robusta-hmf 0.0.1__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.
- robusta_hmf-0.0.1/.gitignore +169 -0
- robusta_hmf-0.0.1/LICENSE +21 -0
- robusta_hmf-0.0.1/PKG-INFO +68 -0
- robusta_hmf-0.0.1/README.md +56 -0
- robusta_hmf-0.0.1/pyproject.toml +37 -0
- robusta_hmf-0.0.1/src/robusta_hmf/__init__.py +11 -0
- robusta_hmf-0.0.1/src/robusta_hmf/_version.py +34 -0
- robusta_hmf-0.0.1/src/robusta_hmf/als.py +45 -0
- robusta_hmf-0.0.1/src/robusta_hmf/convergence.py +64 -0
- robusta_hmf-0.0.1/src/robusta_hmf/frame.py +59 -0
- robusta_hmf-0.0.1/src/robusta_hmf/hmf.py +173 -0
- robusta_hmf-0.0.1/src/robusta_hmf/initialisation.py +91 -0
- robusta_hmf-0.0.1/src/robusta_hmf/likelihoods.py +63 -0
- robusta_hmf-0.0.1/src/robusta_hmf/main.py +478 -0
- robusta_hmf-0.0.1/src/robusta_hmf/py.typed +0 -0
- robusta_hmf-0.0.1/src/robusta_hmf/regularisers.py +23 -0
- robusta_hmf-0.0.1/src/robusta_hmf/rhmf_hogg.py +272 -0
- robusta_hmf-0.0.1/src/robusta_hmf/rotations.py +102 -0
- robusta_hmf-0.0.1/src/robusta_hmf/state.py +37 -0
|
@@ -0,0 +1,169 @@
|
|
|
1
|
+
# vscode
|
|
2
|
+
.vscode/
|
|
3
|
+
|
|
4
|
+
# MacOS
|
|
5
|
+
*.DS_Store
|
|
6
|
+
|
|
7
|
+
# Byte-compiled / optimized / DLL files
|
|
8
|
+
__pycache__/
|
|
9
|
+
*.py[cod]
