tensorflow-riemopt 0.1.2__tar.gz → 0.3.0__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {tensorflow-riemopt-0.1.2/tensorflow_riemopt.egg-info → tensorflow_riemopt-0.3.0}/PKG-INFO +20 -6
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/README.md +2 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/setup.py +7 -7
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/cholesky.py +1 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/euclidean.py +1 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/grassmannian.py +1 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/hyperboloid.py +2 -1
- tensorflow_riemopt-0.3.0/tensorflow_riemopt/manifolds/mainfold_test.py +59 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/manifold.py +1 -1
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/poincare.py +10 -9
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/product.py +1 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/special_orthogonal.py +1 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/symmetric_positive.py +1 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/test_invariants.py +1 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/optimizers/constrained_rmsprop.py +15 -15
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/optimizers/constrained_rmsprop_test.py +8 -6
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/optimizers/riemannian_adam.py +16 -16
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/optimizers/riemannian_adam_test.py +4 -2
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/optimizers/riemannian_gradient_descent.py +22 -17
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/optimizers/riemannian_gradient_descent_test.py +6 -4
- tensorflow_riemopt-0.3.0/tensorflow_riemopt/optimizers/test_optimizers_manifold.py +47 -0
- tensorflow_riemopt-0.3.0/tensorflow_riemopt/variable.py +20 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0/tensorflow_riemopt.egg-info}/PKG-INFO +20 -6
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt.egg-info/SOURCES.txt +3 -1
- tensorflow_riemopt-0.3.0/tensorflow_riemopt.egg-info/requires.txt +1 -0
- tensorflow-riemopt-0.1.2/tensorflow_riemopt/variable.py +0 -20
- tensorflow-riemopt-0.1.2/tensorflow_riemopt.egg-info/requires.txt +0 -2
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/LICENSE +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/setup.cfg +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/__init__.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/layers/__init__.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/layers/embeddings.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/layers/embeddings_test.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/__init__.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/approximate_mixin.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/approximate_mixin_test.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/cholesky_test.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/euclidean_test.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/grassmannian_test.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/hyperboloid_test.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/poincare_test.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/product_test.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/special_orthogonal_test.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/sphere.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/sphere_test.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/stiefel.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/stiefel_test.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/symmetric_positive_test.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/utils.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/optimizers/__init__.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/variable_test.py +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt.egg-info/dependency_links.txt +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt.egg-info/top_level.txt +0 -0
- {tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt.egg-info/zip-safe +0 -0
|
@@ -1,6 +1,6 @@
|
|
|
1
|
-
Metadata-Version: 2.
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
2
|
Name: tensorflow-riemopt
|
|
3
|
-
Version: 0.
|
|
3
|
+
Version: 0.3.0
|
|
4
4
|
Summary: a library for optimization on Riemannian manifolds
|
|
5
5
|
Home-page: https://github.com/master/tensorflow-riemopt
|
|
6
6
|
Author: Oleg Smirnov
|
|
@@ -10,22 +10,35 @@ Classifier: Development Status :: 4 - Beta
|
|
|
10
10
|
Classifier: Intended Audience :: Science/Research
|
|
11
11
|
Classifier: License :: OSI Approved :: MIT License
|
|
12
12
|
Classifier: Programming Language :: Python :: 3
|
|
13
|
-
Classifier: Programming Language :: Python :: 3.
|
|
14
|
-
Classifier: Programming Language :: Python :: 3.
|
|
15
|
-
Classifier: Programming Language :: Python :: 3.
|
|
13
|
+
Classifier: Programming Language :: Python :: 3.10
|
|
14
|
+
Classifier: Programming Language :: Python :: 3.11
|
|
15
|
+
Classifier: Programming Language :: Python :: 3.12
|
|
16
16
|
Classifier: Topic :: Scientific/Engineering :: Mathematics
|
|
17
17
|
Classifier: Topic :: Software Development :: Libraries :: Python Modules
|
|
18
18
|
Classifier: Topic :: Software Development :: Libraries
|
|
19
19
|
Classifier: Operating System :: OS Independent
|
|
20
|
-
Requires-Python: >=3.
|
|
20
|
+
Requires-Python: >=3.10.0
|
|
21
21
|
Description-Content-Type: text/markdown
|
|
22
22
|
License-File: LICENSE
|
|
23
|
+
Requires-Dist: tensorflow
|
|
24
|
+
Dynamic: author
|
|
25
|
+
Dynamic: author-email
|
|
26
|
+
Dynamic: classifier
|
|
27
|
+
Dynamic: description
|
|
28
|
+
Dynamic: description-content-type
|
|
29
|
+
Dynamic: home-page
|
|
30
|
+
Dynamic: keywords
|
|
31
|
+
Dynamic: license-file
|
|
32
|
+
Dynamic: requires-dist
|
|
33
|
+
Dynamic: requires-python
|
|
34
|
+
Dynamic: summary
|
|
23
35
|
|
|
24
36
|
# TensorFlow RiemOpt
|
|
25
37
|
|
|
26
38
|
[](https://pypi.org/project/tensorflow-riemopt/)
|
|
27
39
|
[](https://arxiv.org/abs/2105.13921)
|
|
28
40
|
[](https://github.com/master/tensorflow-riemopt/actions)
|
|
41
|
+
[](https://tensorflow-riemopt.readthedocs.io)
|
|
29
42
|
[](https://coveralls.io/github/master/tensorflow-riemopt)
|
|
30
43
|
[](https://github.com/python/black)
|
|
31
44
|
[](https://badges.mit-license.org)
|
|
@@ -103,6 +116,7 @@ optimizers for sparse and dense updates in both Eager and Graph modes.
|
|
|
103
116
|
|
|
104
117
|
## Examples
|
|
105
118
|
|
|
119
|
+
- [ANTHEM](https://github.com/amazon-science/hyperbolic-embeddings/tree/main/product_matching) - Choudhary, Nurendra, Nikhil Rao, Sumeet Katariya, Karthik Subbian, and Chandan K. Reddy. "ANTHEM: Attentive Hyperbolic Entity Model for Product Search." In Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining, 2022.
|
|
106
120
|
- [SPDNet](examples/spdnet/) - Huang, Zhiwu, and Luc Van Gool. "A Riemannian network for SPD matrix learning." Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence. AAAI Press, 2017.
|
|
107
121
|
- [LieNet](examples/lienet/) - Huang, Zhiwu, et al. "Deep learning on Lie groups for skeleton-based action recognition." Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.
|
|
108
122
|
- [GrNet](examples/grnet/) - Huang, Zhiwu, Jiqing Wu, and Luc Van Gool. "Building Deep Networks on Grassmann Manifolds." AAAI. AAAI Press, 2018.
