synthyverse 0.2.7__tar.gz → 0.2.9__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.
Files changed (113) hide show
  1. {synthyverse-0.2.7 → synthyverse-0.2.9}/PKG-INFO +8 -3
  2. {synthyverse-0.2.7 → synthyverse-0.2.9}/THIRD_PARTY_NOTICES.md +66 -6
  3. {synthyverse-0.2.7 → synthyverse-0.2.9}/pyproject.toml +7 -2
  4. synthyverse-0.2.9/synthyverse/benchmark/__init__.py +14 -0
  5. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/benchmark/synthesis.py +149 -71
  6. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/evaluation/__init__.py +2 -0
  7. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/evaluation/eval.py +43 -32
  8. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/evaluation/fidelity.py +89 -6
  9. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/evaluation/ml.py +87 -28
  10. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/evaluation/privacy.py +4 -1
  11. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/__init__.py +39 -22
  12. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/_optional.py +13 -1
  13. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/arf_generator/model.py +111 -80
  14. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/base.py +91 -58
  15. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/cdtd_generator/cdtd.py +114 -43
  16. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/cdtd_generator/layers.py +16 -13
  17. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/cdtd_generator/utils.py +1 -5
  18. synthyverse-0.2.9/synthyverse/generators/config.py +431 -0
  19. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/ctgan_generator/ct_gan.py +96 -17
  20. synthyverse-0.2.9/synthyverse/generators/ctgan_generator/data_transformer.py +284 -0
  21. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/ctgan_generator/synthesizer.py +15 -6
  22. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/dgm_utils.py +36 -15
  23. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/forestdiffusion_generator/fd.py +17 -5
  24. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/forestdiffusion_generator/model.py +23 -2
  25. synthyverse-0.2.9/synthyverse/generators/tabargn_generator/__init__.py +3 -0
  26. synthyverse-0.2.9/synthyverse/generators/tabargn_generator/encoding.py +200 -0
  27. synthyverse-0.2.9/synthyverse/generators/tabargn_generator/model.py +135 -0
  28. synthyverse-0.2.9/synthyverse/generators/tabargn_generator/tabargn.py +495 -0
  29. synthyverse-0.2.9/synthyverse/generators/tabbyflow_generator/__init__.py +1 -0
  30. synthyverse-0.2.9/synthyverse/generators/tabbyflow_generator/flow_model.py +221 -0
  31. synthyverse-0.2.9/synthyverse/generators/tabbyflow_generator/tabbyflow.py +498 -0
  32. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabcascade_generator/highres.py +1 -1
  33. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabcascade_generator/lowres.py +1 -1
  34. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabcascade_generator/tabcascade.py +145 -58
  35. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabddpm_generator/gaussian_multinomial_diffsuion.py +6 -2
  36. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabddpm_generator/tabddpm.py +150 -46
  37. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabdiff_generator/tabdiff.py +109 -59
  38. synthyverse-0.2.9/synthyverse/generators/tabpfn_generator/__init__.py +1 -0
  39. synthyverse-0.2.9/synthyverse/generators/tabpfn_generator/tabpfn.py +61 -0
  40. synthyverse-0.2.9/synthyverse/generators/tabpfn_generator/tabpfn_extensions/unsupervised.py +779 -0
  41. synthyverse-0.2.9/synthyverse/generators/tabpfn_generator/tabpfn_extensions/utils.py +621 -0
  42. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabsyn_generator/tabsyn.py +126 -68
  43. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tvae_generator/synthesizer.py +18 -6
  44. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tvae_generator/tvae.py +101 -17
  45. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/univariate_generator/univariate.py +22 -7
  46. synthyverse-0.2.9/synthyverse/generators/utrees_generator/__init__.py +1 -0
  47. synthyverse-0.2.9/synthyverse/generators/utrees_generator/_utrees/__init__.py +3 -0
  48. synthyverse-0.2.9/synthyverse/generators/utrees_generator/_utrees/baltobot.py +288 -0
  49. synthyverse-0.2.9/synthyverse/generators/utrees_generator/_utrees/kdi_quantizer.py +353 -0
  50. synthyverse-0.2.9/synthyverse/generators/utrees_generator/_utrees/unmasking_trees.py +381 -0
  51. synthyverse-0.2.9/synthyverse/generators/utrees_generator/kditransform/__init__.py +5 -0
  52. synthyverse-0.2.9/synthyverse/generators/utrees_generator/kditransform/kdi_discretizer.py +714 -0
  53. synthyverse-0.2.9/synthyverse/generators/utrees_generator/kditransform/kdi_transformer.py +576 -0
  54. synthyverse-0.2.9/synthyverse/generators/utrees_generator/kditransform/ksum.py +230 -0
  55. synthyverse-0.2.9/synthyverse/generators/utrees_generator/utrees.py +71 -0
  56. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/xgbddpm_generator/xgbddpm.py +65 -31
  57. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse.egg-info/PKG-INFO +8 -3
  58. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse.egg-info/SOURCES.txt +23 -0
  59. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse.egg-info/requires.txt +5 -0
  60. synthyverse-0.2.7/synthyverse/benchmark/__init__.py +0 -9
  61. {synthyverse-0.2.7 → synthyverse-0.2.9}/LICENSE +0 -0
  62. {synthyverse-0.2.7 → synthyverse-0.2.9}/LICENSES/Apache-2.0.txt +0 -0
  63. {synthyverse-0.2.7 → synthyverse-0.2.9}/LICENSES/CTGAN-BSL-1.1.txt +0 -0
  64. {synthyverse-0.2.7 → synthyverse-0.2.9}/LICENSES/MIT.txt +0 -0
  65. {synthyverse-0.2.7 → synthyverse-0.2.9}/MANIFEST.in +0 -0
  66. {synthyverse-0.2.7 → synthyverse-0.2.9}/NOTICE +0 -0