|
|
10
|
+
*$py.class
|
|
11
|
+
|
|
12
|
+
# C extensions
|
|
13
|
+
*.so
|
|
14
|
+
|
|
15
|
+
# Distribution / packaging
|
|
16
|
+
.Python
|
|
17
|
+
build/
|
|
18
|
+
develop-eggs/
|
|
19
|
+
dist/
|
|
20
|
+
downloads/
|
|
21
|
+
eggs/
|
|
22
|
+
.eggs/
|
|
23
|
+
lib/
|
|
24
|
+
lib64/
|
|
25
|
+
parts/
|
|
26
|
+
sdist/
|
|
27
|
+
var/
|
|
28
|
+
wheels/
|
|
29
|
+
share/python-wheels/
|
|
30
|
+
*.egg-info/
|
|
31
|
+
.installed.cfg
|
|
32
|
+
*.egg
|
|
33
|
+
src/robusta_hmf/_version.py
|
|
34
|
+
# MANIFEST
|
|
35
|
+
|
|
36
|
+
# PyInstaller
|
|
37
|
+
# Usually these files are written by a python script from a template
|
|
38
|
+
# before PyInstaller builds the exe, so as to inject date/other infos into it.
|
|
39
|
+
*.manifest
|
|
40
|
+
*.spec
|
|
41
|
+
|
|
42
|
+
# Installer logs
|
|
43
|
+
pip-log.txt
|
|
44
|
+
pip-delete-this-directory.txt
|
|
45
|
+
|
|
46
|
+
# Unit test / coverage reports
|
|
47
|
+
htmlcov/
|
|
48
|
+
.tox/
|
|
49
|
+
.nox/
|
|
50
|
+
.coverage
|
|
51
|
+
.coverage.*
|
|
52
|
+
.cache
|
|
53
|
+
nosetests.xml
|
|
54
|
+
coverage.xml
|
|
55
|
+
*.cover
|
|
56
|
+
*.py,cover
|
|
57
|
+
.hypothesis/
|
|
58
|
+
.pytest_cache/
|
|
59
|
+
cover/
|
|
60
|
+
|
|
61
|
+
# Translations
|
|
62
|
+
*.mo
|
|
63
|
+
*.pot
|
|
64
|
+
|
|
65
|
+
# Django stuff:
|
|
66
|
+
*.log
|
|
67
|
+
local_settings.py
|
|
68
|
+
db.sqlite3
|
|
69
|
+
db.sqlite3-journal
|
|
70
|
+
|
|
71
|
+
# Flask stuff:
|
|
72
|
+
instance/
|
|
73
|
+
.webassets-cache
|
|
74
|
+
|
|
75
|
+
# Scrapy stuff:
|
|
76
|
+
.scrapy
|
|
77
|
+
|
|
78
|
+
# Sphinx documentation
|
|
79
|
+
docs/_build/
|
|
80
|
+
|
|
81
|
+
# PyBuilder
|
|
82
|
+
.pybuilder/
|
|
83
|
+
target/
|
|
84
|
+
|
|
85
|
+
# Jupyter Notebook
|
|
86
|
+
.ipynb_checkpoints
|
|
87
|
+
|
|
88
|
+
# IPython
|
|
89
|
+
profile_default/
|
|
90
|
+
ipython_config.py
|
|
91
|
+
|
|
92
|
+
# pyenv
|
|
93
|
+
# For a library or package, you might want to ignore these files since the code is
|
|
94
|
+
# intended to run in multiple environments; otherwise, check them in:
|
|
95
|
+
# .python-version
|
|
96
|
+
|
|
97
|
+
# pipenv
|
|
98
|
+
# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
|
|
99
|
+
# However, in case of collaboration, if having platform-specific dependencies or dependencies
|
|
100
|
+
# having no cross-platform support, pipenv may install dependencies that don't work, or not
|
|
101
|
+
# install all needed dependencies.
|
|
102
|
+
#Pipfile.lock
|
|
103
|
+
|
|
104
|
+
# poetry
|
|
105
|
+
# Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
|
|
106
|
+
# This is especially recommended for binary packages to ensure reproducibility, and is more
|
|
107
|
+
# commonly ignored for libraries.
|
|
108
|
+
# https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
|
|
109
|
+
#poetry.lock
|
|
110
|
+
|
|
111
|
+
# pdm
|
|
112
|
+
# Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
|
|
113
|
+
#pdm.lock
|
|
114
|
+
# pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
|
|
115
|
+
# in version control.
|
|
116
|
+
# https://pdm.fming.dev/latest/usage/project/#working-with-version-control
|
|
117
|
+
.pdm.toml
|
|
118
|
+
.pdm-python
|
|
119
|
+
.pdm-build/
|
|
120
|
+
|
|
121
|
+
# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
|
|
122
|
+
__pypackages__/
|
|
123
|
+
|
|
124
|
+
# Celery stuff
|
|
125
|
+
celerybeat-schedule
|
|
126
|
+
celerybeat.pid
|
|
127
|
+
|
|
128
|
+
# SageMath parsed files
|
|
129
|
+
*.sage.py
|
|
130
|
+
|
|
131
|
+
# Environments
|
|
132
|
+
.env
|
|
133
|
+
.venv
|
|
134
|
+
env/
|
|
135
|
+
venv/
|
|
136
|
+
ENV/
|
|
137
|
+
env.bak/
|
|