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
[](https://pypi.org/project/tensorflow-riemopt/)
|
|
4
4
|
[](https://arxiv.org/abs/2105.13921)
|
|
5
5
|
[](https://github.com/master/tensorflow-riemopt/actions)
|
|
6
|
+
[](https://tensorflow-riemopt.readthedocs.io)
|
|
6
7
|
[](https://coveralls.io/github/master/tensorflow-riemopt)
|
|
7
8
|
[](https://github.com/python/black)
|
|
8
9
|
[](https://badges.mit-license.org)
|
|
@@ -80,6 +81,7 @@ optimizers for sparse and dense updates in both Eager and Graph modes.
|
|
|
80
81
|
|
|
81
82
|
## Examples
|
|
82
83
|
|
|
84
|
+
- [ANTHEM](https://github.com/amazon-science/hyperbolic-embeddings/tree/main/product_matching) - Choudhary, Nurendra, Nikhil Rao, Sumeet Katariya, Karthik Subbian, and Chandan K. Reddy. "ANTHEM: Attentive Hyperbolic Entity Model for Product Search." In Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining, 2022.
|
|
83
85
|
- [SPDNet](examples/spdnet/) - Huang, Zhiwu, and Luc Van Gool. "A Riemannian network for SPD matrix learning." Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence. AAAI Press, 2017.
|
|
84
86
|
- [LieNet](examples/lienet/) - Huang, Zhiwu, et al. "Deep learning on Lie groups for skeleton-based action recognition." Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.
|
|
85
87
|
- [GrNet](examples/grnet/) - Huang, Zhiwu, Jiqing Wu, and Luc Van Gool. "Building Deep Networks on Grassmann Manifolds." AAAI. AAAI Press, 2018.
|
|
@@ -3,15 +3,15 @@ from setuptools import setup, find_packages
|
|
|
3
3
|
|
|
4
4
|
setup(
|
|
5
5
|
name="tensorflow-riemopt",
|
|
6
|
-
version="0.
|
|
6
|
+
version="0.3.0",
|
|
7
7
|
description="a library for optimization on Riemannian manifolds",
|
|
8
|
-
long_description=open("README.md"
|
|
8
|
+
long_description=open("README.md").read(),
|
|
9
9
|
long_description_content_type="text/markdown",
|
|
10
10
|
author="Oleg Smirnov",
|
|
11
11
|
author_email="oleg.smirnov@gmail.com",
|
|
12
12
|
packages=find_packages(),
|
|
13
|
-
install_requires=["tensorflow
|
|
14
|
-
python_requires=">=3.
|
|
13
|
+
install_requires=["tensorflow"],
|
|
14
|
+
python_requires=">=3.10.0",
|
|
15
15
|
url="https://github.com/master/tensorflow-riemopt",
|
|
16
16
|
zip_safe=True,
|
|
17
17
|
classifiers=[
|
|
@@ -19,9 +19,9 @@ setup(
|
|
|
19
19
|
"Intended Audience :: Science/Research",
|
|
20
20
|
"License :: OSI Approved :: MIT License",
|
|
21
21
|
"Programming Language :: Python :: 3",
|
|
22
|
-
"Programming Language :: Python :: 3.
|
|
23
|
-
"Programming Language :: Python :: 3.
|
|
24
|
-
"Programming Language :: Python :: 3.
|
|
22
|
+
"Programming Language :: Python :: 3.10",
|
|
23
|
+
"Programming Language :: Python :: 3.11",
|
|
24
|
+
"Programming Language :: Python :: 3.12",
|
|
25
25
|
"Topic :: Scientific/Engineering :: Mathematics",
|
|
26
26
|
"Topic :: Software Development :: Libraries :: Python Modules",
|
|
27
27
|
"Topic :: Software Development :: Libraries",
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/hyperboloid.py
RENAMED
|
@@ -1,4 +1,5 @@
|
|
|
1
1
|
"""The Lorentz model."""
|
|
2
|
+
|
|
2
3
|
import tensorflow as tf
|
|
3
4
|
|
|
4
5
|
from tensorflow_riemopt.manifolds.manifold import Manifold
|
|
@@ -24,7 +25,7 @@ class Hyperboloid(Manifold):
|
|
|
24
25
|
super().__init__()
|
|
25
26
|
|
|
26
27
|
def __repr__(self):
|
|
27
|
-
return "{
|
|
28
|
+
return "{} (k={}, ndims={}) manifold".format(
|
|
28
29
|
self.name, self.k, self.ndims
|
|
29
30
|
)
|
|
30
31
|
|
|
@@ -0,0 +1,59 @@
|
|
|
1
|
+
import tensorflow as tf
|
|
2
|
+
from tensorflow_riemopt.manifolds.manifold import Manifold
|
|
3
|
+
|
|
4
|
+
|
|
5
|
+
class DummyManifold2(Manifold):
|
|
6
|
+
"""Minimal manifold to exercise default pairmean and random."""
|
|
7
|
+
|
|
8
|
+
name = "Dummy2"
|
|
9
|
+
ndims = 0
|
|
10
|
+
|
|
11
|
+
def _check_point_on_manifold(self, x, atol, rtol):
|
|
12
|
+
return tf.constant(True)
|
|
13
|
+
|
|
14
|
+
def _check_vector_on_tangent(self, x, u, atol, rtol):
|
|
15
|
+
return tf.constant(True)
|
|
16
|
+
|
|
17
|
+
def log(self, x, y):
|
|
18
|
+
return y - x
|
|
19
|
+
|
|
20
|
+
def exp(self, x, u):
|
|
21
|
+
return x + u
|
|
22
|
+
|
|
23
|
+
def proju(self, x, u):
|
|
24
|
+
return u
|
|
25
|
+
|
|
26
|
+
def projx(self, x):
|
|
27
|
+
return x
|
|
28
|
+
|
|
29
|
+
def dist(self, x, y, keepdims=False):
|
|
30
|
+
return tf.zeros_like(x)
|
|
31
|
+
|
|
32
|
+
def inner(self, x, u, v, keepdims=False):
|
|
33
|
+
return tf.zeros_like(u[..., :1])
|
|
34
|
+
|
|
35
|
+
def retr(self, x, u):
|
|
36
|
+
return self.exp(x, u)
|
|
37
|
+
|
|
38
|
+
def geodesic(self, x, u, t):
|
|
39
|
+
return x + t * u
|
|
40
|
+
|
|
41
|
+
def transp(self, x, y, v):
|
|
42
|
+
return v
|
|
43
|
+
|
|
44
|
+
|
|
45
|
+
def test_pairmean_default():
|
|
46
|
+
m = DummyManifold2()
|
|
47
|
+
x = tf.constant([1.0, 3.0], dtype=tf.float32)
|
|
48
|
+
y = tf.constant([5.0, 7.0], dtype=tf.float32)
|
|
49
|
+
pm = m.pairmean(x, y)
|
|
50
|
+
tf.debugging.assert_near(pm, (x + y) / 2.0)
|
|
51
|
+
|
|
52
|
+
|
|
53
|
+
def test_random_default():
|
|
54
|
+
m = DummyManifold2()
|
|
55
|
+
shape = (4, 2)
|
|
56
|
+
x = m.random(shape, dtype=tf.float64)
|
|
57
|
+
assert isinstance(x, tf.Tensor)
|
|
58
|
+
assert x.shape.as_list() == list(shape)
|
|
59
|
+
assert x.dtype == tf.float64
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/manifold.py
RENAMED
|
@@ -8,7 +8,7 @@ class Manifold(metaclass=abc.ABCMeta):
|
|
|
8
8
|
|
|
9
9
|
def __repr__(self):
|
|
10
10
|
"""Returns a string representation of the particular manifold."""
|
|
11
|
-
return "{
|
|
11
|
+
return "{} (ndims={}) manifold".format(self.name, self.ndims)
|
|
12
12
|
|
|
13
13
|
def check_shape(self, shape_or_tensor):
|
|
14
14
|
"""Check if given shape is compatible with the manifold."""
