synthyverse 0.2.6__tar.gz → 0.2.7__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 (89) hide show
  1. {synthyverse-0.2.6 → synthyverse-0.2.7}/PKG-INFO +1 -1
  2. {synthyverse-0.2.6 → synthyverse-0.2.7}/pyproject.toml +1 -1
  3. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/benchmark/synthesis.py +24 -33
  4. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/cdtd_generator/cdtd.py +21 -32
  5. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/ctgan_generator/ct_gan.py +5 -19
  6. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/ctgan_generator/synthesizer.py +11 -8
  7. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/dgm_utils.py +14 -6
  8. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/forestdiffusion_generator/fd.py +3 -3
  9. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/synthpop_generator/synthpop.py +78 -20
  10. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabcascade_generator/tabcascade.py +27 -30
  11. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabddpm_generator/tabddpm.py +22 -32
  12. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabdiff_generator/tabdiff.py +21 -30
  13. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabsyn_generator/tabsyn.py +21 -27
  14. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tvae_generator/synthesizer.py +11 -8
  15. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tvae_generator/tvae.py +5 -19
  16. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/xgbddpm_generator/xgbddpm.py +87 -34
  17. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/xgbdiffusion_generator/xgbdiffusion.py +13 -9
  18. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse.egg-info/PKG-INFO +1 -1
  19. {synthyverse-0.2.6 → synthyverse-0.2.7}/LICENSE +0 -0
  20. {synthyverse-0.2.6 → synthyverse-0.2.7}/LICENSES/Apache-2.0.txt +0 -0
  21. {synthyverse-0.2.6 → synthyverse-0.2.7}/LICENSES/CTGAN-BSL-1.1.txt +0 -0
  22. {synthyverse-0.2.6 → synthyverse-0.2.7}/LICENSES/MIT.txt +0 -0
  23. {synthyverse-0.2.6 → synthyverse-0.2.7}/MANIFEST.in +0 -0
  24. {synthyverse-0.2.6 → synthyverse-0.2.7}/NOTICE +0 -0
  25. {synthyverse-0.2.6 → synthyverse-0.2.7}/README.md +0 -0
  26. {synthyverse-0.2.6 → synthyverse-0.2.7}/THIRD_PARTY_NOTICES.md +0 -0
  27. {synthyverse-0.2.6 → synthyverse-0.2.7}/logo/logo.png +0 -0
  28. {synthyverse-0.2.6 → synthyverse-0.2.7}/setup.cfg +0 -0
  29. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/__init__.py +0 -0
  30. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/benchmark/__init__.py +0 -0
  31. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/evaluation/__init__.py +0 -0
  32. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/evaluation/base.py +0 -0
  33. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/evaluation/distances.py +0 -0
  34. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/evaluation/eval.py +0 -0
  35. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/evaluation/fidelity.py +0 -0
  36. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/evaluation/ml.py +0 -0
  37. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/evaluation/preprocessing.py +0 -0
  38. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/evaluation/privacy.py +0 -0
  39. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/evaluation/utility.py +0 -0
  40. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/__init__.py +0 -0
  41. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/_optional.py +0 -0
  42. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/arf_generator/__init__.py +0 -0
  43. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/arf_generator/arf.py +0 -0
  44. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/arf_generator/model.py +0 -0
  45. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/arf_generator/utils.py +0 -0
  46. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/base.py +0 -0
  47. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/cdtd_generator/__init__.py +0 -0
  48. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/cdtd_generator/layers.py +0 -0
  49. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/cdtd_generator/utils.py +0 -0
  50. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/ctgan_generator/__init__.py +0 -0
  51. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/forestdiffusion_generator/__init__.py +0 -0
  52. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/forestdiffusion_generator/model.py +0 -0
  53. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/forestdiffusion_generator/utils/__init__.py +0 -0
  54. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/forestdiffusion_generator/utils/diffusion.py +0 -0
  55. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/forestdiffusion_generator/utils/utils_diffusion.py +0 -0
  56. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/forestdiffusion_generator/utils/writer.py +0 -0
  57. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/persistence.py +0 -0
  58. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/smote_generator/__init__.py +0 -0
  59. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/smote_generator/smote.py +0 -0
  60. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/synthpop_generator/__init__.py +0 -0
  61. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabcascade_generator/__init__.py +0 -0
  62. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabcascade_generator/disttree.py +0 -0
  63. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabcascade_generator/encoder.py +0 -0
  64. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabcascade_generator/highres.py +0 -0
  65. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabcascade_generator/lowres.py +0 -0
  66. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabddpm_generator/__init__.py +0 -0
  67. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabddpm_generator/ddpm_utils.py +0 -0
  68. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabddpm_generator/gaussian_multinomial_diffsuion.py +0 -0
  69. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabddpm_generator/modules.py +0 -0
  70. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabdiff_generator/__init__.py +0 -0
  71. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabdiff_generator/diffusion.py +0 -0
  72. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabdiff_generator/modules.py +0 -0
  73. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabdiff_generator/noise_schedule.py +0 -0
  74. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabdiff_generator/transformer.py +0 -0
  75. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabsyn_generator/__init__.py +0 -0
  76. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabsyn_generator/diffusion.py +0 -0
  77. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tabsyn_generator/vae.py +0 -0
  78. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/tvae_generator/__init__.py +0 -0
  79. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/univariate_generator/__init__.py +0 -0
  80. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/univariate_generator/univariate.py +0 -0
  81. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/xgbddpm_generator/__init__.py +0 -0
  82. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/generators/xgbdiffusion_generator/__init__.py +0 -0
  83. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/utils/__init__.py +0 -0
  84. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/utils/reproducibility.py +0 -0
  85. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse/utils/utils.py +0 -0
  86. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse.egg-info/SOURCES.txt +0 -0
  87. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse.egg-info/dependency_links.txt +0 -0
  88. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse.egg-info/requires.txt +0 -0
  89. {synthyverse-0.2.6 → synthyverse-0.2.7}/synthyverse.egg-info/top_level.txt +0 -0
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.1
2
2
  Name: synthyverse
3
- Version: 0.2.6
3
+ Version: 0.2.7
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
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
4
4
 