138
|
+
venv.bak/
|
|
139
|
+
|
|
140
|
+
# Spyder project settings
|
|
141
|
+
.spyderproject
|
|
142
|
+
.spyproject
|
|
143
|
+
|
|
144
|
+
# Rope project settings
|
|
145
|
+
.ropeproject
|
|
146
|
+
|
|
147
|
+
# mkdocs documentation
|
|
148
|
+
/site
|
|
149
|
+
|
|
150
|
+
# mypy
|
|
151
|
+
.mypy_cache/
|
|
152
|
+
.dmypy.json
|
|
153
|
+
dmypy.json
|
|
154
|
+
|
|
155
|
+
# Pyre type checker
|
|
156
|
+
.pyre/
|
|
157
|
+
|
|
158
|
+
# pytype static type analyzer
|
|
159
|
+
.pytype/
|
|
160
|
+
|
|
161
|
+
# Cython debug symbols
|
|
162
|
+
cython_debug/
|
|
163
|
+
|
|
164
|
+
# PyCharm
|
|
165
|
+
# JetBrains specific template is maintained in a separate JetBrains.gitignore that can
|
|
166
|
+
# be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
|
|
167
|
+
# and can be added to the global gitignore or merged into this file. For a more nuclear
|
|
168
|
+
# option (not recommended) you can uncomment the following to ignore the entire idea folder.
|
|
169
|
+
#.idea/
|
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2025 Thomas Hilder
|
|
4
|
+
|
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
+
in the Software without restriction, including without limitation the rights
|
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
10
|
+
furnished to do so, subject to the following conditions:
|
|
11
|
+
|
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
|
13
|
+
copies or substantial portions of the Software.
|
|
14
|
+
|
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
+
SOFTWARE.
|
|
@@ -0,0 +1,68 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: robusta-hmf
|
|
3
|
+
Version: 0.0.1
|
|
4
|
+
Summary: Robust heteroskedastic matrix factorisation in JAX.
|
|
5
|
+
Author-email: Tom Hilder <tom.hilder.dlhp@gmail.com>, "David W. Hogg" <TODO@TODO.com>
|
|
6
|
+
License-File: LICENSE
|
|
7
|
+
Requires-Python: >=3.13
|
|
8
|
+
Requires-Dist: equinox>=0.13.0
|
|
9
|
+
Requires-Dist: jax>=0.7.1
|
|
10
|
+
Requires-Dist: optax>=0.2.5
|
|
11
|
+
Description-Content-Type: text/markdown
|
|
12
|
+
|
|
13
|
+
# Robusta-HMF
|
|
14
|
+
|
|
15
|
+
`jax` implementation of robust heteroskedastic matrix factorisation. Robusta like the coffee bean, get it?
|
|
16
|
+
|
|
17
|
+
## Installation
|
|
18
|
+
|
|
19
|
+
Easiest is from PyPI either with `pip` (CURRENTLY NOT ON PYPI YET)
|
|
20
|
+
|
|
21
|
+
```sh
|
|
22
|
+
pip install robusta-hmf
|
|
23
|
+
```
|
|
24
|
+
|
|
25
|
+
or `uv` (recommended)
|
|
26
|
+
|
|
27
|
+
```sh
|
|
28
|
+
uv add robusta-hmf
|
|
29
|
+
```
|
|
30
|
+
|
|
31
|
+
Or, you can clone and build from source
|
|
32
|
+
|
|
33
|
+
```sh
|
|
34
|
+
git clone git@github.com:TomHilder/robusta-hmf.git
|
|
35
|
+
cd robusta-hmf
|
|
36
|
+
pip install -e .
|
|
37
|
+
```
|
|
38
|
+
|
|
39
|
+
## Usage
|
|
40
|
+
|
|
41
|
+
TODO
|
|
42
|
+
|
|
43
|
+
## Citation
|
|
44
|
+
|
|
45
|
+
TODO
|
|
46
|
+
|
|
47
|
+
## Help
|
|
48
|
+
|
|
49
|
+
TODO
|
|
50
|
+
|
|
51
|
+
## TODOs
|
|
52
|
+
|
|
53
|
+
- [x] Port Hogg's existing code and make sure it builds/installs*
|
|
54
|
+
- [x] Port to `equinox`*
|
|
55
|
+
- [ ] Type checking with `mypy`*
|
|
56
|
+
- [x] Add dependency injection for the following:*
|
|
57
|
+
- [x] Optimisation method, IRLS, SGD (directly optimising objective, see robust_hmf_notes.pdf)
|
|
58
|
+
- [ ] Potentially `dask` and batching support for SGD
|
|
59
|
+
- [x] w-steps. Each w-step corresponds to a different likelihood. Hogg's is Cauchy. We should let this flexible*
|
|
60
|
+
- [x] Initialisation.
|
|
61
|
+
- [x] Re-orientation. Can easily imagine wanting something cheaper for really big data.