  67. {synthyverse-0.2.7 → synthyverse-0.2.9}/README.md +0 -0
  68. {synthyverse-0.2.7 → synthyverse-0.2.9}/logo/logo.png +0 -0
  69. {synthyverse-0.2.7 → synthyverse-0.2.9}/setup.cfg +0 -0
  70. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/__init__.py +0 -0
  71. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/evaluation/base.py +0 -0
  72. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/evaluation/distances.py +0 -0
  73. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/evaluation/preprocessing.py +0 -0
  74. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/evaluation/utility.py +0 -0
  75. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/arf_generator/__init__.py +0 -0
  76. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/arf_generator/arf.py +0 -0
  77. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/arf_generator/utils.py +0 -0
  78. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/cdtd_generator/__init__.py +0 -0
  79. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/ctgan_generator/__init__.py +0 -0
  80. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/forestdiffusion_generator/__init__.py +0 -0
  81. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/forestdiffusion_generator/utils/__init__.py +0 -0
  82. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/forestdiffusion_generator/utils/diffusion.py +0 -0
  83. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/forestdiffusion_generator/utils/utils_diffusion.py +0 -0
  84. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/forestdiffusion_generator/utils/writer.py +0 -0
  85. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/persistence.py +0 -0
  86. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/smote_generator/__init__.py +0 -0
  87. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/smote_generator/smote.py +0 -0
  88. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/synthpop_generator/__init__.py +0 -0
  89. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/synthpop_generator/synthpop.py +0 -0
  90. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabcascade_generator/__init__.py +0 -0
  91. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabcascade_generator/disttree.py +0 -0
  92. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabcascade_generator/encoder.py +0 -0
  93. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabddpm_generator/__init__.py +0 -0
  94. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabddpm_generator/ddpm_utils.py +0 -0
  95. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabddpm_generator/modules.py +0 -0
  96. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabdiff_generator/__init__.py +0 -0
  97. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabdiff_generator/diffusion.py +0 -0
  98. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabdiff_generator/modules.py +0 -0
  99. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabdiff_generator/noise_schedule.py +0 -0
  100. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabdiff_generator/transformer.py +0 -0
  101. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabsyn_generator/__init__.py +0 -0
  102. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabsyn_generator/diffusion.py +0 -0
  103. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tabsyn_generator/vae.py +0 -0
  104. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/tvae_generator/__init__.py +0 -0
  105. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/univariate_generator/__init__.py +0 -0
  106. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/xgbddpm_generator/__init__.py +0 -0
  107. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/xgbdiffusion_generator/__init__.py +0 -0
  108. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/generators/xgbdiffusion_generator/xgbdiffusion.py +0 -0
  109. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/utils/__init__.py +0 -0
  110. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/utils/reproducibility.py +0 -0
  111. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse/utils/utils.py +0 -0
  112. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse.egg-info/dependency_links.txt +0 -0
  113. {synthyverse-0.2.7 → synthyverse-0.2.9}/synthyverse.egg-info/top_level.txt +0 -0
@@ -1,6 +1,6 @@
1
- Metadata-Version: 2.1
1
+ Metadata-Version: 2.4
2
2
  Name: synthyverse
3
- Version: 0.2.7
3
+ Version: 0.2.9
4
4
  Summary: Synthetic data generation and evaluation library
5
5
  Author: Jim Achterberg, Saif Ul Islam, Zia Ur Rehman
6
6
  License: MIT License
@@ -18,7 +18,7 @@ Classifier: Programming Language :: Python :: 3.11
18
18
  Classifier: Programming Language :: Python :: 3.12
19
19
  Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
20
20
  Classifier: Topic :: Software Development :: Libraries :: Python Modules
21
- Requires-Python: >=3.8
21
+ Requires-Python: >=3.12
22
22
  Description-Content-Type: text/markdown
23
23
  License-File: LICENSE
24
24
  License-File: NOTICE
@@ -41,8 +41,12 @@ Requires-Dist: torch-ema
41
41
  Requires-Dist: einops
42
42
  Requires-Dist: imbalanced-learn
43
43
  Requires-Dist: xgb-diffusion
44
+ Requires-Dist: torchdiffeq>=0.2.5
45
+ Requires-Dist: structured-data-models>=0.2.1
44
46
  Provides-Extra: ctgan
45
47
  Requires-Dist: ctgan==0.12.0; extra == "ctgan"
48
+ Provides-Extra: tabpfn
49
+ Requires-Dist: tabpfn>=9.0.0; extra == "tabpfn"
46
50
  Provides-Extra: docs
47
51
  Requires-Dist: sphinx; extra == "docs"
48
52
  Requires-Dist: furo; extra == "docs"
@@ -51,6 +55,7 @@ Requires-Dist: sphinx-copybutton; extra == "docs"
51
55
  Requires-Dist: sphinx-autodoc-typehints; extra == "docs"
52
56
  Requires-Dist: sphinxcontrib-mermaid; extra == "docs"
53
57
  Requires-Dist: linkify-it-py; extra == "docs"
58
+ Dynamic: license-file
54
59
 