5
5
  [project]
6
6
  name = "synthyverse"
7
- version = "0.2.6"
7
+ version = "0.2.7"
8
8
  description = "Synthetic data generation and evaluation library"
9
9
  readme = "README.md"
10
10
  requires-python = ">=3.8"
@@ -66,7 +66,7 @@ class TabularSynthesisBenchmark:
66
66
  ``"mean"``, ``"median"``, ``"most_frequent"``, and
67
67
  ``"missforest"``. Default: ``"drop"``.
68
68
  monitor_memory (bool): Whether to record peak CPU and, when available,
69
- CUDA memory usage during training and sampling. Default: False.
69
+ CUDA memory used by training. Default: False.
70
70
  reuse_processors (bool): Whether to reuse saved preprocessing artifacts
71
71
  for each train seed when available. Default: True.
72
72
  max_eval_samples (int or None): Default maximum number of rows per real
@@ -654,21 +654,12 @@ class TabularSynthesisBenchmark:
654
654
  )
655
655
  n_train_syn = len(X_train_eval)
656
656
 
657
- memory_monitor = None
658
657
  sampling_start_time = perf_counter()
659
- if self.monitor_memory:
660
- with PeakMemoryMonitor() as memory_monitor:
661
- X_syn = generator.generate(
662
- n_train_syn,
663
- random_state=sampling_seed,
664
- )
665
- X_syn = processor.postprocess(X_syn)
666
- else:
667
- X_syn = generator.generate(
668
- n_train_syn,
669
- random_state=sampling_seed,
670
- )
671
- X_syn = processor.postprocess(X_syn)
658
+ X_syn = generator.generate(
659
+ n_train_syn,
660
+ random_state=sampling_seed,
661
+ )
662
+ X_syn = processor.postprocess(X_syn)
672
663
 