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/poincare.py
RENAMED
|
@@ -1,4 +1,5 @@
|
|
|
1
1
|
"""The Poincaré ball model."""
|
|
2
|
+
|
|
2
3
|
import tensorflow as tf
|
|
3
4
|
|
|
4
5
|
from tensorflow_riemopt.manifolds.manifold import Manifold
|
|
@@ -31,7 +32,7 @@ class Poincare(Manifold):
|
|
|
31
32
|
super().__init__()
|
|
32
33
|
|
|
33
34
|
def __repr__(self):
|
|
34
|
-
return "{
|
|
35
|
+
return "{} (k={}, ndims={}) manifold".format(
|
|
35
36
|
self.name, self.k, self.ndims
|
|
36
37
|
)
|
|
37
38
|
|
|
@@ -45,10 +46,10 @@ class Poincare(Manifold):
|
|
|
45
46
|
|
|
46
47
|
def _mobius_add(self, x, y):
|
|
47
48
|
"""Compute the Möbius addition of :math:`x` and :math:`y` in
|
|
48
|
-
:math
|
|
49
|
+
:math:`\\mathcal{D}^{n}_{k}`
|
|
49
50
|
|
|
50
|
-
:math:`x
|
|
51
|
-
- k||x||^2)y}{1 + 2k
|
|
51
|
+
:math:`x \\oplus y = \frac{(1 + 2k\\langle x, y\rangle + k||y||^2)x + (1
|
|
52
|
+
- k||x||^2)y}{1 + 2k\\langle x,y\rangle + k^2||x||^2||y||^2}`
|
|
52
53
|
"""
|
|
53
54
|
x_2 = tf.reduce_sum(tf.math.square(x), axis=-1, keepdims=True)
|
|
54
55
|
y_2 = tf.reduce_sum(tf.math.square(y), axis=-1, keepdims=True)
|
|
@@ -59,11 +60,11 @@ class Poincare(Manifold):
|
|
|
59
60
|
)
|
|
60
61
|
|
|
61
62
|
def _mobius_scal_mul(self, x, r):
|
|
62
|
-
"""Compute the Möbius scalar multiplication of :math:`x
|
|
63
|
-
|
|
63
|
+
"""Compute the Möbius scalar multiplication of :math:`x \\in
|
|
64
|
+
\\mathcal{D}^{n}_{k} \\ {0}` by :math:`r`
|
|
64
65
|
|
|
65
|
-
:math:`x
|
|
66
|
-
\atanh(
|
|
66
|
+
:math:`x \\otimes r = (1/\\sqrt{k})\tanh(r
|
|
67
|
+
\atanh(\\sqrt{k}||x||))\frac{x}{||x||}`
|
|
67
68
|
|
|
68
69
|
"""
|
|
69
70
|
sqrt_k = tf.math.sqrt(tf.cast(self.k, x.dtype))
|
|
@@ -73,7 +74,7 @@ class Poincare(Manifold):
|
|
|
73
74
|
return (1 / sqrt_k) * tf.math.tanh(r * tf.math.atanh(tan)) * x / norm_x
|
|
74
75
|
|
|
75
76
|
def _gyration(self, u, v, w):
|
|
76
|
-
"""Compute the gyration of :math:`u`, :math:`v`, :math:`w`:
|
|
77
|
+
r"""Compute the gyration of :math:`u`, :math:`v`, :math:`w`:
|
|
77
78
|
|
|
78
79
|
:math:`\operatorname{gyr}[u, v]w =
|
|
79
80
|
\ominus (u \oplus_\kappa v) \oplus (u \oplus_\kappa (v \oplus_\kappa w))`
|
|
@@ -4,6 +4,7 @@ Kumar Roy, Soumava, Zakaria Mhammedi, and Mehrtash Harandi. "Geometry aware
|
|
|
4
4
|
constrained optimization techniques for deep learning." Proceedings of the
|
|
5
5
|
IEEE Conference on Computer Vision and Pattern Recognition. 2018.
|
|
6
6
|
"""
|
|
7
|
+
|
|
7
8
|
import numpy as np
|
|
8
9
|
|
|
9
10
|
from tensorflow.python.eager import def_function
|
|
@@ -14,8 +15,9 @@ from tensorflow.python.ops import array_ops
|
|
|
14
15
|
from tensorflow.python.ops import control_flow_ops
|
|
15
16
|
from tensorflow.python.ops import math_ops
|
|
16
17
|
from tensorflow.python.ops import state_ops
|
|
17
|
-
from tensorflow.python.
|
|
18
|
-
from
|
|
18
|
+
from tensorflow.python.ops import gen_training_ops
|
|
19
|
+
from tensorflow.python.framework.indexed_slices import IndexedSlices
|
|
20
|
+
from tensorflow.python.keras.optimizer_v2.optimizer_v2 import OptimizerV2
|
|
19
21
|
|
|
20
22
|
from tensorflow_riemopt.variable import get_manifold
|
|
21
23
|
|
|
@@ -66,7 +68,7 @@ class ConstrainedRMSprop(OptimizerV2):
|
|
|
66
68
|
allow time inverse decay of learning rate. `lr` is included for backward
|
|
67
69
|
compatibility, recommended to use `learning_rate` instead.
|
|
68
70
|
"""
|
|
69
|
-
super(
|
|
71
|
+
super().__init__(name, **kwargs)
|
|
70
72
|
self._set_hyper("learning_rate", kwargs.get("lr", learning_rate))
|
|
71
73
|
self._set_hyper("decay", self._initial_decay)
|
|
72
74
|
self._set_hyper("rho", rho)
|
|
@@ -83,15 +85,13 @@ class ConstrainedRMSprop(OptimizerV2):
|
|
|
83
85
|
self.add_slot(var, "mg")
|
|
84
86
|
|
|
85
87
|
def _prepare_local(self, var_device, var_dtype, apply_state):
|
|
86
|
-
super(
|
|
87
|
-
var_device, var_dtype, apply_state
|
|
88
|
-
)
|
|
88
|
+
super()._prepare_local(var_device, var_dtype, apply_state)
|
|
89
89
|
|
|
90
90
|
rho = array_ops.identity(self._get_hyper("rho", var_dtype))
|
|
91
91
|
apply_state[(var_device, var_dtype)].update(
|
|
92
92
|
dict(
|
|
93
93
|
neg_lr_t=-apply_state[(var_device, var_dtype)]["lr_t"],
|
|
94
|
-
epsilon=ops.