55
60
  <table align="center" border="0">
56
61
  <tr>
@@ -16,7 +16,7 @@ The MIT license text for these adapted implementations is kept in
16
16
  - Upstream license: MIT License
17
17
  - Upstream copyright notice: `Copyright (c) 2023 Kristin Blesch, Marvin Wright`
18
18
  - Imported or last compared against: `Commit 8b63c1b`
19
- - Modifications: `Fallback when nodes contain single value.`
19
+ - Modifications: `Fallback when nodes contain single value, and fast vectorized sampling.`
20
20
  - Upstream NOTICE text, if any: `None`
21
21
 
22
22
  ### CDTDGenerator
@@ -46,7 +46,7 @@ The MIT license text for these adapted implementations is kept in
46
46
  - Upstream license: MIT License
47
47
  - Upstream copyright notice: `Copyright 2024 Minkai Xu`
48
48
  - Imported or last compared against: `Commit 5ecdb33`
49
- - Modifications: `Refactored API, speed-up masked diffusion by vectorizing loops, fast tensor dataloader.`
49
+ - Modifications: `Refactored API, speed-up masked diffusion by vectorizing loops, fast tensor dataloader, use increasing EMA decay.`
50
50
  - Upstream NOTICE text, if any: `None`
51
51
 
52
52
  ### TabCascadeGenerator
@@ -56,7 +56,7 @@ The MIT license text for these adapted implementations is kept in
56
56
  - Upstream license: MIT License
57
57
  - Upstream copyright notice: `Copyright (c) 2026 Markus Mueller`
58
58
  - Imported or last compared against: `Commit c8f44bb`
59
- - Modifications: `Refactored API, simple ordinal encoding for categories.`
59
+ - Modifications: `Refactored API.`
60
60
  - Upstream NOTICE text, if any: `None`
61
61
 
62
62
  ### ForestDiffusionGenerator
@@ -69,11 +69,51 @@ The MIT license text for these adapted implementations is kept in
69
69
  - Modifications: `Refactored API, allow storing XGBoost models in-memory.`
70
70
  - Upstream NOTICE text, if any: `None`
71
71
 
72
+ ### TabbyFlowGenerator
73
+
74
+ - Synthyverse files: `synthyverse/generators/tabbyflow_generator/*`
75
+ - Upstream project: `ForestDiffusion-MO (https://github.com/andresguzco/ef-vfm/)`
76
+ - Upstream license: MIT License
77
+ - Upstream copyright notice: `Copyright 2024 Andrés Guzmán-Cordero`
78
+ - Imported or last compared against: `Commit 11fa189`
79
+ - Modifications: `Refactored API, fast tensor dataloader, switch adaptive sampler to euler sampler, remove option to remove NaN outputs, use increasing EMA decay.`
80
+ - Upstream NOTICE text, if any: `None`
81
+
72
82
  ## Apache-2.0-licensed adapted implementations
73
83
 
74
84
  The Apache-2.0 license text for these adapted implementations is kept in
75
85
  `LICENSES/Apache-2.0.txt`.
76
86
 
87
+ ### UTreesGenerator
88
+
89
+ - Synthyverse files: `synthyverse/generators/utrees_generator/*`
90
+ - Upstream project: `utrees (https://github.com/calvinmccarter/unmasking-trees/)`
91
+ - Upstream license: Apache License, Version 2.0
92
+ - Upstream copyright notice: `Copyright 2024 Calvin McCarter`
93
+ - Imported or last compared against: `version 0.3.0`
94
+ - Modifications: `Refactored API, omitting TabPFN functionality.`
95
+ - Upstream NOTICE text, if any: `None`
96
+
97
+ ### KDITransform
98
+
99
+ - Synthyverse files: `synthyverse/generators/utrees_generator/kditransform/*`
100
+ - Upstream project: `kditransform (https://github.com/calvinmccarter/kditransform/)`
101
+ - Upstream license: Apache License, Version 2.0
102
+ - Upstream copyright notice: `Copyright 2023-2025 Calvin McCarter`
103
+ - Imported or last compared against: `version 1.1.0`
104
+ - Modifications: `Add support for scikit-learn>=1.6.`
105
+ - Upstream NOTICE text, if any: `None`
106
+
107
+ ### TabARGNGenerator
108
+
109
+ - Synthyverse files: `synthyverse/generators/tabargn_generator/*`
110
+ - Upstream project: `mostlyai-engine (https://github.com/mostly-ai/mostlyai-engine)`
111
+ - Upstream license: Apache License, Version 2.0
112
+ - Upstream copyright notice: `Copyright 2025 MOSTLY AI`
113
+ - Imported or last compared against: `version 2.7.0`
114
+ - Modifications: `Refactored API, omitting several optional features from the upstream engine.`
115
+ - Upstream NOTICE text, if any: `None`
116
+
77
117
  ### TabDDPMGenerator
78
118
 
79
119
  - Synthyverse files: `synthyverse/generators/tabddpm_generator/*`
@@ -81,7 +121,7 @@ The Apache-2.0 license text for these adapted implementations is kept in
81
121
  - Upstream license: Apache License, Version 2.0
82
122
  - Upstream copyright notice: `Copyright vanderschaarlab 2023`
83
123
  - Imported or last compared against: `Commit 23f322f`
84
- - Modifications: `Refactored API, allow manual specification of categorical features, fast tensor dataloader.`
124
+ - Modifications: `Refactored API, allow manual specification of categorical features, fast tensor dataloader, anneal learning rate by optimizer steps instead of epochs.`
85
125
  - Upstream NOTICE text, if any: `None`
86
126
 