673
664
  X_syn = X_syn.reset_index(drop=True)
674
665
  sampling_time = perf_counter() - sampling_start_time
@@ -686,9 +677,6 @@ class TabularSynthesisBenchmark:
686
677
  "set": set_index,
687
678
  }
688
679
  )
689
- result_rows.extend(
690
- memory_metric_rows("sampling", memory_monitor, train_seed, set_index)
691
- )
692
680
  result_rows.extend(
693
681
  self._evaluate_synthetic_datasets(
694
682
  X_train_eval=X_train_eval,
@@ -1081,22 +1069,22 @@ def training_timeout_rows(
1081
1069
 
1082
1070
 
1083
1071
  class PeakMemoryMonitor:
1084
- """Sample process memory while a benchmark step is running."""
1072
+ """Sample memory used above the pre-training baseline."""
1085
1073
 
1086
1074
  def __init__(self, interval_seconds: float = 1.0):
1087
1075
  self.interval_seconds = interval_seconds
1088
1076
  self.peak_memory_mb: Optional[float] = None
1089
1077
  self.peak_cuda_memory_mb: Optional[float] = None
1090
1078
  self._baseline_rss_bytes: Optional[int] = None
1091
- self._baseline_cuda_bytes: Optional[int] = None
1079
+ self._baseline_cuda_memory_bytes: Optional[int] = None
1092
1080
  self._peak_rss_bytes: Optional[int] = None
1093
1081
  self._stop_event = threading.Event()
1094
1082
  self._thread: Optional[threading.Thread] = None
1095
1083
 
1096
1084
  def __enter__(self) -> "PeakMemoryMonitor":
1097
1085
  self._baseline_rss_bytes = _current_process_memory_bytes()
1098
- self._sample_once()
1099
- self._baseline_cuda_bytes = _cuda_memory_bytes()
1086
+ self._peak_rss_bytes = self._baseline_rss_bytes
1087
+ self._baseline_cuda_memory_bytes = _current_cuda_memory_bytes()
1100
1088
  _reset_cuda_peak_memory()
1101
1089
  self._thread = threading.Thread(target=self._sample_until_stopped, daemon=True)
1102
1090
  self._thread.start()
@@ -1109,10 +1097,13 @@ class PeakMemoryMonitor:
1109
1097
  self._thread.join(timeout=1.0)
1110
1098
  self._sample_once()
1111
1099
  self.peak_memory_mb = _bytes_to_mib(
1112
- _subtract_baseline(self._peak_rss_bytes, self._baseline_rss_bytes)
1100
+ _memory_delta_bytes(self._peak_rss_bytes, self._baseline_rss_bytes)
1113
1101
  )
1114
1102
  self.peak_cuda_memory_mb = _bytes_to_mib(
1115
- _subtract_baseline(_cuda_peak_memory_bytes(), self._baseline_cuda_bytes)
1103
+ _memory_delta_bytes(
1104
+ _cuda_peak_memory_bytes(),
1105
+ self._baseline_cuda_memory_bytes,
1106
+ )
1116
1107
  )
1117
1108
 
1118
1109
  def _sample_until_stopped(self) -> None:
@@ -1133,13 +1124,13 @@ def _bytes_to_mib(value: Optional[int]) -> Optional[float]:
1133
1124
  return value / BYTES_PER_MIB
1134
1125
 
1135
1126
 
1136
- def _subtract_baseline(
1137
- value: Optional[int],
1138
- baseline: Optional[int],
1127
+ def _memory_delta_bytes(
1128
+ peak_bytes: Optional[int],
1129
+ baseline_bytes: Optional[int],
1139
1130
  ) -> Optional[int]:
1140
- if value is None or baseline is None:
1131
+ if peak_bytes is None or baseline_bytes is None:
1141
1132
  return None
1142
- return max(0, value - baseline)
1133
+ return max(peak_bytes - baseline_bytes, 0)
1143
1134
 
1144
1135
 
1145
1136
  def _current_process_memory_bytes() -> Optional[int]:
@@ -1256,26 +1247,26 @@ def _reset_cuda_peak_memory() -> None:
1256
1247
  return
1257
1248
 
