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` acting like a diagonal matrix.""" 16 17from __future__ import absolute_import 18from __future__ import division 19from __future__ import print_function 20 21from tensorflow.python.framework import ops 22from tensorflow.python.ops import array_ops 23from tensorflow.python.ops import check_ops 24from tensorflow.python.ops import math_ops 25from tensorflow.python.ops.linalg import linalg_impl as linalg 26from tensorflow.python.ops.linalg import linear_operator 27from tensorflow.python.ops.linalg import linear_operator_util 28from tensorflow.python.util.tf_export import tf_export 29 30__all__ = ["LinearOperatorDiag",] 31 32 33@tf_export("linalg.LinearOperatorDiag") 34class LinearOperatorDiag(linear_operator.LinearOperator): 35 """`LinearOperator` acting like a [batch] square diagonal matrix. 36 37 This operator acts like a [batch] diagonal matrix `A` with shape 38 `[B1,...,Bb, N, N]` for some `b >= 0`. The first `b` indices index a 39 batch member. For every batch index `(i1,...,ib)`, `A[i1,...,ib, : :]` is 40 an `N x N` matrix. This matrix `A` is not materialized, but for 41 purposes of broadcasting this shape will be relevant. 42 43 `LinearOperatorDiag` is initialized with a (batch) vector. 44 45 ```python 46 # Create a 2 x 2 diagonal linear operator. 47 diag = [1., -1.] 48 operator = LinearOperatorDiag(diag) 49 50 operator.to_dense() 51 ==> [[1., 0.] 52 [0., -1.]] 53 54 operator.shape 55 ==> [2, 2] 56 57 operator.log_abs_determinant() 58 ==> scalar Tensor 59 60 x = ... Shape [2, 4] Tensor 61 operator.matmul(x) 62 ==> Shape [2, 4] Tensor 63 64 # Create a [2, 3] batch of 4 x 4 linear operators. 65 diag = tf.random.normal(shape=[2, 3, 4]) 66 operator = LinearOperatorDiag(diag) 67 68 # Create a shape [2, 1, 4, 2] vector. Note that this shape is compatible 69 # since the batch dimensions, [2, 1], are broadcast to 70 # operator.batch_shape = [2, 3]. 71 y = tf.random.normal(shape=[2, 1, 4, 2]) 72 x = operator.solve(y) 73 ==> operator.matmul(x) = y 74 ``` 75 76 #### Shape compatibility 77 78 This operator acts on [batch] matrix with compatible shape. 79 `x` is a batch matrix with compatible shape for `matmul` and `solve` if 80 81 ``` 82 operator.shape = [B1,...,Bb] + [N, N], with b >= 0 83 x.shape = [C1,...,Cc] + [N, R], 84 and [C1,...,Cc] broadcasts with [B1,...,Bb] to [D1,...,Dd] 85 ``` 86 87 #### Performance 88 89 Suppose `operator` is a `LinearOperatorDiag` of shape `[N, N]`, 90 and `x.shape = [N, R]`. Then 91 92 * `operator.matmul(x)` involves `N * R` multiplications. 93 * `operator.solve(x)` involves `N` divisions and `N * R` multiplications. 94 * `operator.determinant()` involves a size `N` `reduce_prod`. 95 96 If instead `operator` and `x` have shape `[B1,...,Bb, N, N]` and 97 `[B1,...,Bb, N, R]`, every operation increases in complexity by `B1*...*Bb`. 98 99 #### Matrix property hints 100 101 This `LinearOperator` is initialized with boolean flags of the form `is_X`, 102 for `X = non_singular, self_adjoint, positive_definite, square`. 103 These have the following meaning: 104 105 * If `is_X == True`, callers should expect the operator to have the 106 property `X`. This is a promise that should be fulfilled, but is *not* a 107 runtime assert. For example, finite floating point precision may result 108 in these promises being violated. 109 * If `is_X == False`, callers should expect the operator to not have `X`. 110 * If `is_X == None` (the default), callers should have no expectation either 111 way. 112 """ 113 114 def __init__(self, 115 diag, 116 is_non_singular=None, 117 is_self_adjoint=None, 118 is_positive_definite=None, 119 is_square=None, 120 name="LinearOperatorDiag"): 121 r"""Initialize a `LinearOperatorDiag`. 122 123 Args: 124 diag: Shape `[B1,...,Bb, N]` `Tensor` with `b >= 0` `N >= 0`. 125 The diagonal of the operator. Allowed dtypes: `float16`, `float32`, 126 `float64`, `complex64`, `complex128`. 127 is_non_singular: Expect that this operator is non-singular. 128 is_self_adjoint: Expect that this operator is equal to its hermitian 129 transpose. If `diag.dtype` is real, this is auto-set to `True`. 130 is_positive_definite: Expect that this operator is positive definite, 131 meaning the quadratic form `x^H A x` has positive real part for all 132 nonzero `x`. Note that we do not require the operator to be 133 self-adjoint to be positive-definite. See: 134 https://en.wikipedia.org/wiki/Positive-definite_matrix#Extension_for_non-symmetric_matrices 135 is_square: Expect that this operator acts like square [batch] matrices. 