87
127
  ### TabSynGenerator
@@ -94,6 +134,26 @@ The Apache-2.0 license text for these adapted implementations is kept in
94
134
  - Modifications: `Refactor API, fast tensor dataloader.`
95
135
  - Upstream NOTICE text, if any: `Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.`
96
136
 
137
+ ### TabPFNGenerator
138
+
139
+ - Synthyverse files: `synthyverse/generators/tabpfn_generator/*`
140
+ - Upstream project: `TabPFN (https://github.com/PriorLabs/TabPFN/)`
141
+ - Upstream license: Apache License, Version 2.0
142
+ - Upstream copyright notice: `Copyright (c) 2026 Prior Labs GmbH`
143
+ - Imported or last compared against: `Commit 15f5e6b`
144
+ - Modifications: `None; wrapped API.`
145
+ - Upstream NOTICE text, if any: `TabPFN. Copyright (c) 2026 Prior Labs GmbH. Third-party attribution notices are provided in THIRD-PARTY-NOTICES.md.`
146
+
147
+ ### TabPFN-extensions
148
+
149
+ - Synthyverse files: `synthyverse/generators/tabpfn_generator/tabpfn_extensions/*`
150
+ - Upstream project: `tabpfn-extensions (https://github.com/priorlabs/tabpfn-extensions)`
151
+ - Upstream license: Apache License, Version 2.0
152
+ - Upstream copyright notice: `Copyright (c) 2026 Prior Labs GmbH`
153
+ - Imported or last compared against: `Commit abcafe3`
154
+ - Modifications: `None.`
155
+ - Upstream NOTICE text, if any: `None`
156
+
97
157
  ### AlphaPrecisionBetaRecall fidelity metric
98
158
 
99
159
  - Synthyverse files: `synthyverse/evaluation/fidelity.py`
@@ -101,7 +161,7 @@ The Apache-2.0 license text for these adapted implementations is kept in
101
161
  - Upstream license: Apache License, Version 2.0
102
162
  - Upstream copyright notice: `Copyright vanderschaarlab 2023`
103
163
  - Imported or last compared against: `Commit 23f322f`
104
- - Modifications: `Refactored API, allow manual specification k in Beta-Recall.`
164
+ - Modifications: `Refactored API, allow manual specification k in Beta-Recall, and use Gower distance.`
105
165
  - Upstream NOTICE text, if any: `None`
106
166
 
107
167
 
@@ -118,7 +178,7 @@ not under the synthyverse MIT License.
118
178
  - Upstream copyright notice: `The Licensed Work is (c) DataCebo, Inc.; license text copyright (c) 2017 MariaDB Corporation Ab, All Rights Reserved.`
119
179
  - License text: `LICENSES/CTGAN-BSL-1.1.txt`
120
180
  - Imported or last compared against: `version 0.12.0`
121
- - Modifications: `No ctgan source code is vendored.`
181
+ - Modifications: `enable time-based early stopping, and validation every k steps/epochs.`
122
182
  - Upstream NOTICE text, if any: `The Business Source License (this document, or the "License") is not an Open Source license. However, the Licensed Work will eventually be made available under an Open Source License, as stated in this License.`
123
183
 
124
184
  ## Non-vendored dependencies
@@ -4,10 +4,10 @@ build-backend = "setuptools.build_meta"
4
4
 
5
5
  [project]
6
6
  name = "synthyverse"
7
- version = "0.2.7"
7
+ version = "0.2.9"
8
8
  description = "Synthetic data generation and evaluation library"
9
9
  readme = "README.md"
10
- requires-python = ">=3.8"
10
+ requires-python = ">=3.12"
11
11
  license = {text = "MIT License"}
12
12
  authors = [
13
13
  {name = "Jim Achterberg"},
@@ -44,6 +44,8 @@ dependencies = [
44
44
  "einops",
45
45
  "imbalanced-learn",
46
46
  "xgb-diffusion",
47
+ "torchdiffeq>=0.2.5",
48
+ "structured-data-models>=0.2.1",
47
49
  ]
48
50
 
49
51
  [tool.setuptools]
@@ -63,6 +65,9 @@ Documentation = "https://synthyverse.readthedocs.io"
63
65
  ctgan = [
64
66
  "ctgan==0.12.0",
65
67
  ]
68
+ tabpfn = [
69
+ "tabpfn>=9.0.0",
70
+ ]
66
71
  docs = [
67
72
  "sphinx",
68
73
  "furo",
@@ -0,0 +1,14 @@
1
+ from importlib import import_module
2
+
3
+
4
+ def __getattr__(name: str):
5
+ if name in globals():
6
+ return globals()[name]
7
+ if name == "TabularSynthesisBenchmark":
8
+ benchmark = import_module(".synthesis", __name__).TabularSynthesisBenchmark
9
+ globals()[name] = benchmark
10
+ return benchmark
11
+ raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
12
+
13
+
14
+ __all__ = ["TabularSynthesisBenchmark"]
@@ -13,10 +13,10 @@ from typing import Any, Dict, Iterable, Optional, Union
13
13
  import pandas as pd
14
14
  from sklearn.model_selection import train_test_split
15
15
 