|
|
62
|
+
- [ ] Add a save and restore method. Probably avoid pickle/dill and instead encapsulate info in serialisable way and then rebuild model upon loading
|
|
63
|
+
- [ ] Eh maybe, maybe not
|
|
64
|
+
- [x] Tests!*
|
|
65
|
+
- [ ] CI, automated tests, automated relases, and PyPI*
|
|
66
|
+
- [ ] Relax version requirements since uv by default is newest everything
|
|
67
|
+
|
|
68
|
+
(*) = Priority
|
|
@@ -0,0 +1,56 @@
|
|
|
1
|
+
# Robusta-HMF
|
|
2
|
+
|
|
3
|
+
`jax` implementation of robust heteroskedastic matrix factorisation. Robusta like the coffee bean, get it?
|
|
4
|
+
|
|
5
|
+
## Installation
|
|
6
|
+
|
|
7
|
+
Easiest is from PyPI either with `pip` (CURRENTLY NOT ON PYPI YET)
|
|
8
|
+
|
|
9
|
+
```sh
|
|
10
|
+
pip install robusta-hmf
|
|
11
|
+
```
|
|
12
|
+
|
|
13
|
+
or `uv` (recommended)
|
|
14
|
+
|
|
15
|
+
```sh
|
|
16
|
+
uv add robusta-hmf
|
|
17
|
+
```
|
|
18
|
+
|
|
19
|
+
Or, you can clone and build from source
|
|
20
|
+
|
|
21
|
+
```sh
|
|
22
|
+
git clone git@github.com:TomHilder/robusta-hmf.git
|
|
23
|
+
cd robusta-hmf
|
|
24
|
+
pip install -e .
|
|
25
|
+
```
|
|
26
|
+
|
|
27
|
+
## Usage
|
|
28
|
+
|
|
29
|
+
TODO
|
|
30
|
+
|
|
31
|
+
## Citation
|
|
32
|
+
|
|
33
|
+
TODO
|
|
34
|
+
|
|
35
|
+
## Help
|
|
36
|
+
|
|
37
|
+
TODO
|
|
38
|
+
|
|
39
|
+
## TODOs
|
|
40
|
+
|
|
41
|
+
- [x] Port Hogg's existing code and make sure it builds/installs*
|
|
42
|
+
- [x] Port to `equinox`*
|
|
43
|
+
- [ ] Type checking with `mypy`*
|
|
44
|
+
- [x] Add dependency injection for the following:*
|
|
45
|
+
- [x] Optimisation method, IRLS, SGD (directly optimising objective, see robust_hmf_notes.pdf)
|
|
46
|
+
- [ ] Potentially `dask` and batching support for SGD
|
|
47
|
+
- [x] w-steps. Each w-step corresponds to a different likelihood. Hogg's is Cauchy. We should let this flexible*
|
|
48
|
+
- [x] Initialisation.
|
|
49
|
+
- [x] Re-orientation. Can easily imagine wanting something cheaper for really big data.
|
|
50
|
+
- [ ] Add a save and restore method. Probably avoid pickle/dill and instead encapsulate info in serialisable way and then rebuild model upon loading
|
|
51
|
+
- [ ] Eh maybe, maybe not
|
|
52
|
+
- [x] Tests!*
|
|
53
|
+
- [ ] CI, automated tests, automated relases, and PyPI*
|
|
54
|
+
- [ ] Relax version requirements since uv by default is newest everything
|
|
55
|
+
|
|
56
|
+
(*) = Priority
|
|
@@ -0,0 +1,37 @@
|
|
|
1
|
+
[project]
|
|
2
|
+
name = "robusta-hmf"
|
|
3
|
+
dynamic = ["version"]
|
|
4
|
+
description = "Robust heteroskedastic matrix factorisation in JAX."