1258
1249
 
1259
- def _cuda_peak_memory_bytes() -> Optional[int]:
1250
+ def _current_cuda_memory_bytes() -> Optional[int]:
1260
1251
  try:
1261
1252
  import torch
1262
1253
 
1263
1254
  if not torch.cuda.is_available():
1264
1255
  return None
1265
1256
  torch.cuda.synchronize()
1266
- return int(torch.cuda.max_memory_allocated())
1257
+ return int(torch.cuda.memory_allocated())
1267
1258
  except Exception:
1268
1259
  return None
1269
1260
 
1270
1261
 
1271
- def _cuda_memory_bytes() -> Optional[int]:
1262
+ def _cuda_peak_memory_bytes() -> Optional[int]:
1272
1263
  try:
1273
1264
  import torch
1274
1265
 
1275
1266
  if not torch.cuda.is_available():
1276
1267
  return None
1277
1268
  torch.cuda.synchronize()
1278
- return int(torch.cuda.memory_allocated())
1269
+ return int(torch.cuda.max_memory_allocated())
1279
1270
  except Exception:
1280
1271
  return None
1281
1272
 
@@ -12,7 +12,6 @@ from ..dgm_utils import (
12
12
  FastTensorDataLoader,
13
13
  QuantileStandardScaler,
14
14
  clone_state_dict,
15
- split_validation,
16
15
  validate_c2st,
17
16
  )
18
17
  from .layers import MLP, MixedTypeDiffusion
@@ -54,9 +53,7 @@ class CDTDGenerator(BaseGenerator):
54
53
  ema_decay (float): Exponential moving average decay. Default: 0.999.
55
54
  log_steps (int): Steps between logging. Default: 100.
56
55
  cap_train_time (float): Time limit in seconds for training. Default: None.
57
- val_size (float): Fraction of training rows reserved for validation set early stopping. Default: 0.0.
58
- val_steps (int): Steps between validation. Default: 5000.
59
- target_column (str): Name of the target column, potentially used for stratified validation splitting. Default: None.
56
+ val_steps (int): Steps between training-set C2ST validation. Set to <=0 to disable validation. Default: 5000.
60
57
 
61
58
  Example:
62
59
  >>> import pandas as pd
@@ -106,9 +103,7 @@ class CDTDGenerator(BaseGenerator):
106
103
  random_state: int = 0,
107
104
  full_determinism: bool = False,
108
105
  cap_train_time: Optional[float] = None,
109
- val_size: float = 0.0,
110
106
  val_steps: int = 5000,
111
- target_column: Optional[str] = None,
112
107
  ):
113
108
  super().__init__(random_state=random_state, full_determinism=full_determinism)
114
109
  self.cat_emb_dim = cat_emb_dim
@@ -132,9 +127,7 @@ class CDTDGenerator(BaseGenerator):
132
127
  self.ema_decay = ema_decay
133
128
  self.log_steps = log_steps
134
129
  self.cap_train_time = cap_train_time
135
- self.val_size = val_size
136
130
  self.val_steps = val_steps
137
- self.target_column = target_column
138
131
 
139
132
  def _fit(self, X: pd.DataFrame, discrete_features: list):
140
133
  self.discrete_features = discrete_features
@@ -143,23 +136,15 @@ class CDTDGenerator(BaseGenerator):
143
136
  ]
144
137
  self.col_order = X.columns
145
138
 
146
- X = X.copy()
139
+ X_train = X.copy()
147
140
 
148
- X, X_val = split_validation(
149
- X,
150
- self.val_size,
151
- self.target_column,
152
- self.discrete_features,
153
- self.random_state,
141
+ self.quant_encoder = QuantileStandardScaler(X_train.shape[0], self.random_state)
142
+ X_train[self.numerical_features] = self.quant_encoder.fit_transform(
143
+ X_train[self.numerical_features].astype(float)
154
144
  )
155
145
 