16
+ from synthyverse import evaluation
16
17
  from synthyverse.generators import DataProcessor, get_generator
17
18
  from synthyverse.utils.utils import memory_guarded
18
19
 
19
-
20
20
  RESULT_COLUMNS = ["metric name", "metric value", "train_seed", "set"]
21
21
  BYTES_PER_MIB = 1024 * 1024
22
22
  _PROCESS_MEMORY_READER = None
@@ -36,10 +36,9 @@ class TabularSynthesisBenchmark:
36
36
  * call ``train()`` and ``eval()`` separately to reuse saved models or run
37
37
  different metrics later.
38
38
 
39
- When ``model_save_dir`` is provided, the benchmark stores the fitted data
40
- processor and trained generator there. When ``model_save_dir`` is ``None``,
41
- models and processors are not saved. Results are returned as a long-format
42
- ``pandas.DataFrame`` and can also be written to CSV.
39
+ When ``model_save_dir`` is provided, the benchmark caches fitted processors
40
+ there and saves trained generators when ``save_models=True``. Results are
41
+ returned as a long-format ``pandas.DataFrame`` and can also be written to CSV.
43
42
 
44
43
  Args:
45
44
  X (pd.DataFrame): Real tabular dataset to benchmark on.
@@ -52,10 +51,11 @@ class TabularSynthesisBenchmark:
52
51
  target_column (str or None): Column used for supervised metrics and,
53
52
  when categorical, stratified splits. Use ``None`` when there is no
54
53
  target column.
55
- model_save_dir (str or Path or None): Directory used for saved models
56
- and processors. Models are saved under ``model_save_dir/models`` and
57
- preprocessing artifacts are saved under ``model_save_dir/processors``.
58
- When ``None``, models and processors are not saved. Default: None.
54
+ model_save_dir (str or Path or None): Directory used for cached processors
55
+ under ``model_save_dir/processors`` and optional saved models under
56
+ ``model_save_dir/models``. When ``None``, neither is saved.
57
+ Use a separate directory
58
+ for each dataset and split/preprocessing configuration. Default: None.
59
59
  random_state (int): First seed used for train splits and synthetic set
60
60
  sampling. Additional train seeds are consecutive integers starting
61
61
  from this value. Default: 42.
@@ -67,6 +67,8 @@ class TabularSynthesisBenchmark:
67
67
  ``"missforest"``. Default: ``"drop"``.
68
68
  monitor_memory (bool): Whether to record peak CPU and, when available,
69
69
  CUDA memory used by training. Default: False.
70
+ save_models (bool): Whether to save trained generators when
71
+ ``model_save_dir`` is set. Default: False.
70
72
  reuse_processors (bool): Whether to reuse saved preprocessing artifacts
71
73
  for each train seed when available. Default: True.
72
74
  max_eval_samples (int or None): Default maximum number of rows per real
@@ -75,6 +77,9 @@ class TabularSynthesisBenchmark:
75
77
  the same varying seed used for sampling synthetic data. Each
76
78
  synthetic set is generated at the (capped) real-train evaluation
77
79
  size. Default: 100_000.
80
+ allow_generated_missings (bool): Whether to evaluate generated samples
81
+ with missing numerical values. When False, skipped sets receive an
82
+ ``evaluation_status`` result explaining the skip. Default: False.
78
83
  dataset_save_dir (str or Path or None): Directory for saving sampled
79
84
  synthetic datasets as parquet files. When provided, datasets are
80
85
  saved under ``dataset_save_dir/train_seed/sampling_seed``.
@@ -93,7 +98,8 @@ class TabularSynthesisBenchmark:
93
98
  ... generator_params={"epochs": 300},
94
99
  ... categorical_features=["sex", "mortality"],
95
100
  ... target_column="mortality",
96
- ... model_save_dir="dataset/ctgan",
101
+ ... model_save_dir="dataset",
102
+ ... save_models=True,
97
103
  ... random_state=42,
98
104
  ... )
99
105
  >>>
@@ -109,7 +115,8 @@ class TabularSynthesisBenchmark:
109
115
  ... generator_params={"epochs": 300},
110
116
  ... categorical_features=["sex", "mortality"],
111
117
  ... target_column="mortality",
112
- ... model_save_dir="dataset/ctgan",
118
+ ... model_save_dir="dataset",
119
+ ... save_models=True,
113
120
  ... random_state=42,
114
121
  ... )
115
122
  >>> benchmark.train(n_train_seeds=3)
@@ -120,7 +127,8 @@ class TabularSynthesisBenchmark:
120
127
  ... generator_params={"epochs": 300},
121
128
  ... categorical_features=["sex", "mortality"],
122
129
  ... target_column="mortality",
123
- ... model_save_dir="dataset/ctgan",
130
+ ... model_save_dir="dataset",
131
+ ... save_models=True,
124
132
  ... random_state=42,
125
133
  ... )
126
134
  >>> results = reloaded_benchmark.eval(
@@ -141,8 +149,10 @@ class TabularSynthesisBenchmark:
141
149
  constraints: Optional[Union[list[str], str]] = None,
142
150
  missing_imputation_method: str = "drop",
143
151
  monitor_memory: bool = False,
152
+ save_models: bool = False,
144
153
  reuse_processors: bool = True,
145
154
  max_eval_samples: Optional[int] = 100_000,
155
+ allow_generated_missings: bool = False,
146
156
  dataset_save_dir: Optional[Union[str, Path]] = None,
147
157
  cap_train_time: Optional[float] = None,
148
158
  ):
@@ -157,8 +167,10 @@ class TabularSynthesisBenchmark:
157
167
  self.constraints = self._normalize_constraints(constraints)
158
168
  self.missing_imputation_method = missing_imputation_method
159
169
  self.monitor_memory = monitor_memory
170
+ self.save_models = save_models
160
171
  self.reuse_processors = reuse_processors
161
172
  self.max_eval_samples = self._validate_max_eval_samples(max_eval_samples)
173
+ self.allow_generated_missings = allow_generated_missings
162
174
  self.dataset_save_dir = (
163
175
  None if dataset_save_dir is None else Path(dataset_save_dir)
164
176
  )
@@ -178,7 +190,8 @@ class TabularSynthesisBenchmark:
178
190
  For each train seed, this method creates the train/test split, fits or
179
191
  loads the corresponding ``DataProcessor``, trains the
180
192
  configured generator, and saves the trained generator under
181
- ``model_save_dir/models`` when ``model_save_dir`` is set.
193
+ ``model_save_dir/models`` when ``save_models=True`` and ``model_save_dir``
194
+ is set.
182
195
 
