arborenum 0.2.4__tar.gz → 0.2.6__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.
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: arborenum
3
- Version: 0.2.4
3
+ Version: 0.2.6
4
4
  Summary: Exact, approximate, and anytime decision tree Rashomon sets over binary and continuous features
5
5
  Requires-Python: >=3.9
6
6
  Description-Content-Type: text/markdown
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
4
4
 
5
5
  [project]
6
6
  name = "arborenum"
7
- version = "0.2.4"
7
+ version = "0.2.6"
8
8
  description = "Exact, approximate, and anytime decision tree Rashomon sets over binary and continuous features"
9
9
  requires-python = ">=3.9"
10
10
  dependencies = [
@@ -776,6 +776,7 @@ class ArborEnum:
776
776
  self.multiplicative_slack_ = float(multiplicative_slack)
777
777
  self.lookahead_k_ = int(lookahead_k)
778
778
  self.eta_defer_ = float(eta_defer)
779
+ self._model.set_additive(bool(additive))
779
780
 
780
781
  if X_initial is None:
781
782
  if early_stopping and X_num.shape[1] > 0:
@@ -856,8 +857,7 @@ class ArborEnum:
856
857
  memory_limit_mb=memory_limit_mb,
857
858
  use_deferral=use_deferral,
858
859
  eta_defer=eta_defer,
859
- additive=additive,
860
-
860
+ additive=additive,
861
861
  )
862
862
 
863
863
  return self.fit_prepared(
@@ -890,7 +890,6 @@ class ArborEnum:
890
890
  stronger_rollout=stronger_rollout,
891
891
  use_deferral=use_deferral,
892
892
  eta_defer=eta_defer,
893
- additive=additive,
894
893
  )
895
894
 
896
895
 
@@ -1363,6 +1362,8 @@ class ArborEnum:
1363
1362
 
1364
1363
  key_mode_parsed = parse_key_mode(key_mode)
1365
1364
 
1365
+ self._model.set_additive(bool(additive))
1366
+
1366
1367
  self._model.fit_then_extend(
1367
1368
  X,
1368
1369
  y,
@@ -1394,7 +1395,6 @@ class ArborEnum:
1394
1395
  bool(use_deferral),
1395
1396
  float(eta_defer),
1396
1397
  bb_pred_vec,
1397
- bool(additive)
1398
1398
  )
1399
1399
 
1400
1400
  return self
@@ -1434,6 +1434,7 @@ class ArborEnum:
1434
1434
  use_deferral=False,
1435
1435
  eta_defer=0.0,
1436
1436
  bb_pred=None,
1437
+ additive=False,
1437
1438
  ):
1438
1439
  X = np.asarray(X, dtype=np.uint8)
1439
1440
  y = np.asarray(y, dtype=int)
@@ -1554,6 +1555,8 @@ class ArborEnum:
1554
1555
  )
1555
1556
  key_mode_parsed = parse_key_mode(key_mode)
1556
1557
 
1558
+ self._model.set_additive(bool(additive))
1559
+
1557
1560
  self._model.fit_anytime(
1558
1561
  X,
1559
1562
  y,
@@ -1946,6 +1949,8 @@ class ArborEnum:
1946
1949
 
1947
1950
  key_mode_parsed = parse_key_mode(key_mode)
1948
1951
 
1952
+ self._model.set_additive(bool(additive))
1953
+
1949
1954
  self._model.fit_prepared_then_extend(
1950
1955
  float(lambda_reg),
1951
1956
  int(depth_budget),
@@ -1972,7 +1977,6 @@ class ArborEnum:
1972
1977
  float(memory_limit_mb),
1973
1978
  bool(use_deferral),
1974
1979
  float(eta_defer),
1975
- bool(additive)
1976
1980
  )
1977
1981
 
1978
1982
  return self
@@ -2008,6 +2012,7 @@ class ArborEnum:
2008
2012
  memory_limit_mb=-1.0,
2009
2013
  use_deferral=False,
2010
2014
  eta_defer=0.0,
2015
+ additive=False,
2011
2016
  ):
2012
2017
  rashomon_mult = float(rashomon_mult)
2013
2018
 
@@ -2065,6 +2070,8 @@ class ArborEnum:
2065
2070
 
2066
2071
  key_mode_parsed = parse_key_mode(key_mode)
2067
2072
 
2073
+ self._model.set_additive(bool(additive))
2074
+
2068
2075
  self._model.fit_prepared_anytime(
2069
2076
  float(lambda_reg),
2070
2077
  int(depth_budget),
@@ -2276,6 +2283,109 @@ class ArborEnum:
2276
2283
  copy=False,
2277
2284
  )
2278
2285
 