156
- self.quant_encoder = QuantileStandardScaler(X.shape[0], self.random_state)
157
- X[self.numerical_features] = self.quant_encoder.fit_transform(
158
- X[self.numerical_features].astype(float)
159
- )
160
-
161
- X_discrete = torch.tensor(X[self.discrete_features].to_numpy()).long()
162
- X_numerical = torch.tensor(X[self.numerical_features].to_numpy()).float()
146
+ X_discrete = torch.tensor(X_train[self.discrete_features].to_numpy()).long()
147
+ X_numerical = torch.tensor(X_train[self.numerical_features].to_numpy()).float()
163
148
 
164
149
  # --- build diffusion model ---
165
150
  self.num_cat_features = X_discrete.shape[1]
@@ -245,9 +230,10 @@ class CDTDGenerator(BaseGenerator):
245
230
  best_val_score = float("inf")
246
231
  best_val_model = None
247
232
 
248
- start_time = time.monotonic()
233
+ train_time = 0.0
249
234
  with tqdm(initial=current_step, total=self.training_steps) as pbar:
250
235
  while current_step < self.training_steps:
236
+ step_start_time = time.monotonic()
251
237
  optimizer.zero_grad()
252
238
 
253
239
  inputs = next(train_iter)
@@ -277,8 +263,16 @@ class CDTDGenerator(BaseGenerator):
277
263
  for param_group in optimizer.param_groups:
278
264
  param_group["lr"] = scheduler(current_step)
279
265
 
266
+ train_time += time.monotonic() - step_start_time
267
+ if (
268
+ self.cap_train_time is not None
269
+ and train_time > self.cap_train_time
270
+ ):
271
+ print(f"Training timed out after {self.cap_train_time} seconds.")
272
+ break
273
+
280
274
  if (
281
- X_val is not None
275
+ self.val_steps > 0
282
276
  and current_step % self.val_steps == 0
283
277
  and current_step > 0
284
278
  ):
@@ -286,7 +280,9 @@ class CDTDGenerator(BaseGenerator):
286
280
  self.diff_model.eval()
287
281
  ema_diff_model.store()
288
282
  ema_diff_model.copy_to()
289
- score = validate_c2st(self, X_val, random_state=self.random_state)
283
+
284
+ score = validate_c2st(self, X, random_state=self.random_state)
285
+ print(f"Validation score at step {current_step}: {score}")
290
286
 
291
287
  if score < best_val_score:
292
288
  best_val_score = score
@@ -301,13 +297,6 @@ class CDTDGenerator(BaseGenerator):
301
297
  if stop_training:
302
298
  break
303
299
 
304
- if (
305
- self.cap_train_time is not None
306
- and (time.monotonic() - start_time) > self.cap_train_time
307
- ):
308
- print(f"Training timed out after {self.cap_train_time} seconds.")
309
- break
310
-
311
300
  if best_val_model is None:
312
301
  ema_diff_model.copy_to()
313
302
  else:
@@ -7,7 +7,7 @@ import pandas as pd
7
7
 
8
8
  from .._optional import require_ctgan
9
9
  from ..base import BaseGenerator
10
- from ..dgm_utils import clone_state_dict, split_validation, validate_c2st
10
+ from ..dgm_utils import clone_state_dict, validate_c2st
11
11
  from ...utils.utils import resolve_epochs_from_training_steps
12
12
 
13
13
 
@@ -41,9 +41,7 @@ class CTGANGenerator(BaseGenerator):
41
41
  discriminator. Default: 10.
42
42
  cuda (bool): Whether to use CUDA if available. Default: True.
43
43
  cap_train_time (float): Time limit in seconds for training. Default: None.
44
- val_size (float): Fraction of training rows reserved for validation set early stopping. Default: 0.0.
45
- val_steps (int): Epochs between validation, or training steps when ``training_steps`` is provided. Default: 50.
46
- target_column (str): Name of the target column, potentially used for stratified validation splitting. Default: None.
44
+ val_steps (int): Epochs between training-set C2ST validation, or training steps when ``training_steps`` is provided. Set to <=0 to disable validation. Default: 50.
47
45
 