183
196
  Use this method when you want to train models now and evaluate them
184
197
  later with ``eval()``. Use the same ``test_size`` during evaluation so
@@ -236,7 +249,7 @@ class TabularSynthesisBenchmark:
236
249
  if write_results:
237
250
  self._save_results(result_rows, results_path)
238
251
  break
239
- if self.model_save_dir is not None:
252
+ if self.save_models and self.model_save_dir is not None:
240
253
  generator.save(self._model_dir(train_seed))
241
254
 
242
255
  rows = [
@@ -247,11 +260,14 @@ class TabularSynthesisBenchmark:
247
260
  "set": pd.NA,
248
261
  }
249
262
  ]
263
+ rows.extend(trainable_param_rows(generator, train_seed))
264
+ rows.extend(trained_steps_epoch_rows(generator, train_seed))
250
265
  rows.extend(memory_metric_rows("training", memory_monitor, train_seed))
251
266
  result_rows.extend(rows)
252
267
  new_rows.extend(rows)
253
268
  if write_results:
254
269
  self._save_results(result_rows, results_path)
270
+ del generator, processor, processed, split
255
271
 
256
272
  return pd.DataFrame(new_rows, columns=RESULT_COLUMNS)
257
273
 
@@ -326,7 +342,7 @@ class TabularSynthesisBenchmark:
326
342
  processor, processed = self._load_or_fit_processor(train_seed, split)
327
343
  generator = self._load_generator(generator_cls, train_seed)
328
344
 
329
- rows = self._evaluate_generator(
345
+ for rows in self._evaluate_generator(
330
346
  generator=generator,
331
347
  processor=processor,
332
348
  processed=processed,
@@ -334,10 +350,11 @@ class TabularSynthesisBenchmark:
334
350
  n_sets=n_sets,
335
351
  train_seed=train_seed,
336
352
  max_eval_samples=max_eval_samples,
337
- )
338
- result_rows.extend(rows)
339
- new_rows.extend(rows)
340
- self._save_results(result_rows, results_path)
353
+ ):
354
+ result_rows.extend(rows)
355
+ new_rows.extend(rows)
356
+ self._save_results(result_rows, results_path)
357
+ del generator, processor, processed, split
341
358
 
342
359
  return pd.DataFrame(new_rows, columns=RESULT_COLUMNS)
343
360
 
@@ -411,6 +428,8 @@ class TabularSynthesisBenchmark:
411
428
  result_rows.extend(rows)
412
429
  new_rows.extend(rows)
413
430
  self._save_results(result_rows, results_path)
431
+ del X_syn, X_train_eval, X_test_eval
432
+ del processor, processed, split
414
433
 
415
434
  return pd.DataFrame(new_rows, columns=RESULT_COLUMNS)
416
435
 
@@ -486,7 +505,7 @@ class TabularSynthesisBenchmark:
486
505
  new_rows.extend(rows)
487
506
  self._save_results(result_rows, results_path)
488
507
  break
489
- if self.model_save_dir is not None:
508
+ if self.save_models and self.model_save_dir is not None:
490
509
  generator.save(self._model_dir(train_seed))
491
510
 
492
511
  rows = [
@@ -497,12 +516,14 @@ class TabularSynthesisBenchmark:
497
516
  "set": pd.NA,
498
517
  }
499
518
  ]
519
+ rows.extend(trainable_param_rows(generator, train_seed))
520
+ rows.extend(trained_steps_epoch_rows(generator, train_seed))
500
521
  rows.extend(memory_metric_rows("training", memory_monitor, train_seed))
501
522
  result_rows.extend(rows)
502
523
  new_rows.extend(rows)
503
524
  self._save_results(result_rows, results_path)
504
525
 
505
- rows = self._evaluate_generator(
526
+ for rows in self._evaluate_generator(
506
527
  generator=generator,
507
528
  processor=processor,
508
529
  processed=processed,
@@ -510,10 +531,11 @@ class TabularSynthesisBenchmark:
510
531
  n_sets=n_sets,
511
532
  train_seed=train_seed,
512
533
  max_eval_samples=max_eval_samples,
513
- )
514
- result_rows.extend(rows)
515
- new_rows.extend(rows)
516
- self._save_results(result_rows, results_path)
534
+ ):
535
+ result_rows.extend(rows)
536
+ new_rows.extend(rows)
537
+ self._save_results(result_rows, results_path)
538
+ del generator, processor, processed, split
517
539
 
