self._set_graph_parents([self._matrix]) 160 161 def _check_matrix(self, matrix): 162 """Static check of the `matrix` argument.""" 163 allowed_dtypes = [ 164 dtypes.float16, 165 dtypes.float32, 166 dtypes.float64, 167 dtypes.complex64, 168 dtypes.complex128, 169 ] 170 171 matrix = ops.convert_to_tensor_v2_with_dispatch(matrix, name="matrix") 172 173 dtype = matrix.dtype 174 if dtype not in allowed_dtypes: 175 raise TypeError( 176 "Argument matrix must have dtype in %s. Found: %s" 177 % (allowed_dtypes, dtype)) 178 179 if matrix.shape.ndims is not None and matrix.shape.ndims < 2: 180 raise ValueError( 181 "Argument matrix must have at least 2 dimensions. Found: %s" 182 % matrix) 183 184 def _shape(self): 185 return self._matrix.shape 186 187 def _shape_tensor(self): 188 return array_ops.shape(self._matrix) 189 190 def _matmul(self, x, adjoint=False, adjoint_arg=False): 191 return math_ops.matmul( 192 self._matrix, x, adjoint_a=adjoint, adjoint_b=adjoint_arg) 193 194 def _solve(self, rhs, adjoint=False, adjoint_arg=False): 195 return self._dense_solve(rhs, adjoint=adjoint, adjoint_arg=adjoint_arg) 196 197 def _to_dense(self): 198 return self._matrix 199