48
46
  Example:
49
47
  >>> import pandas as pd
@@ -85,9 +83,7 @@ class CTGANGenerator(BaseGenerator):
85
83
  pac=10,
86
84
  cuda=True,
87
85
  cap_train_time: Optional[float] = None,
88
- val_size: float = 0.0,
89
86
  val_steps: int = 50,
90
- target_column: Optional[str] = None,
91
87
  random_state: int = 0,
92
88
  full_determinism: bool = False,
93
89
  ):
@@ -114,21 +110,13 @@ class CTGANGenerator(BaseGenerator):
114
110
  raise ValueError("pac must be >= 1.")
115
111
  self.cuda = cuda
116
112
  self.cap_train_time = cap_train_time
117
- self.val_size = val_size
118
113
  self.val_steps = val_steps
119
- self.target_column = target_column
120
114
 
121
115
  def _fit(self, X: pd.DataFrame, discrete_features: list):
122
116
  from .synthesizer import CTGAN
123
117
 
124
118
  self.discrete_features = list(discrete_features)
125
- X, X_val = split_validation(
126
- X,
127
- self.val_size,
128
- self.target_column,
129
- self.discrete_features,
130
- self.random_state,
131
- )
119
+ X = X.copy()
132
120
 
133
121
  self.batch_size = self._round_batch_size(self.batch_size, self.pac)
134
122
  epochs = resolve_epochs_from_training_steps(
@@ -164,7 +152,7 @@ class CTGANGenerator(BaseGenerator):
164
152
  def validate():
165
153
  nonlocal best_val_score, best_val_model
166
154
  self.model._generator.eval()
167
- score = validate_c2st(self, X_val, random_state=self.random_state)
155
+ score = validate_c2st(self, X, random_state=self.random_state)
168
156
  self.model._generator.train()
169
157
  if score < best_val_score:
170
158
  best_val_score = score
@@ -173,7 +161,7 @@ class CTGANGenerator(BaseGenerator):
173
161
  return True
174
162
 
175
163
  def validate_callback(step, epoch, epoch_end):
176
- if X_val is None:
164
+ if self.val_steps <= 0:
177
165
  return False
178
166
  if self.training_steps is not None:
179
167
  return (
@@ -220,9 +208,7 @@ class CTGANGenerator(BaseGenerator):
220
208
  "pac": self.pac,
221
209
  "cuda": self.cuda,
222
210
  "cap_train_time": self.cap_train_time,
223
- "val_size": self.val_size,
224
211
  "val_steps": self.val_steps,
225
- "target_column": self.target_column,
226
212
  "discrete_features": getattr(self, "discrete_features", None),
227
213
  }
228
214
 
@@ -424,12 +424,13 @@ class CTGAN(BaseSynthesizer):
424
424
  epoch_iterator.set_description(description.format(gen=0, dis=0))
425
425
 
426
426
  steps_per_epoch = max(len(train_data) // self._batch_size, 1)
427
- start_time = time.monotonic()
427
+ train_time = 0.0
428
428
  step = 0
429
429
  timed_out = False
430
430
  stop_training = False
431
431
  for i in epoch_iterator:
432
432
  for id_ in range(steps_per_epoch):
433
+ step_start_time = time.monotonic()
433
434
  for n in range(self._discriminator_steps):
434
435
  fakez = torch.normal(mean=mean, std=std)
435
436
 
@@ -507,18 +508,19 @@ class CTGAN(BaseSynthesizer):
507
508
  loss_g.backward()
508
509
  optimizerG.step()
509
510
  step += 1
510
- if validate_callback is not None and validate_callback(
511
- step, i + 1, False
512
- ):
513
- stop_training = True
514
- break
511
+ train_time += time.monotonic() - step_start_time
515
512
  if (
516
513
  self.cap_train_time is not None
517
- and time.monotonic() - start_time > self.cap_train_time
514
+ and train_time > self.cap_train_time
518
515
  ):
519
516
  print(f"Training timed out after {self.cap_train_time} seconds.")
520
517
  timed_out = True
521
518
  break
519
+ if validate_callback is not None and validate_callback(
520
+ step, i + 1, False
521
+ ):
522
+ stop_training = True
523
+ break
522
524
 