518
540
  return pd.DataFrame(new_rows, columns=RESULT_COLUMNS)
519
541
 
@@ -561,14 +583,10 @@ class TabularSynthesisBenchmark:
561
583
  processor.save(processor_dir)
562
584
  print(f"Saved DataProcessor to {processor_file}")
563
585
 
564
- X_train_model = processed
565
- X_test_model = processor.preprocess(X=split["test"])
566
-
567
586
  return processor, {
568
- "train_model": X_train_model,
569
- "test_model": X_test_model,
570
- "train_eval": processor.postprocess(X_train_model),
571
- "test_eval": processor.postprocess(X_test_model),
587
+ "train_model": processed,
588
+ "train_eval": processor.imputer.transform(split["train"]),
589
+ "test_eval": processor.imputer.transform(split["test"]),
572
590
  }
573
591
 
574
592
  def _fit_generator(
@@ -642,8 +660,7 @@ class TabularSynthesisBenchmark:
642
660
  n_sets: int,
643
661
  train_seed: int,
644
662
  max_eval_samples: Optional[int],
645
- ) -> list[dict[str, Any]]:
646
- result_rows = []
663
+ ) -> Iterable[list[dict[str, Any]]]:
647
664
  for set_index in memory_guarded(range(n_sets)):
648
665
  sampling_seed = self.random_state + set_index
649
666
  X_train_eval = self._subsample_eval_dataset(
@@ -669,15 +686,15 @@ class TabularSynthesisBenchmark:
669
686
  X_syn,
670
687
  )
671
688
 
672
- result_rows.append(
689
+ rows = [
673
690
  {
674
691
  "metric name": "sampling_time_seconds",
675
692
  "metric value": sampling_time,
676
693
  "train_seed": train_seed,
677
694
  "set": set_index,
678
695
  }
679
- )
680
- result_rows.extend(
696
+ ]
697
+ rows.extend(
681
698
  self._evaluate_synthetic_datasets(
682
699
  X_train_eval=X_train_eval,
683
700
  X_test_eval=X_test_eval,
@@ -688,8 +705,8 @@ class TabularSynthesisBenchmark:
688
705
  sampling_seed=sampling_seed,
689
706
  )
690
707
  )
691
-
692
- return result_rows
708
+ del X_syn, X_train_eval, X_test_eval
709
+ yield rows
693
710
 
694
711
  def _evaluate_synthetic_datasets(
695
712
  self,
@@ -701,12 +718,26 @@ class TabularSynthesisBenchmark:
701
718
  set_index: int,
702
719
  sampling_seed: int,
703
720
  ) -> list[dict[str, Any]]:
704
- from synthyverse.evaluation.eval import TabularMetricEvaluator
705
-
706
721
  eval_categorical_features = [
707
722
  col for col in self.categorical_features if col in X_train_eval.columns
708
723
  ]
709
- evaluator = TabularMetricEvaluator(
724
+ numerical_features = [
725
+ col for col in X_syn.columns if col not in eval_categorical_features
726
+ ]
727
+ if (
728
+ not self.allow_generated_missings
729
+ and X_syn[numerical_features].isna().any().any()
730
+ ):
731
+ return [
732
+ {
733
+ "metric name": "evaluation_status",
734
+ "metric value": "skipped due to generated missing values",
735
+ "train_seed": train_seed,
736
+ "set": set_index,
737
+ }
738
+ ]
739
+
740
+ evaluator = evaluation.TabularMetricEvaluator(
710
741
  metrics=metrics,
711
742
  target_column=self.target_column,
712
743
  random_state=sampling_seed,
@@ -783,8 +814,9 @@ class TabularSynthesisBenchmark:
783
814
  return generator
784
815
 
785
816
  legacy_name = re.split(r"[\W_]+", generator, maxsplit=1)[0]
786
- if legacy_name != generator and TabularSynthesisBenchmark._is_registered_generator(
787
- legacy_name
817
+ if (
818
+ legacy_name != generator
819
+ and TabularSynthesisBenchmark._is_registered_generator(legacy_name)
788
820
  ):
789
821
  return legacy_name
790
822
  return generator
@@ -834,9 +866,7 @@ class TabularSynthesisBenchmark:
834
866
  return self.model_save_dir / "processors" / str(train_seed)
835
867
 
836
868
  def _model_dir(self, train_seed: int) -> Path:
837
- return (
838
- self.model_save_dir / "models" / self.generator / f"train_seed_{train_seed}"
839
- )
869
+ return self.model_save_dir / "models" / self.generator / str(train_seed)
840
870
 
841
871
  def _dataset_dir(self, train_seed: int, sampling_seed: int) -> Path:
842
872
  return self.dataset_save_dir / str(train_seed) / str(sampling_seed)
@@ -872,7 +902,7 @@ class TabularSynthesisBenchmark:
872
902
  return
873
903
  dataset_dir = self._dataset_dir(train_seed, sampling_seed)
874
904
  dataset_dir.mkdir(parents=True, exist_ok=True)
875
- X_syn.to_parquet(dataset_dir / "synthetic_train.parquet", index=False)
905
+ X_syn.to_parquet(dataset_dir / "syn.parquet", index=False)
876
906
 
