forked from tensorflow/tensorflow
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathlinalg_grad.py
More file actions
134 lines (119 loc) · 4.88 KB
/
Copy pathlinalg_grad.py
File metadata and controls
134 lines (119 loc) · 4.88 KB
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
43
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
96
97
98
99
100
101
102
103
104
105
106
107
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
# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Gradients for operators defined in linalg_ops.py.
Useful reference for derivative formulas is
An extended collection of matrix derivative results for forward and reverse
mode algorithmic differentiation by Mike Giles:
http://eprints.maths.ox.ac.uk/1079/1/NA-08-01.pdf
A detailed derivation of formulas for backpropagating through spectral layers
(SVD and Eig) by Ionescu, Vantzos & Sminchisescu:
https://arxiv.org/pdf/1509.07838v4.pdf
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from tensorflow.python.framework import ops
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops import linalg_ops
from tensorflow.python.ops import math_ops
@ops.RegisterGradient("MatrixInverse")
def _MatrixInverseGrad(op, grad):
"""Gradient for MatrixInverse."""
ainv = op.outputs[0]
return -math_ops.batch_matmul(
ainv, math_ops.batch_matmul(
grad, ainv, adj_y=True), adj_x=True)
@ops.RegisterGradient("MatrixDeterminant")
def _MatrixDeterminantGrad(op, grad):
"""Gradient for MatrixDeterminant."""
a = op.inputs[0]
c = op.outputs[0]
a_adj_inv = linalg_ops.matrix_inverse(a, adjoint=True)
multipliers = array_ops.reshape(
grad * c, array_ops.concat(0, [array_ops.shape(c), [1, 1]]))
return multipliers * a_adj_inv
@ops.RegisterGradient("Cholesky")
def _CholeskyGrad(op, grad):
"""Gradient for Cholesky."""
return linalg_ops.cholesky_grad(op.outputs[0], grad)
@ops.RegisterGradient("MatrixSolve")
def _MatrixSolveGrad(op, grad):
"""Gradient for MatrixSolve."""
a = op.inputs[0]
adjoint_a = op.get_attr("adjoint")
c = op.outputs[0]
grad_b = linalg_ops.matrix_solve(a, grad, adjoint=not adjoint_a)
if adjoint_a:
grad_a = -math_ops.batch_matmul(c, grad_b, adj_y=True)
else:
grad_a = -math_ops.batch_matmul(grad_b, c, adj_y=True)
return (grad_a, grad_b)
@ops.RegisterGradient("MatrixTriangularSolve")
def _MatrixTriangularSolveGrad(op, grad):
"""Gradient for MatrixTriangularSolve."""
a = op.inputs[0]
adjoint_a = op.get_attr("adjoint")
lower_a = op.get_attr("lower")
c = op.outputs[0]
grad_b = linalg_ops.matrix_triangular_solve(
a, grad, lower=lower_a, adjoint=not adjoint_a)
if adjoint_a:
grad_a = -math_ops.batch_matmul(c, grad_b, adj_y=True)
else:
grad_a = -math_ops.batch_matmul(grad_b, c, adj_y=True)
if lower_a:
grad_a = array_ops.matrix_band_part(grad_a, -1, 0)
else:
grad_a = array_ops.matrix_band_part(grad_a, 0, -1)
return (grad_a, grad_b)
@ops.RegisterGradient("SelfAdjointEigV2")
def _SelfAdjointEigV2Grad(op, grad_e, grad_v):
"""Gradient for SelfAdjointEigV2."""
e = op.outputs[0]
v = op.outputs[1]
# a = op.inputs[0], which satisfies
# a[...,:,:] * v[...,:,i] = e[...,i] * v[...,i]
with ops.control_dependencies([grad_e.op, grad_v.op]):
if grad_v is not None:
# Construct the matrix f(i,j) = (i != j ? 1 / (e_i - e_j) : 0).
# Notice that because of the term involving f, the gradient becomes
# infinite (or NaN in practice) when eigenvalues are not unique.
# Mathematically this should not be surprising, since for (k-fold)
# degenerate eigenvalues, the corresponding eigenvectors are only defined
# up to arbitrary rotation in a (k-dimensional) subspace.
f = array_ops.matrix_set_diag(
math_ops.inv(
array_ops.expand_dims(e, -2) - array_ops.expand_dims(e, -1)),
array_ops.zeros_like(e))
grad_a = math_ops.batch_matmul(
v,
math_ops.batch_matmul(
array_ops.matrix_diag(grad_e) + f * math_ops.batch_matmul(
v, grad_v, adj_x=True),
v,
adj_y=True))
else:
grad_a = math_ops.batch_matmul(
v,
math_ops.batch_matmul(
array_ops.matrix_diag(grad_e), v, adj_y=True))
# The forward op only depends on the lower triangular part of a, so here we
# symmetrize and take the lower triangle
grad_a = array_ops.matrix_band_part(
grad_a + array_ops.matrix_transpose(grad_a), -1, 0)
grad_a = array_ops.matrix_set_diag(grad_a, 0.5 *
array_ops.matrix_diag_part(grad_a))
return grad_a