523
525
  generator_loss = loss_g.detach().cpu().item()
524
526
  discriminator_loss = loss_d.detach().cpu().item()
@@ -542,7 +544,8 @@ class CTGAN(BaseSynthesizer):
542
544
  description.format(gen=generator_loss, dis=discriminator_loss)
543
545
  )
544
546
  if (
545
- not stop_training
547
+ not timed_out
548
+ and not stop_training
546
549
  and validate_callback is not None
547
550
  and validate_callback(step, i + 1, True)
548
551
  ):
@@ -1,6 +1,6 @@
1
1
  import random
2
2
  from contextlib import contextmanager
3
-
3
+ from math import ceil
4
4
  import numpy as np
5
5
  import torch
6
6
  import torch.nn as nn
@@ -30,16 +30,22 @@ def preserve_rng_state(generator):
30
30
 
31
31
 
32
32
  def validate_c2st(
33
- self, X_val: pd.DataFrame, n_sets: int = 3, nfolds: int = 3, random_state: int = 0
33
+ self,
34
+ X: pd.DataFrame,
35
+ nfolds: int = 3,
36
+ random_state: int = 0,
37
+ max_samples: int = 50_000,
34
38
  ):
35
- # preserves rng state for training the DGM
39
+ n = min(len(X), max_samples)
40
+ n_sets = int(max(min(ceil(max_samples / len(X)), 10), 3))
36
41
  with preserve_rng_state(self):
37
42
  result = 0
38
43
  for i in range(n_sets):
39
44
  state = random_state + i + 1
40
- syn = self.generate(len(X_val), random_state=state)
45
+ x = X.sample(n, replace=False, ignore_index=True, random_state=state)
46
+ syn = self.generate(n, random_state=state)
41
47
  score = ClassifierTwoSampleTest(nfold=nfolds, random_state=state).evaluate(
42
- X_val, syn, discrete_features=self.discrete_features
48
+ x, syn, discrete_features=self.discrete_features
43
49
  )["c2st.auc"]
44
50
  result += abs(score - 0.5)
45
51
  return result / n_sets
@@ -162,7 +168,9 @@ class FastTensorDataLoader:
162
168
 
163
169
  n_batches, remainder = divmod(self.dataset_len, self.batch_size)
164
170
  self.n_batches = n_batches if drop_last else n_batches + (remainder > 0)
165
- self.iter_len = self.n_batches * self.batch_size if drop_last else self.dataset_len
171
+ self.iter_len = (
172
+ self.n_batches * self.batch_size if drop_last else self.dataset_len
173
+ )
166
174
 
167
175
  def __iter__(self):
168
176
  self.indices = torch.randperm(self.dataset_len) if self.shuffle else None
@@ -29,15 +29,15 @@ class ForestDiffusionGenerator(BaseGenerator):
29
29
  trained XGBoost models are kept in memory. Default: None.
30
30
  num_timesteps: The number of noise levels to use. Default: 50.
31
31
  diffusion_type: The type of diffusion to use. Default: "vp".
32
- multi_output: Whether to use multi-output XGBoost ensembles. Default: False.
32
+ multi_output: Whether to use multi-output XGBoost. Default: False.
33
33
  xgb_hypers: The hyperparameters for the XGBoost models. Default: {"n_estimators": 100, "max_depth": 6, "early_stopping_rounds": 50}.
34
34
  noise_samples_per_row: The number of different noise samples per real data sample. Default: 100.
35
35
  eps: The timestep to stop generation at. Default: 0.001.
36
36
  beta_min: The minimum beta for the diffusion. Default: 0.1.
37
37
  beta_max: The maximum beta for the diffusion. Default: 8.
38
38
  n_jobs: The number of parallel jobs to use. Default: -1.
39
- backend: The backend to use. Default: "loky".
40
- n_batch: The number of batches to use. Default: -1.
39
+ backend: The joblib backend to use. Default: "loky".
40
+ n_batch: The number of batches to use in XGBoost data iterator. Set to <=0 to disable data iterator. Default: -1.
41
41
 
42
42
  Example:
43
43
  >>> import pandas as pd