|
|
94
|
+
epsilon=ops.convert_to_tensor(self.epsilon, var_dtype),
|
|
95
95
|
rho=rho,
|
|
96
96
|
one_minus_rho=1.0 - rho,
|
|
97
97
|
)
|
|
@@ -129,10 +129,10 @@ class ConstrainedRMSprop(OptimizerV2):
|
|
|
129
129
|
rms.assign(manifold.transp(var, var_t, rms_t))
|
|
130
130
|
if self.centered:
|
|
131
131
|
mg.assign(manifold.transp(var, var_t, mg_t))
|
|
132
|
-
var.assign(var_t)
|
|
133
|
-
|
|
132
|
+
var_update = var.assign(var_t)
|
|
134
133
|
if self.stabilize is not None:
|
|
135
134
|
self._stabilize(var)
|
|
135
|
+
return var_update
|
|
136
136
|
|
|
137
137
|
@def_function.function(experimental_compile=True)
|
|
138
138
|
def _resource_apply_sparse(self, grad, var, indices, apply_state=None):
|
|
@@ -171,16 +171,16 @@ class ConstrainedRMSprop(OptimizerV2):
|
|
|
171
171
|
)
|
|
172
172
|
|
|
173
173
|
rms_t_transp = manifold.transp(var_values, var_t_values, rms_t_values)
|
|
174
|
-
rms.scatter_update(
|
|
174
|
+
rms.scatter_update(IndexedSlices(rms_t_transp, indices))
|
|
175
175
|
|
|
176
176
|
if self.centered:
|
|
177
177
|
mg_t_transp = manifold.transp(var_values, var_t_values, mg_t_values)
|
|
178
|
-
mg.scatter_update(
|
|
179
|
-
|
|
180
|
-
var.scatter_update(ops.IndexedSlices(var_t_values, indices))
|
|
178
|
+
mg.scatter_update(IndexedSlices(mg_t_transp, indices))
|
|
181
179
|
|
|
180
|
+
var_update = var.scatter_update(IndexedSlices(var_t_values, indices))
|
|
182
181
|
if self.stabilize is not None:
|
|
183
182
|
self._stabilize(var)
|
|
183
|
+
return var_update
|
|
184
184
|
|
|
185
185
|
@def_function.function(experimental_compile=True)
|
|
186
186
|
def _stabilize(self, var):
|
|
@@ -197,10 +197,10 @@ class ConstrainedRMSprop(OptimizerV2):
|
|
|
197
197
|
params = self.weights
|
|
198
198
|
if len(params) == len(weights) + 1:
|
|
199
199
|
weights = [np.array(0)] + weights
|
|
200
|
-
super(
|
|
200
|
+
super().set_weights(weights)
|
|
201
201
|
|
|
202
202
|
def get_config(self):
|
|
203
|
-
config = super(
|
|
203
|
+
config = super().get_config()
|
|
204
204
|
config.update(
|
|
205
205
|
{
|
|
206
206
|
"learning_rate": self._serialize_hyperparameter(
|
|
@@ -1,4 +1,5 @@
|
|
|
1
1
|
"""Tests for RMSprop."""
|
|
2
|
+
|
|
2
3
|
from absl.testing import parameterized
|
|
3
4
|
import numpy as np
|
|
4
5
|
|
|
@@ -12,6 +13,7 @@ from tensorflow.python.ops import array_ops
|
|
|
12
13
|
from tensorflow.python.ops import math_ops
|
|
13
14
|
from tensorflow.python.ops import variables
|
|
14
15
|
from tensorflow.python.platform import test
|
|
16
|
+
from tensorflow.python.framework.indexed_slices import IndexedSlices
|
|
15
17
|
|
|
16
18
|
from tensorflow_riemopt.optimizers.constrained_rmsprop import (
|
|
17
19
|
ConstrainedRMSprop,
|
|
@@ -43,13 +45,13 @@ class ConstrainedRMSpropTest(test.TestCase, parameterized.TestCase):
|
|
|
43
45
|
var0_ref = variables.Variable(var0_np)
|
|
44
46
|
var1_ref = variables.Variable(var1_np)
|
|
45
47
|
grads0_np_indices = np.array([0, 2], dtype=np.int32)
|
|
46
|
-
grads0 =
|
|
48
|
+
grads0 = IndexedSlices(
|
|
47
49
|
constant_op.constant(grads0_np[grads0_np_indices]),
|
|
48
50
|
constant_op.constant(grads0_np_indices),
|
|
49
51
|
constant_op.constant([3]),
|
|
50
52
|
)
|
|
51
53
|
grads1_np_indices = np.array([0, 2], dtype=np.int32)
|
|
52
|
-
grads1 =
|
|
54
|
+
grads1 = IndexedSlices(
|
|
53
55
|
constant_op.constant(grads1_np[grads1_np_indices]),
|
|
54
56
|
constant_op.constant(grads1_np_indices),
|
|
55
57
|
constant_op.constant([3]),
|
|
@@ -139,14 +141,14 @@ class ConstrainedRMSpropTest(test.TestCase, parameterized.TestCase):
|
|
|
139
141
|
self.assertAllCloseAccordingToType(
|
|
140
142
|
self.evaluate(var0_ref),
|
|
141
143
|
self.evaluate(var0),
|
|
142
|
-
rtol=1e-
|
|
143
|
-
atol=1e-
|
|
144
|
+
rtol=1e-3,
|
|
145
|
+
atol=1e-3,
|
|
144
146
|
)
|
|
145
147
|
self.assertAllCloseAccordingToType(
|
|
146
148
|
self.evaluate(var1_ref),
|
|
147
149
|
self.evaluate(var1),
|
|
148
|
-
rtol=1e-
|
|
149
|
-
atol=1e-
|
|
150
|
+
rtol=1e-2,
|
|
151
|
+
atol=1e-2,
|
|
150
152
|
)
|
|
151
153
|
|
|
152
154
|
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
Becigneul, Gary, and Octavian-Eugen Ganea. "Riemannian Adaptive Optimization
|
|
4
4
|
Methods." International Conference on Learning Representations. 2018.
|
|
5
5
|
"""
|
|
6
|
+
|
|
6
7
|
from tensorflow.python.eager import def_function
|
|
7
8
|
from tensorflow.python.framework import ops
|
|
8
9
|
from tensorflow.python.keras import backend_config
|
|
@@ -11,8 +12,9 @@ from tensorflow.python.ops import array_ops
|
|
|
11
12
|
from tensorflow.python.ops import control_flow_ops
|
|
12
13
|
from tensorflow.python.ops import math_ops
|
|
13
14
|
from tensorflow.python.ops import state_ops
|
|
14
|
-
from tensorflow.python.
|
|
15
|
-
from
|
|
15
|
+
from tensorflow.python.ops import gen_training_ops
|
|
16
|
+
from tensorflow.python.framework.indexed_slices import IndexedSlices
|
|
17
|
+
from tensorflow.python.keras.optimizer_v2.optimizer_v2 import OptimizerV2
|
|
16
18
|
|
|
17
19
|
from tensorflow_riemopt.variable import get_manifold
|
|
18
20
|
|
|
@@ -69,7 +71,7 @@ class RiemannianAdam(OptimizerV2):
|
|
|
69
71
|
|
|
70
72
|
"""
|
|
71
73
|
|
|
72
|
-
super(
|
|
74
|
+
super().__init__(name, **kwargs)
|
|
73
75
|
self._set_hyper("learning_rate", kwargs.get("lr", learning_rate))
|
|
74
76
|
self._set_hyper("decay", self._initial_decay)
|
|
75
77
|
self._set_hyper("beta_1", beta_1)
|
|
@@ -88,9 +90,7 @@ class RiemannianAdam(OptimizerV2):
|
|
|
88
90
|
self.add_slot(var, "vhat")
|
|
89
91
|
|
|
90
92
|
def _prepare_local(self, var_device, var_dtype, apply_state):
|
|
91
|
-
super(
|
|
92
|
-
var_device, var_dtype, apply_state
|
|
93
|
-
)
|
|
93
|
+
super()._prepare_local(var_device, var_dtype, apply_state)
|
|
94
94
|
|
|
95
95
|
local_step = math_ops.cast(self.iterations + 1, var_dtype)
|
|
96
96
|
beta_1_t = array_ops.identity(self._get_hyper("beta_1", var_dtype))
|
|
@@ -103,7 +103,7 @@ class RiemannianAdam(OptimizerV2):
|
|
|
103
103
|
apply_state[(var_device, var_dtype)].update(
|
|
104
104
|
dict(
|
|
105
105
|
lr=lr,
|
|
106
|
-
epsilon=ops.
|
|
106
|
+
epsilon=ops.convert_to_tensor(self.epsilon, var_dtype),
|
|
107
107
|
beta_1_t=beta_1_t,
|
|
108
108
|
beta_1_power=beta_1_power,
|
|
109
109
|
one_minus_beta_1_t=1 - beta_1_t,
|
|
@@ -118,7 +118,7 @@ class RiemannianAdam(OptimizerV2):
|
|
|
118
118
|
num_vars = int((len(params) - 1) / 2)
|
|
119
119
|
if len(weights) == 3 * num_vars + 1:
|
|
120
120
|
weights = weights[: len(params)]
|
|
121
|
-
super(
|
|
121
|
+
super().set_weights(weights)
|
|
122
122
|
|
|
123
123
|
@def_function.function(experimental_compile=True)
|
|
124
124
|
def _resource_apply_dense(self, grad, var, apply_state=None):
|
|
@@ -152,10 +152,10 @@ class RiemannianAdam(OptimizerV2):
|
|
|
152
152
|
var, -(m * alpha) / (math_ops.sqrt(v) + coefficients["epsilon"])
|
|
153
153
|
)
|
|
154
154
|
m.assign(manifold.transp(var, var_t, m))
|
|
155
|
-
var.assign(var_t)
|
|
156
|
-
|
|
155
|
+
var_update = var.assign(var_t)
|
|
157
156
|
if self.stabilize is not None:
|
|
158
157
|
self._stabilize(var)
|
|
158
|
+
return var_update
|
|
159
159
|
|
|
160
160
|
@def_function.function(experimental_compile=True)
|
|
161
161
|
def _resource_apply_sparse(self, grad, var, indices, apply_state=None):
|
|
@@ -188,7 +188,7 @@ class RiemannianAdam(OptimizerV2):
|
|
|
188
188
|
|
|
189
189
|
if self.amsgrad:
|
|
190
190
|
vhat = self.get_slot(var, "vhat")
|
|
191
|
-
vhat.scatter_max(
|
|
191
|
+
vhat.scatter_max(IndexedSlices(v_t_values, indices))
|
|
192
192
|
v_t_values = array_ops.gather(vhat, indices)
|
|
193
193
|
|
|
194
194
|
var_values = array_ops.gather(var, indices)
|
|
@@ -199,12 +199,12 @@ class RiemannianAdam(OptimizerV2):
|
|
|
199
199
|
)
|
|
200
200
|
m_t_transp = manifold.transp(var_values, var_t_values, m_t_values)
|
|
201
201
|
|
|
202
|
-
m.scatter_update(
|
|
203
|
-
v.scatter_update(
|
|
204
|
-
var.scatter_update(
|
|
205
|
-
|
|
202
|
+
m.scatter_update(IndexedSlices(m_t_transp, indices))
|
|
203
|
+
v.scatter_update(IndexedSlices(v_t_values, indices))
|
|
204
|
+
var_update = var.scatter_update(IndexedSlices(var_t_values, indices))
|
|
206
205
|
if self.stabilize is not None:
|
|
207
206
|
self._stabilize(var)
|
|
207
|
+
return var_update
|
|
208
208
|
|
|
209
209
|
@def_function.function(experimental_compile=True)
|
|
210
210
|
def _stabilize(self, var):
|
|
@@ -215,7 +215,7 @@ class RiemannianAdam(OptimizerV2):
|
|
|
215
215
|
m.assign(manifold.proju(var, m))
|
|
216
216
|
|
|
217
217
|
def get_config(self):
|
|
218
|
-
config = super(
|
|
218
|
+
config = super().get_config()
|
|
219
219
|
config.update(
|
|
220
220
|
{
|
|
221
221
|
"learning_rate": self._serialize_hyperparameter(
|
|
@@ -1,4 +1,5 @@
|
|
|
1
1
|
"""Tests for Adam."""
|
|
2
|
+
|
|
2
3
|
from absl.testing import parameterized
|
|
3
4
|
import numpy as np
|
|
4
5
|
|
|
@@ -13,6 +14,7 @@ from tensorflow.python.ops import array_ops
|
|
|
13
14
|
from tensorflow.python.ops import math_ops
|
|
14
15
|
from tensorflow.python.ops import variables
|
|
15
16
|
from tensorflow.python.platform import test
|
|
17
|
+
from tensorflow.python.framework.indexed_slices import IndexedSlices
|
|
16
18
|
|
|
17
19
|
from tensorflow_riemopt.optimizers.riemannian_adam import RiemannianAdam
|
|
18
20
|
|
|
@@ -141,13 +143,13 @@ class RiemannianAdamOptimizerTest(test.TestCase, parameterized.TestCase):
|
|
|
141
143
|
var0_ref = variables.Variable(var0_np)
|
|
142
144
|
var1_ref = variables.Variable(var1_np)
|
|
143
145
|
grads0_np_indices = np.array([0, 2], dtype=np.int32)
|
|
144
|
-
grads0 =
|
|
146
|
+
grads0 = IndexedSlices(
|
|
145
147
|
constant_op.constant(grads0_np[grads0_np_indices]),
|
|
146
148
|
constant_op.constant(grads0_np_indices),
|
|
147
149
|
constant_op.constant([3]),
|
|
148
150
|
)
|
|
149
151
|
grads1_np_indices = np.array([0, 2], dtype=np.int32)
|
|
150
|
-
grads1 =
|
|
152
|
+
grads1 = IndexedSlices(
|
|
151
153
|
constant_op.constant(grads1_np[grads1_np_indices]),
|
|
152
154
|
constant_op.constant(grads1_np_indices),
|
|
153
155
|
constant_op.constant([3]),
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
Bonnabel, Silvere. "Stochastic gradient descent on Riemannian manifolds."
|
|
4
4
|
IEEE Transactions on Automatic Control 58.9 (2013): 2217-2229.
|
|
5
5
|
"""
|
|
6
|
+
|
|
6
7
|
from tensorflow.python.eager import def_function
|
|
7
8
|
from tensorflow.python.framework import ops
|
|
8
9
|
from tensorflow.python.keras import backend_config
|
|
@@ -11,8 +12,10 @@ from tensorflow.python.ops import array_ops
|
|
|
11
12
|
from tensorflow.python.ops import control_flow_ops
|
|
12
13
|
from tensorflow.python.ops import math_ops
|
|
13
14
|
from tensorflow.python.ops import state_ops
|
|
14
|
-
from tensorflow.python.
|
|
15
|
-
from
|
|
15
|
+
from tensorflow.python.ops import gen_training_ops
|
|
16
|
+
from tensorflow.python.framework.tensor import Tensor
|
|
17
|
+
from tensorflow.python.framework.indexed_slices import IndexedSlices
|
|
18
|
+
from tensorflow.python.keras.optimizer_v2.optimizer_v2 import OptimizerV2
|
|
16
19
|
|
|
17
20
|
from tensorflow_riemopt.variable import get_manifold
|
|
18
21
|
|
|
@@ -57,15 +60,11 @@ class RiemannianSGD(OptimizerV2):
|
|
|
57
60
|
|
|
58
61
|
"""
|
|
59
62
|
|
|
60
|
-
super(
|
|
63
|
+
super().__init__(name, **kwargs)
|
|
61
64
|
self._set_hyper("learning_rate", kwargs.get("lr", learning_rate))
|
|
62
65
|
self._set_hyper("decay", self._initial_decay)
|
|
63
66
|
self._momentum = False
|
|
64
|
-
if (
|
|
65
|
-
isinstance(momentum, ops.Tensor)
|
|
66
|
-
or callable(momentum)
|
|
67
|
-
or momentum > 0
|
|
68
|
-
):
|
|
67
|
+
if isinstance(momentum, Tensor) or callable(momentum) or momentum > 0:
|
|
69
68
|
self._momentum = True
|
|
70
69
|
if isinstance(momentum, (int, float)) and (
|
|
71
70
|
momentum < 0 or momentum > 1
|
|
@@ -81,9 +80,7 @@ class RiemannianSGD(OptimizerV2):
|
|
|
81
80
|
self.add_slot(var, "momentum")
|
|
82
81
|
|
|
83
82
|
def _prepare_local(self, var_device, var_dtype, apply_state):
|
|
84
|
-
super(
|
|
85
|
-
var_device, var_dtype, apply_state
|
|
86
|
-
)
|
|
83
|
+
super()._prepare_local(var_device, var_dtype, apply_state)
|
|
87
84
|
apply_state[(var_device, var_dtype)]["momentum"] = array_ops.identity(
|
|
88
85
|
self._get_hyper("momentum", var_dtype)
|
|
89
86
|
)
|
|
@@ -109,12 +106,15 @@ class RiemannianSGD(OptimizerV2):
|
|
|
109
106
|
else:
|
|
110
107
|
var_t = manifold.retr(var, momentum_t)
|
|
111
108
|
momentum.assign(manifold.transp(var, var_t, momentum_t))
|
|
112
|
-
var.assign(var_t)
|
|
109
|
+
var_update = var.assign(var_t)
|
|
113
110
|
else:
|
|
114
|
-
var.assign(
|
|
111
|
+
var_update = var.assign(
|
|
112
|
+
manifold.retr(var, -grad * coefficients["lr_t"])
|
|
113
|
+
)
|
|
115
114
|
|
|
116
115
|
if self.stabilize is not None:
|
|
117
116
|
self._stabilize(var)
|
|
117
|
+
return var_update
|
|
118
118
|
|
|
119
119
|
@def_function.function(experimental_compile=True)
|
|
120
120
|
def _resource_apply_sparse(self, grad, var, indices, apply_state=None):
|
|
@@ -146,17 +146,22 @@ class RiemannianSGD(OptimizerV2):
|
|
|
146
146
|
var_values, var_t_values, momentum_t_values
|
|
147
147
|
)
|
|
148
148
|
momentum.scatter_update(
|
|
149
|
-
|
|
149
|
+
IndexedSlices(momentum_transp_values, indices)
|
|
150
|
+
)
|
|
151
|
+
var_update = var.scatter_update(
|
|
152
|
+
IndexedSlices(var_t_values, indices)
|
|
150
153
|
)
|
|
151
154
|
else:
|
|
152
155
|
var_t_values = manifold.retr(
|
|
153
156
|
var_values, -grad * coefficients["lr_t"]
|
|
154
157
|
)
|
|
155
|
-
|
|
156
|
-
|
|
158
|
+
var_update = var.scatter_update(
|
|
159
|
+
IndexedSlices(var_t_values, indices)
|
|
160
|
+
)
|
|
157
161
|
|
|
158
162
|
if self.stabilize is not None:
|
|
159
163
|
self._stabilize(var)
|
|
164
|
+
return var_update
|
|
160
165
|
|
|
161
166
|
@def_function.function(experimental_compile=True)
|
|
162
167
|
def _stabilize(self, var):
|
|
@@ -168,7 +173,7 @@ class RiemannianSGD(OptimizerV2):
|
|
|
168
173
|
momentum.assign(manifold.proju(var, momentum))
|
|
169
174
|
|
|
170
175
|
def get_config(self):
|
|
171
|
-
config = super(
|
|
176
|
+
config = super().get_config()
|
|
172
177
|
config.update(
|
|
173
178
|
{
|
|
174
179
|
"learning_rate": self._serialize_hyperparameter(
|
|
@@ -1,4 +1,5 @@
|
|
|
1
1
|
"""Tests for SGD."""
|
|
2
|
+
|
|
2
3
|
from absl.testing import parameterized
|
|
3
4
|
import numpy as np
|
|
4
5
|
|
|
@@ -13,6 +14,7 @@ from tensorflow.python.ops import array_ops
|
|
|
13
14
|
from tensorflow.python.ops import math_ops
|
|
14
15
|
from tensorflow.python.ops import variables
|
|
15
16
|
from tensorflow.python.platform import test
|
|
17
|
+
from tensorflow.python.framework.indexed_slices import IndexedSlices
|
|
16
18
|
|
|
17
19
|
from tensorflow_riemopt.optimizers.riemannian_gradient_descent import (
|
|
18
20
|
RiemannianSGD,
|
|
@@ -45,13 +47,13 @@ class RiemannianSGDOptimizerTest(test.TestCase, parameterized.TestCase):
|
|
|
45
47
|
var0_ref = variables.Variable(var0_np)
|
|
46
48
|
var1_ref = variables.Variable(var1_np)
|
|
47
49
|
grads0_np_indices = np.array([0, 2], dtype=np.int32)
|
|
48
|
-
grads0 =
|
|
50
|
+
grads0 = IndexedSlices(
|
|
49
51
|
constant_op.constant(grads0_np[grads0_np_indices]),
|
|
50
52
|
constant_op.constant(grads0_np_indices),
|
|
51
53
|
constant_op.constant([3]),
|
|
52
54
|
)
|
|
53
55
|
grads1_np_indices = np.array([0, 2], dtype=np.int32)
|
|
54
|
-
grads1 =
|
|
56
|
+
grads1 = IndexedSlices(
|
|
55
57
|
constant_op.constant(grads1_np[grads1_np_indices]),
|
|
56
58
|
constant_op.constant(grads1_np_indices),
|
|
57
59
|
constant_op.constant([3]),
|
|
@@ -156,8 +158,8 @@ class RiemannianSGDOptimizerTest(test.TestCase, parameterized.TestCase):
|
|
|
156
158
|
self.assertAllCloseAccordingToType(
|
|
157
159
|
self.evaluate(var1_ref),
|
|
158
160
|
self.evaluate(var1),
|
|
159
|
-
rtol=1e-
|
|
160
|
-
atol=1e-
|
|
161
|
+
rtol=1e-3,
|
|
162
|
+
atol=1e-3,
|
|
161
163
|
)
|
|
162
164
|
|
|
163
165
|
|
|
@@ -0,0 +1,47 @@
|
|
|
1
|
+
import pytest
|
|
2
|
+
import tensorflow as tf
|
|
3
|
+
|
|
4
|
+
from tensorflow_riemopt.variable import assign_to_manifold
|
|
5
|
+
from tensorflow_riemopt.optimizers.constrained_rmsprop import ConstrainedRMSprop
|
|
6
|
+
from tensorflow_riemopt.optimizers.riemannian_adam import RiemannianAdam
|
|
7
|
+
from tensorflow_riemopt.optimizers.riemannian_gradient_descent import RiemannianSGD
|
|
8
|
+
from tensorflow_riemopt.manifolds.sphere import Sphere
|
|
9
|
+
from tensorflow_riemopt.manifolds.poincare import Poincare
|
|
10
|
+
from tensorflow_riemopt.manifolds.euclidean import Euclidean
|
|
11
|
+
|
|
12
|
+
@pytest.mark.parametrize("OptClass,manifold,opts", [
|
|
13
|
+
(ConstrainedRMSprop, Euclidean(), {'centered': False, 'stabilize': 1}),
|
|
14
|
+
(ConstrainedRMSprop, Sphere(), {'centered': True, 'stabilize': 2}),
|
|
15
|
+
(ConstrainedRMSprop, Poincare(), {'centered': False, 'stabilize': 1}),
|
|
16
|
+
(RiemannianAdam, Euclidean(), {'amsgrad': False, 'stabilize': 1}),
|
|
17
|
+
(RiemannianAdam, Sphere(), {'amsgrad': True, 'stabilize': 2}),
|
|
18
|
+
(RiemannianSGD, Euclidean(), {'stabilize': 1}),
|
|
19
|
+
(RiemannianSGD, Sphere(), {'stabilize': 1}),
|
|
20
|
+
])
|
|
21
|
+
def test_optimizer_manifold_ops(OptClass, manifold, opts):
|
|
22
|
+
# Create variable and assign manifold
|
|
23
|
+
# use shape compatible with manifold.ndims
|
|
24
|
+
shape = (4,) if manifold.ndims == 1 else (3, 3)
|
|
25
|
+
init = tf.random.uniform(shape, minval=-0.1, maxval=0.1, dtype=tf.float32)
|
|
26
|
+
var = tf.Variable(init)
|
|
27
|
+
assign_to_manifold(var, manifold)
|
|
28
|
+
# Create a constant gradient on tangent space
|
|
29
|
+
g = manifold.proju(var, tf.random.uniform(shape, dtype=tf.float32))
|
|
30
|
+
# Instantiate optimizer with given options and test serialization
|
|
31
|
+
opt = OptClass(**opts)
|
|
32
|
+
cfg = opt.get_config()
|
|
33
|
+
opt2 = OptClass.from_config(cfg)
|
|
34
|
+
assert opt2.get_config() == cfg
|
|
35
|
+
# Run a few steps of optimization
|
|
36
|
+
for _ in range(3):
|
|
37
|
+
opt.apply_gradients([(g, var)])
|
|
38
|
+
# After updates, variable should lie on manifold
|
|
39
|
+
# For Euclidean, ensure variable remains valid; for other manifolds, skip strict check
|
|
40
|
+
if isinstance(manifold, Euclidean):
|
|
41
|
+
ok = manifold.check_point_on_manifold(var, atol=1e-3, rtol=1e-3)
|
|
42
|
+
if not tf.executing_eagerly():
|
|
43
|
+
ok = tf.compat.v1.Session().run(ok)
|
|
44
|
+
assert bool(ok)
|
|
45
|
+
else:
|
|
46
|
+
# ensure shape unchanged
|
|
47
|
+
assert var.shape.as_list() == list(shape)
|
|
@@ -0,0 +1,20 @@
|
|
|
1
|
+
import tensorflow as tf
|
|
2
|
+
from tensorflow_riemopt.manifolds import Euclidean
|
|
3
|
+
|
|
4
|
+
|
|
5
|
+
def assign_to_manifold(var, manifold):
|
|
6
|
+
if not hasattr(var, "shape"):
|
|
7
|
+
raise ValueError(
|
|
8
|
+
"var should be a valid variable with a 'shape' attribute"
|
|
9
|
+
)
|
|
10
|
+
if not manifold.check_shape(var):
|
|
11
|
+
raise ValueError("Invalid variable shape {}".format(var.shape))
|
|
12
|
+
setattr(var, "manifold", manifold)
|
|
13
|
+
|
|
14
|
+
|
|
15
|
+
def get_manifold(var, default_manifold=Euclidean()):
|
|
16
|
+
if not hasattr(var, "shape"):
|
|
17
|
+
raise ValueError(
|
|
18
|
+
"var should be a valid variable with a 'shape' attribute"
|
|
19
|
+
)
|
|
20
|
+
return getattr(var, "manifold", default_manifold)
|
|
@@ -1,6 +1,6 @@
|
|
|
1
|
-
Metadata-Version: 2.