136 name: A name for this `LinearOperator`. 137 138 Raises: 139 TypeError: If `diag.dtype` is not an allowed type. 140 ValueError: If `diag.dtype` is real, and `is_self_adjoint` is not `True`. 141 """ 142 parameters = dict( 143 diag=diag, 144 is_non_singular=is_non_singular, 145 is_self_adjoint=is_self_adjoint, 146 is_positive_definite=is_positive_definite, 147 is_square=is_square, 148 name=name 149 ) 150 151 with ops.name_scope(name, values=[diag]): 152 self._diag = linear_operator_util.convert_nonref_to_tensor( 153 diag, name="diag") 154 self._check_diag(self._diag) 155 156 # Check and auto-set hints. 157 if not self._diag.dtype.is_complex: 158 if is_self_adjoint is False: 159 raise ValueError("A real diagonal operator is always self adjoint.") 160 else: 161 is_self_adjoint = True 162 163 if is_square is False: 164 raise ValueError("Only square diagonal operators currently supported.") 165 is_square = True 166 167 super(LinearOperatorDiag, self).__init__( 168 dtype=self._diag.dtype, 169 is_non_singular=is_non_singular, 170 is_self_adjoint=is_self_adjoint, 171 is_positive_definite=is_positive_definite, 172 is_square=is_square, 173 parameters=parameters, 174 name=name) 175 # TODO(b/143910018) Remove graph_parents in V3. 176 self._set_graph_parents([self._diag]) 177 178 def _check_diag(self, diag): 179 """Static check of diag.""" 180 if diag.shape.ndims is not None and diag.shape.ndims < 1: 181 raise ValueError("Argument diag must have at least 1 dimension. " 182 "Found: %s" % diag) 183 184 def _shape(self): 185 # If d_shape = [5, 3], we return [5, 3, 3]. 186 d_shape = self._diag.shape 187 return d_shape.concatenate(d_shape[-1:]) 188 189 def _shape_tensor(self): 190 d_shape = array_ops.shape(self._diag) 191 k = d_shape[-1] 192 return array_ops.concat((d_shape, [k]), 0) 193 194 @property 195 def diag(self): 196 return self._diag 197 198 def _assert_non_singular(self): 199 return linear_operator_util.assert_no_entries_with_modulus_zero( 200 self._diag, 201 message="Singular operator: Diagonal contained zero values.") 202 203 def _assert_positive_definite(self): 204 if self.dtype.is_complex: 205 message = ( 206 "Diagonal operator had diagonal entries with non-positive real part, " 207 "thus was not positive definite.") 208 else: 209 message = ( 210 "Real diagonal operator had non-positive diagonal entries, " 211 "thus was not positive definite.") 212 213 return check_ops.assert_positive( 214 math_ops.real(self._diag), 215 message=message) 216 217 def _assert_self_adjoint(self): 218 return linear_operator_util.assert_zero_imag_part( 219 self._diag, 220 message=( 221 "This diagonal operator contained non-zero imaginary values. " 222 " Thus it was not self-adjoint.")) 223 224 def _matmul(self, x, adjoint=False, adjoint_arg=False): 225 diag_term = math_ops.conj(self._diag) if adjoint else self._diag 226 x = linalg.adjoint(x) if adjoint_arg else x 227 diag_mat = array_ops.expand_dims(diag_term, -1) 228 return diag_mat * x 229 230 def _matvec(self, x, adjoint=False): 231 diag_term = math_ops.conj(self._diag) if adjoint else self._diag 232 return diag_term * x 233 234 def _determinant(self): 235 return math_ops.reduce_prod(self._diag, axis=[-1]) 236 237 def _log_abs_determinant(self): 238 log_det = math_ops.reduce_sum( 239 math_ops.log(math_ops.abs(self._diag)), axis=[-1]) 240 if self.dtype.is_complex: 241 log_det = math_ops.cast(log_det, dtype=self.dtype) 242 return log_det 243 244 def _solve(self, rhs, adjoint=False, adjoint_arg=False): 245 diag_term = math_ops.conj(self._diag) if adjoint else self._diag 246 rhs = linalg.adjoint(rhs) if adjoint_arg else rhs 247 inv_diag_mat = array_ops.expand_dims(1. / diag_term, -1) 248 return rhs * inv_diag_mat 249 250 def _to_dense(self): 251 return array_ops.matrix_diag(self._diag) 252 253 def _diag_part(self): 254 return self.diag 255 256 def _add_to_tensor(self, x): 257 x_diag = array_ops.matrix_diag_part(x) 258 new_diag = self._diag + x_diag 259 return array_ops.matrix_set_diag(x, new_diag) 260 261 def _eigvals(self): 262 return ops.convert_to_tensor_v2_with_dispatch(self.diag) 263 264 def _cond(self): 265 abs_diag = math_ops.abs(self.diag) 266 return (math_ops.reduce_max(abs_diag, axis=-1) / 267 math_ops.reduce_min(abs_diag, axis=-1)) 268