2286
+ def get_exact_replacement_importance_intervals(
2287
+ self,
2288
+ X_eval,
2289
+ y_eval,
2290
+ budget_override=None,
2291
+ ):
2292
+ """
2293
+ Exact global [min, max] subtractive model-reliance interval
2294
+ over the fitted Rashomon set, evaluated on X_eval, y_eval.
2295
+ X_eval may be completely different from the training data.
2296
+ """
2297
+
2298
+ if self.n_features_in_ is None:
2299
+ raise RuntimeError("Fit the model first.")
2300
+
2301
+ X_internal = np.ascontiguousarray(
2302
+ self.transform(X_eval),
2303
+ dtype=np.uint8,
2304
+ )
2305
+
2306
+ y_eval = np.ascontiguousarray(
2307
+ np.asarray(y_eval, dtype=np.int32)
2308
+ )
2309
+
2310
+ if y_eval.ndim != 1:
2311
+ raise ValueError("y_eval must be 1D.")
2312
+
2313
+ if y_eval.shape[0] != X_internal.shape[0]:
2314
+ raise ValueError(
2315
+ "y_eval must have the same number of rows as X_eval."
2316
+ )
2317
+
2318
+ # map each original feature to its internal threshold columns.
2319
+ variable_columns = [
2320
+ [] for _ in range(self.n_features_in_)
2321
+ ]
2322
+
2323
+ # low-cardinality features created directly in python.
2324
+ for internal_feature, spec in enumerate(
2325
+ self.binary_feature_specs_
2326
+ ):
2327
+ original_feature = int(spec["original_feature"])
2328
+ variable_columns[original_feature].append(
2329
+ int(internal_feature)
2330
+ )
2331
+
2332
+ # continuous threshold groups created in cpp
2333
+ starts = self.get_continuous_starts()
2334
+
2335
+ for group, original_feature in enumerate(
2336
+ self.continuous_feature_indices_
2337
+ ):
2338
+ original_feature = int(original_feature)
2339
+
2340
+ start = int(starts[group])
2341
+ end = int(self.get_continuous_group_end(group))
2342
+
2343
+ variable_columns[original_feature].extend(
2344
+ range(start, end)
2345
+ )
2346
+
2347
+ # constant training features have no internal columns and
2348
+ # therefore importance exactly zero for every fitted tree.
2349
+ active_features = [
2350
+ j
2351
+ for j, cols in enumerate(variable_columns)
2352
+ if cols
2353
+ ]
2354
+
2355
+ active_variable_columns = [
2356
+ variable_columns[j]
2357
+ for j in active_features
2358
+ ]
2359
+
2360
+ budget = (
2361
+ -1
2362
+ if budget_override is None
2363
+ else int(budget_override)
2364
+ )
2365
+
2366
+ active_intervals = np.asarray(
2367
+ self._model
2368
+ .get_exact_replacement_importance_intervals_packed_trie(
2369
+ X_internal,
2370
+ y_eval,
2371
+ budget,
2372
+ active_variable_columns,
2373
+ np.empty(0, dtype=np.int32),
2374
+ False, # global min/max, not per-sample
2375
+ ),
2376
+ dtype=float,
2377
+ )
2378
+
2379
+ intervals = np.zeros(
2380
+ (self.n_features_in_, 2),
2381
+ dtype=float,
2382
+ )
2383
+
2384
+ if active_features:
2385
+ intervals[active_features] = active_intervals
2386
+
2387
+ return intervals
2388
+
2279
2389
 
2280
2390
  # WARNING: 1-indexed unlike features
2281
2391
  def get_tree_paths(self, tree_index: int):
@@ -2060,6 +2060,58 @@ PYBIND11_MODULE(_core, m) {
2060
2060
  "continuous feature group."
2061
2061
  )
2062
2062
 
2063
+ .def(
2064
+ "get_exact_replacement_importance_intervals_packed_trie",
2065
+ [](const ArborEnum& self,
2066
+ py::array_t<
2067
+ uint8_t,
2068
+ py::array::c_style | py::array::forcecast
2069
+ > X_eval,
2070
+ py::array_t<
2071
+ int,
2072
+ py::array::c_style | py::array::forcecast
2073
+ > y_eval,
2074
+ int budget_override,
2075
+ const std::vector<std::vector<int>>& variable_columns,
2076
+ py::array_t<
2077
+ int,
2078
+ py::array::c_style | py::array::forcecast
2079
+ > bb_pred_eval,
2080
+ bool sum_samplewise_extrema) {
2081
+
2082
+ auto X_vec =
2083
+ numpy_uint8_2d_to_row_major(X_eval, "X_eval");
2084
+
2085
+ auto y_vec =
2086
+ numpy_int_1d_to_vector(y_eval, "y_eval");
2087
+
2088
+ auto bb_vec =
2089
+ numpy_int_1d_to_vector(
2090
+ bb_pred_eval,
2091
+ "bb_pred_eval"
2092
+ );
2093
+
2094
+ return self
2095
+ .get_exact_replacement_importance_intervals_packed_trie(
2096
+ X_vec,
2097
+ y_vec,
2098
+ budget_override,
2099
+ variable_columns,
2100
+ bb_vec,
2101
+ {}, // matched_group_of_row_by_variable_eval
2102
+ {}, // matched_group_size_by_variable_eval
2103
+ sum_samplewise_extrema
2104
+ );
2105
+ },
2106
+ py::arg("X_eval"),
2107
+ py::arg("y_eval"),
2108
+ py::arg("budget_override") = -1,
2109
+ py::arg("variable_columns") =
2110
+ std::vector<std::vector<int>>{},
2111
+ py::arg("bb_pred_eval") = py::array_t<int>(0),
2112
+ py::arg("sum_samplewise_extrema") = false
2113
+ )
2114
+
2063
2115
 
2064
2116
  .def(
2065
2117
  "training_samples_with_multiple_reachable_predictions",