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
2
|
Name: tensorflow-riemopt
|
|
3
|
-
Version: 0.
|
|
3
|
+
Version: 0.3.0
|
|
4
4
|
Summary: a library for optimization on Riemannian manifolds
|
|
5
5
|
Home-page: https://github.com/master/tensorflow-riemopt
|
|
6
6
|
Author: Oleg Smirnov
|
|
@@ -10,22 +10,35 @@ Classifier: Development Status :: 4 - Beta
|
|
|
10
10
|
Classifier: Intended Audience :: Science/Research
|
|
11
11
|
Classifier: License :: OSI Approved :: MIT License
|
|
12
12
|
Classifier: Programming Language :: Python :: 3
|
|
13
|
-
Classifier: Programming Language :: Python :: 3.
|
|
14
|
-
Classifier: Programming Language :: Python :: 3.
|
|
15
|
-
Classifier: Programming Language :: Python :: 3.
|
|
13
|
+
Classifier: Programming Language :: Python :: 3.10
|
|
14
|
+
Classifier: Programming Language :: Python :: 3.11
|
|
15
|
+
Classifier: Programming Language :: Python :: 3.12
|
|
16
16
|
Classifier: Topic :: Scientific/Engineering :: Mathematics
|
|
17
17
|
Classifier: Topic :: Software Development :: Libraries :: Python Modules
|
|
18
18
|
Classifier: Topic :: Software Development :: Libraries
|
|
19
19
|
Classifier: Operating System :: OS Independent
|
|
20
|
-
Requires-Python: >=3.
|
|
20
|
+
Requires-Python: >=3.10.0
|
|
21
21
|
Description-Content-Type: text/markdown
|
|
22
22
|
License-File: LICENSE
|
|
23
|
+
Requires-Dist: tensorflow
|
|
24
|
+
Dynamic: author
|
|
25
|
+
Dynamic: author-email
|
|
26
|
+
Dynamic: classifier
|
|
27
|
+
Dynamic: description
|
|
28
|
+
Dynamic: description-content-type
|
|
29
|
+
Dynamic: home-page
|
|
30
|
+
Dynamic: keywords
|
|
31
|
+
Dynamic: license-file
|
|
32
|
+
Dynamic: requires-dist
|
|
33
|
+
Dynamic: requires-python
|
|
34
|
+
Dynamic: summary
|
|
23
35
|
|
|
24
36
|
# TensorFlow RiemOpt
|
|
25
37
|
|
|
26
38
|
[](https://pypi.org/project/tensorflow-riemopt/)
|
|
27
39
|
[](https://arxiv.org/abs/2105.13921)
|
|
28
40
|
[](https://github.com/master/tensorflow-riemopt/actions)
|
|
41
|
+
[](https://tensorflow-riemopt.readthedocs.io)
|
|
29
42
|
[](https://coveralls.io/github/master/tensorflow-riemopt)
|
|
30
43
|
[](https://github.com/python/black)
|
|
31
44
|
[](https://badges.mit-license.org)
|
|
@@ -103,6 +116,7 @@ optimizers for sparse and dense updates in both Eager and Graph modes.
|
|
|
103
116
|
|
|
104
117
|
## Examples
|
|
105
118
|
|
|
119
|
+
- [ANTHEM](https://github.com/amazon-science/hyperbolic-embeddings/tree/main/product_matching) - Choudhary, Nurendra, Nikhil Rao, Sumeet Katariya, Karthik Subbian, and Chandan K. Reddy. "ANTHEM: Attentive Hyperbolic Entity Model for Product Search." In Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining, 2022.
|
|
106
120
|
- [SPDNet](examples/spdnet/) - Huang, Zhiwu, and Luc Van Gool. "A Riemannian network for SPD matrix learning." Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence. AAAI Press, 2017.
|
|
107
121
|
- [LieNet](examples/lienet/) - Huang, Zhiwu, et al. "Deep learning on Lie groups for skeleton-based action recognition." Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.
|
|
108
122
|
- [GrNet](examples/grnet/) - Huang, Zhiwu, Jiqing Wu, and Luc Van Gool. "Building Deep Networks on Grassmann Manifolds." AAAI. AAAI Press, 2018.
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt.egg-info/SOURCES.txt
RENAMED
|
@@ -25,6 +25,7 @@ tensorflow_riemopt/manifolds/grassmannian.py
|
|
|
25
25
|
tensorflow_riemopt/manifolds/grassmannian_test.py
|
|
26
26
|
tensorflow_riemopt/manifolds/hyperboloid.py
|
|
27
27
|
tensorflow_riemopt/manifolds/hyperboloid_test.py
|
|
28
|
+
tensorflow_riemopt/manifolds/mainfold_test.py
|
|
28
29
|
tensorflow_riemopt/manifolds/manifold.py
|
|
29
30
|
tensorflow_riemopt/manifolds/poincare.py
|
|
30
31
|
tensorflow_riemopt/manifolds/poincare_test.py
|
|
@@ -46,4 +47,5 @@ tensorflow_riemopt/optimizers/constrained_rmsprop_test.py
|
|
|
46
47
|
tensorflow_riemopt/optimizers/riemannian_adam.py
|
|
47
48
|
tensorflow_riemopt/optimizers/riemannian_adam_test.py
|
|
48
49
|
tensorflow_riemopt/optimizers/riemannian_gradient_descent.py
|
|
49
|
-
tensorflow_riemopt/optimizers/riemannian_gradient_descent_test.py
|
|
50
|
+
tensorflow_riemopt/optimizers/riemannian_gradient_descent_test.py
|
|
51
|
+
tensorflow_riemopt/optimizers/test_optimizers_manifold.py
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
tensorflow
|
|
@@ -1,20 +0,0 @@
|
|
|
1
|
-
import tensorflow as tf
|
|
2
|
-
|
|
3
|
-
from tensorflow_riemopt.manifolds import Euclidean
|
|
4
|
-
|
|
5
|
-
|
|
6
|
-
def assign_to_manifold(var, manifold):
|
|
7
|
-
if not isinstance(var, tf.Variable):
|
|
8
|
-
raise ValueError("var should be a TensorFlow variable")
|
|
9
|
-
if not manifold.check_shape(var):
|
|
10
|
-
raise ValueError("Invalid variable shape {}".format(var.shape))
|
|
11
|
-
setattr(var, "manifold", manifold)
|
|
12
|
-
|
|
13
|
-
|
|
14
|
-
def get_manifold(var, default_manifold=Euclidean()):
|
|
15
|
-
if not isinstance(var, tf.Variable):
|
|
16
|
-
raise ValueError("var should be a TensorFlow variable")
|
|
17
|
-
if hasattr(var, "manifold"):
|
|
18
|
-
return var.manifold
|
|
19
|
-
else:
|
|
20
|
-
return default_manifold
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/layers/embeddings.py
RENAMED
|
File without changes
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/layers/embeddings_test.py
RENAMED
|
File without changes
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/__init__.py
RENAMED
|
File without changes
|
|
File without changes
|
|
File without changes
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/cholesky_test.py
RENAMED
|
File without changes
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/euclidean_test.py
RENAMED
|
File without changes
|
|
File without changes
|
|
File without changes
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/poincare_test.py
RENAMED
|
File without changes
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/product_test.py
RENAMED
|
File without changes
|
|
File without changes
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/sphere.py
RENAMED
|
File without changes
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/sphere_test.py
RENAMED
|
File without changes
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/stiefel.py
RENAMED
|
File without changes
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/manifolds/stiefel_test.py
RENAMED
|
File without changes
|
|
File without changes
|
|
File without changes
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt/optimizers/__init__.py
RENAMED
|
File without changes
|
|
File without changes
|
|
File without changes
|
{tensorflow-riemopt-0.1.2 → tensorflow_riemopt-0.3.0}/tensorflow_riemopt.egg-info/top_level.txt
RENAMED
|
File without changes
|
|
File without changes
|