|
|
5
|
+
readme = "README.md"
|
|
6
|
+
authors = [
|
|
7
|
+
{ name = "Tom Hilder", email = "tom.hilder.dlhp@gmail.com" },
|
|
8
|
+
{ name = "David W. Hogg", email = "TODO@TODO.com" },
|
|
9
|
+
]
|
|
10
|
+
requires-python = ">=3.13"
|
|
11
|
+
dependencies = ["equinox>=0.13.0", "jax>=0.7.1", "optax>=0.2.5"]
|
|
12
|
+
|
|
13
|
+
[build-system]
|
|
14
|
+
requires = ["hatchling", "hatch-vcs"]
|
|
15
|
+
build-backend = "hatchling.build"
|
|
16
|
+
|
|
17
|
+
[tool.hatch.build]
|
|
18
|
+
include = ["src/robusta_hmf/py.typed"]
|
|
19
|
+
exclude = ["**/__pycache__/**"]
|
|
20
|
+
|
|
21
|
+
[tool.hatch.build.targets.sdist]
|
|
22
|
+
include = ["src/**", "README.md", "LICENSE"]
|
|
23
|
+
|
|
24
|
+
[tool.hatch.build.targets.wheel]
|
|
25
|
+
packages = ["src/robusta_hmf"]
|
|
26
|
+
|
|
27
|
+
[tool.hatch.build.hooks.vcs]
|
|
28
|
+
version-file = "src/robusta_hmf/_version.py"
|
|
29
|
+
|
|
30
|
+
[tool.hatch.version]
|
|
31
|
+
source = "vcs"
|
|
32
|
+
|
|
33
|
+
[tool.ruff]
|
|
34
|
+
line-length = 99 # Sorry PEP8 I'm a rebel
|
|
35
|
+
|
|
36
|
+
[dependency-groups]
|
|
37
|
+
dev = ["mypy>=1.18.2", "pytest>=8.4.2", "pytest-cov>=7.0.0"]
|
|
@@ -0,0 +1,34 @@
|
|
|
1
|
+
# file generated by setuptools-scm
|
|
2
|
+
# don't change, don't track in version control
|
|
3
|
+
|
|
4
|
+
__all__ = [
|
|
5
|
+
"__version__",
|
|
6
|
+
"__version_tuple__",
|
|
7
|
+
"version",
|
|
8
|
+
"version_tuple",
|
|
9
|
+
"__commit_id__",
|
|
10
|
+
"commit_id",
|
|
11
|
+
]
|
|
12
|
+
|
|
13
|
+
TYPE_CHECKING = False
|
|
14
|
+
if TYPE_CHECKING:
|
|
15
|
+
from typing import Tuple
|
|
16
|
+
from typing import Union
|
|
17
|
+
|
|
18
|
+
VERSION_TUPLE = Tuple[Union[int, str], ...]
|
|
19
|
+
COMMIT_ID = Union[str, None]
|
|
20
|
+
else:
|
|
21
|
+
VERSION_TUPLE = object
|
|
22
|
+
COMMIT_ID = object
|
|
23
|
+
|
|
24
|
+
version: str
|
|
25
|
+
__version__: str
|
|
26
|
+
__version_tuple__: VERSION_TUPLE
|
|
27
|
+
version_tuple: VERSION_TUPLE
|
|
28
|
+
commit_id: COMMIT_ID
|
|
29
|
+
__commit_id__: COMMIT_ID
|
|
30
|
+
|
|
31
|
+
__version__ = version = '0.0.1'
|
|
32
|
+
__version_tuple__ = version_tuple = (0, 0, 1)
|
|
33
|
+
|
|
34
|
+
__commit_id__ = commit_id = None
|
|
@@ -0,0 +1,45 @@
|
|
|
1
|
+
# als.py
|
|
2
|
+
|
|
3
|
+
import equinox as eqx
|
|
4
|
+
import jax
|
|
5
|
+
import jax.numpy as jnp
|
|
6
|
+
|
|
7
|
+
from .state import RHMFState, update_state
|
|
8
|
+
|
|
9
|
+
|
|
10
|
+
class WeightedAStep(eqx.Module):
|
|
11
|
+
ridge: float | None = eqx.field(static=True, default=None)
|
|
12
|
+
|
|
13
|
+
def __call__(self, Y, W, state: RHMFState):
|
|
14
|
+
G = state.G
|
|
15
|
+
|
|
16
|
+
def solve_row(y_i, w_i):
|
|
17
|
+
s = jnp.sqrt(w_i)
|
|
18
|
+
Gw = G * s[:, None]
|
|
19
|
+
M = Gw.T @ Gw
|
|
20
|
+
if self.ridge is not None:
|
|
21
|
+
M = M + self.ridge * jnp.eye(G.shape[1], dtype=G.dtype)
|
|
22
|
+
b = Gw.T @ (s * y_i)
|
|
23
|
+
return jnp.linalg.solve(M, b)
|
|
24
|
+
|
|
25
|
+
A_new = jax.vmap(solve_row)(Y, W) # [N, K]
|
|
26
|
+
return update_state(state, A=A_new)
|
|
27
|
+
|
|
28
|
+
|
|
29
|
+
class WeightedGStep(eqx.Module):
|
|
30
|
+
ridge: float | None = eqx.field(static=True, default=None)
|
|
31
|
+
|
|
32
|
+
def __call__(self, Y, W, state: RHMFState):
|
|
33
|
+
A = state.A
|
|
34
|
+
|
|
35
|
+
def solve_col(y_j, w_j):
|
|
36
|
+
s = jnp.sqrt(w_j)
|
|
37
|
+
Aw = A * s[:, None]
|
|
38
|
+
M = Aw.T @ Aw
|
|
39
|
+
if self.ridge is not None:
|
|
40
|
+
M = M + self.ridge * jnp.eye(A.shape[1], dtype=A.dtype)
|
|
41
|
+
b = Aw.T @ (s * y_j)
|
|
42
|
+
return jnp.linalg.solve(M, b)
|
|
43
|
+
|
|
44
|
+
G_new = jax.vmap(solve_col)(Y.T, W.T) # [D, K]
|
|
45
|
+
return update_state(state, G=G_new)
|
|
@@ -0,0 +1,64 @@
|
|
|
1
|
+
# convergence.py
|
|
2
|
+
|
|
3
|
+
from dataclasses import dataclass, field
|
|
4
|
+
from typing import Literal, get_args
|
|
5
|
+
|
|
6
|
+
import jax
|
|
7
|
+
import jax.numpy as jnp
|
|
8
|
+
from jaxtyping import Array
|
|
9
|
+
|
|
10
|
+
from .state import RHMFState
|
|
11
|
+
|
|
12
|
+
jax.config.update("jax_enable_x64", True)
|
|
13
|
+
|
|
14
|
+
ConvStrategy = Literal["max_frac_G", "max_frac_A", "rel_frac_loss", "none"]
|
|
15
|
+
|
|
16
|
+
DEFAULT_STRATEGY = "max_frac_G"
|
|
17
|
+
DEFAULT_TOL = 1e-6
|
|
18
|
+
|
|
19
|
+
|
|
20
|
+
def max_frac_mat(old_mat: Array, new_mat: Array, tol: float) -> bool:
|
|
21
|
+
d_mat = old_mat - new_mat
|
|
22
|
+
if jnp.max(d_mat * d_mat) / jnp.mean(new_mat * new_mat) < tol:
|
|
23
|
+
return True
|
|
24
|
+
else:
|
|
25
|
+
return False
|
|
26
|
+
|
|
27
|
+
|
|
28
|
+
def rel_frac_loss(old_loss: float, new_loss: float, tol: float) -> bool:
|
|
29
|
+
if jnp.abs(old_loss - new_loss) / jnp.abs(new_loss) < tol:
|
|
30
|
+
return True
|
|
31
|
+
else:
|
|
32
|
+
return False
|
|
33
|
+
|
|
34
|
+
|
|
35
|
+
def never_converged(*args, **kwargs) -> bool:
|
|
36
|
+
return False
|
|
37
|
+
|
|
38
|
+
|
|
39
|
+
@dataclass(frozen=True)
|
|
40
|
+
class ConvergenceTester:
|
|
41
|
+
strategy: ConvStrategy = field(default=DEFAULT_STRATEGY)
|
|
42
|
+
tol: float = field(default=DEFAULT_TOL)
|
|
43
|
+
|
|
44
|
+
def __post_init__(self):
|
|
45
|
+
if self.strategy not in get_args(ConvStrategy):
|
|
46
|
+
raise ValueError(
|
|
47
|
+
f"Invalid strategy: {self.strategy}. Must be one of {get_args(ConvStrategy)}."
|
|
48
|
+
)
|
|
49
|
+
|
|
50
|
+
def is_converged(
|
|
51
|
+
self,
|
|
52
|
+
old_state: RHMFState,
|
|
53
|
+
new_state: RHMFState,
|
|
54
|
+
old_loss: float,
|
|
55
|
+
new_loss: float,
|
|
56
|
+
) -> bool:
|
|
57
|
+
if self.strategy == "max_frac_G":
|
|
58
|
+
return max_frac_mat(old_state.G, new_state.G, tol=self.tol)
|
|
59
|
+
elif self.strategy == "max_frac_A":
|
|
60
|
+
return max_frac_mat(old_state.A, new_state.A, tol=self.tol)
|
|
61
|
+
elif self.strategy == "rel_frac_loss":
|
|
62
|
+
return rel_frac_loss(old_loss, new_loss, tol=self.tol)
|
|
63
|
+
elif self.strategy == "none":
|
|
64
|
+
return never_converged()
|
|
@@ -0,0 +1,59 @@
|
|
|
1
|
+
# frame.py
|
|
2
|
+
|
|
3
|
+
from copy import deepcopy
|
|
4
|
+
from dataclasses import dataclass
|
|
5
|
+
|
|
6
|
+
import jax.numpy as jnp
|
|
7
|
+
from jaxtyping import Array
|
|
8
|
+
|
|
9
|
+
from .convergence import ConvergenceTester
|
|
10
|
+
from .hmf import HMF
|
|
11
|
+
from .state import RHMFState
|
|
12
|
+
|
|
13
|
+
|
|
14
|
+
@dataclass
|
|
15
|
+
class OptFrame:
|
|
16
|
+
method: HMF
|
|
17
|
+
conv_tester: ConvergenceTester
|
|
18
|
+
|
|
19
|
+
def run(
|
|
20
|
+
self,
|
|
21
|
+
Y: Array,
|
|
22
|
+
W: Array,
|
|
23
|
+
init_state: RHMFState,
|
|
24
|
+
rotation_cadence: int,
|
|
25
|
+
conv_check_cadence: int,
|
|
26
|
+
max_iter: int,
|
|
27
|
+
):
|
|
28
|
+
# Get the step function from the method
|
|
29
|
+
step_fn = self.method.get_stepper()
|
|
30
|
+
|
|
31
|
+
# Always rotate for als
|
|
32
|
+
if self.method.opt_method == "als":
|
|
33
|
+
rotation_cadence = 1
|
|
34
|
+
|
|
35
|
+
# To store the loss history
|
|
36
|
+
loss_history = []
|
|
37
|
+
|
|
38
|
+
# Initialise states and loss
|
|
39
|
+
state = init_state
|
|
40
|
+
prev_state = deepcopy(init_state)
|
|
41
|
+
prev_loss = jnp.inf
|
|
42
|
+
|
|
43
|
+
# Training loop
|
|
44
|
+
for i in range(max_iter):
|
|
45
|
+
# Do we rotate this iteration?
|
|
46
|
+
rot = True if (i % rotation_cadence == 0 and i != 0) else False
|
|
47
|
+
# Take an optimization step and record the loss
|
|
48
|
+
state, loss = step_fn(Y=Y, W_data=W, state=state, rotate=rot)
|
|
49
|
+
loss_history.append(loss)
|
|
50
|
+
# Check convergence and print loss every conv_check_cadence iterations
|
|
51
|
+
if i % conv_check_cadence == 0 and i != 0:
|
|
52
|
+
if self.conv_tester.is_converged(prev_state, state, prev_loss, loss):
|
|
53
|
+
print(f"Converged at iteration {i}")
|
|
54
|
+
break
|
|
55
|
+
prev_state = deepcopy(state)
|
|
56
|
+
prev_loss = loss
|
|
57
|
+
print(f"iter {state.it:03d} | loss {loss:.4f}", flush=True)
|
|
58
|
+
|
|
59
|
+
return state, jnp.array(loss_history)
|