877
907
  def _load_synthetic_dataset(
878
908
  self,
@@ -881,7 +911,7 @@ class TabularSynthesisBenchmark:
881
911
  processor: DataProcessor,
882
912
  ) -> pd.DataFrame:
883
913
  dataset_dir = self._dataset_dir(train_seed, sampling_seed)
884
- train_path = dataset_dir / "synthetic_train.parquet"
914
+ train_path = dataset_dir / "syn.parquet"
885
915
  if not train_path.exists():
886
916
  raise FileNotFoundError(
887
917
  f"No saved synthetic datasets found for train_seed={train_seed}, "
@@ -947,34 +977,26 @@ class TabularSynthesisBenchmark:
947
977
 
948
978
 
949
979
  def make_csv_safe(value: Any) -> Any:
980
+ value = _json_safe(value)
981
+ if isinstance(value, (dict, list)):
982
+ return json.dumps(value)
983
+ return pd.NA if value is None else value
984
+
985
+
986
+ def _json_safe(value: Any) -> Any:
950
987
  if isinstance(value, dict):
951
- return json.dumps(
952
- {str(key): make_csv_safe(item) for key, item in value.items()}
953
- )
988
+ return {str(key): _json_safe(item) for key, item in value.items()}
954
989
  if isinstance(value, (list, tuple, set)):
955
- return json.dumps([make_csv_safe(item) for item in value])
990
+ return [_json_safe(item) for item in value]
956
991
  if isinstance(value, Path):
957
992
  return str(value)
958
993
  if isinstance(value, pd.DataFrame):
959
- return value.to_json(orient="records")
960
- if isinstance(value, pd.Series):
961
- return make_csv_safe(value.to_list())
994
+ return json.loads(value.to_json(orient="records"))
962
995
  if hasattr(value, "tolist") and not isinstance(value, (str, bytes)):
963
- try:
964
- return make_csv_safe(value.tolist())
965
- except TypeError:
966
- pass
996
+ return _json_safe(value.tolist())
967
997
  if hasattr(value, "item") and not isinstance(value, (str, bytes)):
968
- try:
969
- return make_csv_safe(value.item())
970
- except (TypeError, ValueError):
971
- pass
972
- try:
973
- if pd.isna(value):
974
- return pd.NA
975
- except (TypeError, ValueError):
976
- pass
977
- return value
998
+ return _json_safe(value.item())
999
+ return None if pd.isna(value) else value
978
1000
 
979
1001
 
980
1002
  def flatten_metrics(metrics: Any, prefix: str = "") -> Iterable[tuple[str, Any]]:
@@ -1068,10 +1090,39 @@ def training_timeout_rows(
1068
1090
  ]
1069
1091
 
1070
1092
 
1093
+ def trainable_param_rows(generator, train_seed: int) -> list[dict[str, Any]]:
1094
+ get_trainable_params = getattr(generator, "get_trainable_params", None)
1095
+ if not callable(get_trainable_params):
1096
+ return []
1097
+ return [
1098
+ {
1099
+ "metric name": "trainable_params",
1100
+ "metric value": get_trainable_params(),
1101
+ "train_seed": train_seed,
1102
+ "set": pd.NA,
1103
+ }
1104
+ ]
1105
+
1106
+
1107
+ def trained_steps_epoch_rows(generator, train_seed: int) -> list[dict[str, Any]]:
1108
+ get_trained_steps_epochs = getattr(generator, "get_trained_steps_epochs", None)
1109
+ if not callable(get_trained_steps_epochs):
1110
+ return []
1111
+ return [
1112
+ {
1113
+ "metric name": name,
1114
+ "metric value": make_csv_safe(value),
1115
+ "train_seed": train_seed,
1116
+ "set": pd.NA,
1117
+ }
1118
+ for name, value in flatten_metrics(get_trained_steps_epochs())
1119
+ ]
1120
+
1121
+
1071
1122
  class PeakMemoryMonitor:
1072
1123
  """Sample memory used above the pre-training baseline."""
1073
1124
 
1074
- def __init__(self, interval_seconds: float = 1.0):
1125
+ def __init__(self, interval_seconds: float = 0.1):
1075
1126
  self.interval_seconds = interval_seconds
1076
1127
  self.peak_memory_mb: Optional[float] = None
1077
1128
  self.peak_cuda_memory_mb: Optional[float] = None
@@ -1162,7 +1213,34 @@ def _make_psutil_memory_reader():
1162
1213
  except Exception:
1163
1214
  return None
1164
1215
 
1216
+ def process_memory_bytes(current_process) -> Optional[int]:
1217
+ try:
1218
+ memory_info = current_process.memory_full_info()
1219
+ except Exception:
1220
+ try:
1221
+ return int(current_process.memory_info().rss)
1222
+ except Exception:
1223
+ return None
1224
+
1225
+ for field in ("pss", "uss", "rss"):
1226
+ value = getattr(memory_info, field, None)
1227
+ if value is not None:
1228
+ return int(value)
1229
+ return None
1230
+
1165
1231
  def read_memory_bytes() -> Optional[int]:
1232
+ try:
1233
+ processes = [process, *process.children(recursive=True)]
1234
+ except Exception:
1235
+ processes = [process]
1236
+
1237
+ total_bytes = 0
1238
+ for current_process in processes:
1239
+ memory_bytes = process_memory_bytes(current_process)
1240
+ if memory_bytes is not None:
1241
+ total_bytes += memory_bytes
1242
+ if total_bytes:
1243
+ return total_bytes
1166
1244
  try:
1167
1245
  return int(process.memory_info().rss)
1168
1246
  except Exception: