autogluon.core 1.2.1b20250113__py3-none-any.whl → 1.2.1b20250114__py3-none-any.whl

autogluon/core/models/ensemble/bagged_ensemble_model.py

@@ -18,7 +18,7 @@ from autogluon.common.utils.distribute_utils import DistributedContext
 from autogluon.common.utils.log_utils import DuplicateFilter
 from autogluon.common.utils.try_import import try_import_ray
 
-from ...constants import MULTICLASS, QUANTILE, REFIT_FULL_SUFFIX, REGRESSION, SOFTCLASS
+from ...constants import BINARY, MULTICLASS, QUANTILE, REFIT_FULL_SUFFIX, REGRESSION, SOFTCLASS
 from ...hpo.exceptions import EmptySearchSpace
 from ...pseudolabeling.pseudolabeling import assert_pseudo_column_match
 from ...utils.exceptions import TimeLimitExceeded
@@ -106,6 +106,9 @@ class BaggedEnsembleModel(AbstractModel):
             # 'refit_folds': False,  # [Advanced, Experimental] Whether to refit bags immediately to a refit_full model in a single .fit call.
             # 'num_folds' None,  # Number of bagged folds per set. If specified, overrides .fit `k_fold` value.
             # 'max_sets': None,  # Maximum bagged repeats to allow, if specified, will set `self.can_fit()` to `self._n_repeats_finished < max_repeats`
+            "stratify": "auto",
+            "bin": "auto",
+            "n_bins": None,
         }
         for param, val in default_params.items():
             self._set_default_param_value(param, val)
@@ -125,11 +128,32 @@ class BaggedEnsembleModel(AbstractModel):
     def can_infer(self):
         return self.is_fit() and self.params.get("save_bag_folds", True)
 
-    def is_stratified(self):
-        if self.problem_type in [REGRESSION, QUANTILE, SOFTCLASS]:
-            return False
+    def is_stratified(self) -> bool:
+        """
+        Returns whether to stratify on the label during KFold splits
+        """
+        stratify = self.params.get("stratify", "auto")
+        if isinstance(stratify, str) and stratify == "auto":
+            return self.problem_type in [
+                BINARY,
+                MULTICLASS,
+
+                # Commented out due to inconclusive results on whether this is helpful when combined with binning
+                # REGRESSION,
+                # QUANTILE,
+            ]
         else:
-            return True
+            return stratify
+
+    def is_binned(self) -> bool:
+        """
+        Returns whether to bin the label during stratified KFold splits
+        """
+        bin = self.params.get("bin", "auto")
+        if isinstance(bin, str) and bin == "auto":
+            return self.problem_type in [REGRESSION, QUANTILE]
+        else:
+            return bin
 
     def is_fit(self) -> bool:
         return self.n_children != 0
@@ -188,8 +212,16 @@ class BaggedEnsembleModel(AbstractModel):
         else:
             return X
 
-    def _get_cv_splitter(self, n_splits, n_repeats, groups=None):
-        return CVSplitter(n_splits=n_splits, n_repeats=n_repeats, groups=groups, stratified=self.is_stratified(), random_state=self._random_state)
+    def _get_cv_splitter(self, n_splits: int, n_repeats: int, groups=None) -> CVSplitter:
+        return CVSplitter(
+            n_splits=n_splits,
+            n_repeats=n_repeats,
+            groups=groups,
+            stratify=self.is_stratified(),
+            bin=self.is_binned(),
+            n_bins=self.params.get("n_bins", None),
+            random_state=self._random_state,
+        )
 
     def _fit(
         self,

autogluon/core/utils/utils.py

@@ -13,7 +13,8 @@ import pandas as pd
 import scipy.stats
 from numpy.typing import ArrayLike
 from pandas import DataFrame, Series
-from sklearn.model_selection import LeaveOneGroupOut, RepeatedKFold, RepeatedStratifiedKFold, train_test_split
+from sklearn.model_selection import BaseCrossValidator, LeaveOneGroupOut, RepeatedKFold, RepeatedStratifiedKFold, train_test_split
+from sklearn.preprocessing import KBinsDiscretizer
 
 from autogluon.common.utils.resource_utils import ResourceManager
 
@@ -34,12 +35,54 @@ from .miscs import warning_filter
 logger = logging.getLogger(__name__)
 
 
+# TODO: Add binned stratification support for regression in train/val split (non CV)
 class CVSplitter:
-    def __init__(self, splitter_cls=None, n_splits=5, n_repeats=1, random_state=0, stratified=False, groups=None):
+    def __init__(
+        self,
+        splitter_cls=None,
+        n_splits: int = 5,
+        n_repeats: int = 1,
+        random_state: int | None = 0,
+        stratify: bool = False,
+        bin: bool = False,
+        n_bins: int | None = None,
+        groups: pd.Series = None,
+    ):
+        """
+        Wrapper around splitter objects to perform KFold splits.
+        Supports regression stratification via the `bin` and `n_bins` argument.
+
+        Parameters
+        ----------
+        splitter_cls, default None
+            The class to use for splitting.
+            If None, will automatically be determined based off of `stratify`, `groups`, and `n_repeats`.
+        n_splits : int, default 5
+            The number of splits to perform.
+            Ignored if `groups` is specified.
+        n_repeats: int, default 1
+            The number of repeated splits to perform.
+            Ignored if `groups` is specified.
+        random_state : int, default 0
+            The seed to use when splitting the data.
+        stratify : bool, default False
+            If True, will stratify the splits on `y`.
+        bin : bool, default False
+            If True and `stratify` is True, will bin `y` into `n_bins` bins for stratification.
+            Should only be used for regression and quantile tasks.
+        n_bins : int, default None
+            The number of bins to use when `bin` is True.
+            If None, defaults to `np.floor(n_samples / n_splits)`.
+        groups : pd.Series, default None
+            If specified, splitter_cls will default to LeaveOneGroupOut.
+
+        """
         self.n_splits = n_splits
         self.n_repeats = n_repeats
         self.random_state = random_state
-        self.stratified = stratified
+        self.stratify = stratify
+        self.bin = bin
+        self.n_bins = n_bins
         self.groups = groups
         if splitter_cls is None:
             splitter_cls = self._get_splitter_cls()
@@ -53,13 +96,13 @@ class CVSplitter:
             self.n_splits = num_groups
             splitter_cls = LeaveOneGroupOut
             # pass
-        elif self.stratified:
+        elif self.stratify:
             splitter_cls = RepeatedStratifiedKFold
         else:
             splitter_cls = RepeatedKFold
         return splitter_cls
 
-    def _get_splitter(self, splitter_cls):
+    def _get_splitter(self, splitter_cls) -> BaseCrossValidator:
         if splitter_cls == LeaveOneGroupOut:
             return splitter_cls()
         elif splitter_cls in [RepeatedKFold, RepeatedStratifiedKFold]:
@@ -67,19 +110,38 @@ class CVSplitter:
         else:
             raise AssertionError(f"{splitter_cls} is not supported as a valid `splitter_cls` input to CVSplitter.")
 
-    def split(self, X, y):
-        if isinstance(self._splitter, RepeatedStratifiedKFold):
+    def split(self, X: pd.DataFrame, y: pd.Series) -> list[tuple[np.ndarray, np.ndarray]]:
+        splitter = self._splitter
+        if isinstance(splitter, RepeatedStratifiedKFold):
+            if self.bin:
+                if self.n_bins is None:
+                    n_splits = splitter.get_n_splits()
+                    n_samples = len(y)
+
+                    # ensure at least n_splits samples per bin
+                    n_bins = int(np.floor(n_samples / n_splits))
+                else:
+                    n_bins = self.n_bins
+
+                if n_bins > 1:
+                    k_bins_discretizer = KBinsDiscretizer(n_bins=n_bins, encode='ordinal', random_state=self.random_state)
+                    y_bin = k_bins_discretizer.fit_transform(y.to_frame())[:, 0]
+                    y = pd.Series(data=y_bin, index=y.index, name=y.name)
+                else:
+                    # Don't stratify, can't bin!
+                    splitter = self._get_splitter(splitter_cls=RepeatedKFold)
+
             # FIXME: There is a bug in sklearn that causes an incorrect ValueError if performing stratification and all classes have fewer than n_splits samples.
             #  This is hacked by adding a dummy class with n_splits samples, performing the kfold split, then removing the dummy samples from all resulting indices.
             #  This is very inefficient and complicated and ideally should be fixed in sklearn.
             with warning_filter():
                 try:
-                    out = [[train_index, test_index] for train_index, test_index in self._splitter.split(X, y)]
+                    out = [[train_index, test_index] for train_index, test_index in splitter.split(X, y)]
                 except:
                     y_dummy = pd.concat([y, pd.Series([-1] * self.n_splits)], ignore_index=True)
                     X_dummy = pd.concat([X, X.head(self.n_splits)], ignore_index=True)
                     invalid_index = set(list(y_dummy.tail(self.n_splits).index))
-                    out = [[train_index, test_index] for train_index, test_index in self._splitter.split(X_dummy, y_dummy)]
+                    out = [[train_index, test_index] for train_index, test_index in splitter.split(X_dummy, y_dummy)]
                     len_out = len(out)
                     for i in range(len_out):
                         train_index, test_index = out[i]
@@ -87,7 +149,7 @@ class CVSplitter:
                         out[i][1] = [index for index in test_index if index not in invalid_index]
             return out
         else:
-            return [[train_index, test_index] for train_index, test_index in self._splitter.split(X, y, groups=self.groups)]
+            return [[train_index, test_index] for train_index, test_index in splitter.split(X, y, groups=self.groups)]
 
 
 def setup_compute(nthreads_per_trial, ngpus_per_trial):

autogluon/core/version.py

@@ -1,3 +1,3 @@
 """This is the autogluon version file."""
-__version__ = '1.2.1b20250113'
+__version__ = '1.2.1b20250114'
 __lite__ = False

{autogluon.core-1.2.1b20250113.dist-info → autogluon.core-1.2.1b20250114.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: autogluon.core
-Version: 1.2.1b20250113
+Version: 1.2.1b20250114
 Summary: Fast and Accurate ML in 3 Lines of Code
 Home-page: https://github.com/autogluon/autogluon
 Author: AutoGluon Community
@@ -43,12 +43,12 @@ Requires-Dist: tqdm<5,>=4.38
 Requires-Dist: requests
 Requires-Dist: matplotlib<3.11,>=3.7.0
 Requires-Dist: boto3<2,>=1.10
-Requires-Dist: autogluon.common==1.2.1b20250113
+Requires-Dist: autogluon.common==1.2.1b20250114
 Provides-Extra: all
 Requires-Dist: ray[default,tune]<2.41,>=2.10.0; extra == "all"
-Requires-Dist: hyperopt<0.2.8,>=0.2.7; extra == "all"
-Requires-Dist: pyarrow>=15.0.0; extra == "all"
 Requires-Dist: ray[default]<2.41,>=2.10.0; extra == "all"
+Requires-Dist: pyarrow>=15.0.0; extra == "all"
+Requires-Dist: hyperopt<0.2.8,>=0.2.7; extra == "all"
 Provides-Extra: ray
 Requires-Dist: ray[default]<2.41,>=2.10.0; extra == "ray"
 Provides-Extra: raytune
@@ -56,11 +56,11 @@ Requires-Dist: pyarrow>=15.0.0; extra == "raytune"
 Requires-Dist: ray[default,tune]<2.41,>=2.10.0; extra == "raytune"
 Requires-Dist: hyperopt<0.2.8,>=0.2.7; extra == "raytune"
 Provides-Extra: tests
-Requires-Dist: types-requests; extra == "tests"
-Requires-Dist: pytest-mypy; extra == "tests"
+Requires-Dist: flake8; extra == "tests"
 Requires-Dist: pytest; extra == "tests"
+Requires-Dist: pytest-mypy; extra == "tests"
+Requires-Dist: types-requests; extra == "tests"
 Requires-Dist: types-setuptools; extra == "tests"
-Requires-Dist: flake8; extra == "tests"
 
 
 

{autogluon.core-1.2.1b20250113.dist-info → autogluon.core-1.2.1b20250114.dist-info}/RECORD

@@ -1,9 +1,9 @@
-autogluon.core-1.2.1b20250113-py3.8-nspkg.pth,sha256=cQGwpuGPqg1GXscIwt-7PmME1OnSpD-7ixkikJ31WAY,554
+autogluon.core-1.2.1b20250114-py3.8-nspkg.pth,sha256=cQGwpuGPqg1GXscIwt-7PmME1OnSpD-7ixkikJ31WAY,554
 autogluon/core/__init__.py,sha256=8KfvvHzXX3a4q6z43Dw1yE7VtbAoiSMaglVpKDy6Xeg,245
 autogluon/core/_setup_utils.py,sha256=NqlGK6So0KG5M0LbBJNT1TI3iAmG93kd_6Brih6y2gQ,6935
 autogluon/core/constants.py,sha256=nEVLdSFJ-5O-tz3jUD3qPX65RMp7g8qOR38XlurbP4Y,3403
 autogluon/core/problem_type.py,sha256=XJmMgeNBgS7u43pDK-spTivatPyh_INOXveEXwQt-Rw,2993
-autogluon/core/version.py,sha256=xp_3RhUZoc8du55mJV_5b3-ffuLY-6zGvseBIwCs1Z0,90
+autogluon/core/version.py,sha256=Um-j8InMDDRIJ5WoL607BMYYvkzsVWU8cxXJ2aXfjbQ,90
 autogluon/core/augmentation/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 autogluon/core/augmentation/distill_utils.py,sha256=JBlp2WOMNKoJv8aKVwJVRQSalSk8jx36HM7-k_VvkhY,9404
 autogluon/core/calibrate/__init__.py,sha256=eU6qLj7DKUhaz2HHNHDrfroRaLM-mhuSncK_v1UP4F8,62
@@ -47,7 +47,7 @@ autogluon/core/models/dummy/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMp
 autogluon/core/models/dummy/_dummy_quantile_regressor.py,sha256=i-ZW2flJ60jsMfMK24IP39Xwc55-UlBDvHmqanIf29Q,664
 autogluon/core/models/dummy/dummy_model.py,sha256=at2FZSM2_LuAQ78E2YrRCRt3UaKMyyOnc6p2rtZgA2w,1414
 autogluon/core/models/ensemble/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-autogluon/core/models/ensemble/bagged_ensemble_model.py,sha256=HuyRqdtsdN2z_t9Fa9qWN3U5dz3O7MGYn2qe4BKU9Go,71600
+autogluon/core/models/ensemble/bagged_ensemble_model.py,sha256=p3lkm0Cweyu_maOeboirGmQIjIaaiDHXS-XHDc-wYYw,72610
 autogluon/core/models/ensemble/fold_fitting_strategy.py,sha256=01vzNVvE4FIFgD6YqbhK63XoUlSztnVFsrDdsoqm75U,47021
 autogluon/core/models/ensemble/ray_parallel_fold_fitting_strategy.py,sha256=8RASa-eV6n9kUgbqQHNt7k4IrvuB9NdrunIMLYOLwgA,2068
 autogluon/core/models/ensemble/stacker_ensemble_model.py,sha256=DuDXgozvG9JYYkRvGACA7EXDAtj3Tz_uAjXTfxu5tFg,18041
@@ -85,15 +85,15 @@ autogluon/core/utils/infer_utils.py,sha256=In-u47ELeMLol-O02-bbo79Ak2TYo6iuQ7SJk
 autogluon/core/utils/miscs.py,sha256=Kh5tXxyGHWfEEb36t3iBpdhJI7V4Pgih6xN_bpiQwFg,369
 autogluon/core/utils/plots.py,sha256=ambgOSY7Hp--CDU6raghZx6O8m4a49spdHRG3zEMypk,12238
 autogluon/core/utils/time.py,sha256=WaTonKWCSO5BoSIPo_hdgreGiQxoYE2dYWKO3aPFHj4,3841
-autogluon/core/utils/utils.py,sha256=K05ewQuGauLnVaYwccNDk1moUDg2EEzdSlq8gsw6JVM,55071
+autogluon/core/utils/utils.py,sha256=FMa9kIUAxA3IIBbATmBnNEVObSAivehZ2_zCy3PRR-c,57660
 autogluon/core/utils/version_utils.py,sha256=5-r8hLRKTaZbj5qo2uzE_2E4casH49Ye3WyeHlgHuz4,3252
 autogluon/core/utils/loaders/__init__.py,sha256=W5FAdQvpDcn_uisqJrlSAObWVta-YjJLKGN3NCbEgIo,109
 autogluon/core/utils/savers/__init__.py,sha256=bGWciSxAkj6u06vOC4pTvr22f_1ey0glgvmjCMEOm78,89
-autogluon.core-1.2.1b20250113.dist-info/LICENSE,sha256=CeipvOyAZxBGUsFoaFqwkx54aPnIKEtm9a5u2uXxEws,10142
-autogluon.core-1.2.1b20250113.dist-info/METADATA,sha256=j-_5gd4PPTvZsdGC2daWT2OV-6A0Pdf4C20Qq-8_oC4,12328
-autogluon.core-1.2.1b20250113.dist-info/NOTICE,sha256=7nPQuj8Kp-uXsU0S5so3-2dNU5EctS5hDXvvzzehd7E,114
-autogluon.core-1.2.1b20250113.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
-autogluon.core-1.2.1b20250113.dist-info/namespace_packages.txt,sha256=giERA4R78OkJf2ijn5slgjURlhRPzfLr7waIcGkzYAo,10
-autogluon.core-1.2.1b20250113.dist-info/top_level.txt,sha256=giERA4R78OkJf2ijn5slgjURlhRPzfLr7waIcGkzYAo,10
-autogluon.core-1.2.1b20250113.dist-info/zip-safe,sha256=AbpHGcgLb-kRsJGnwFEktk7uzpZOCcBY74-YBdrKVGs,1
-autogluon.core-1.2.1b20250113.dist-info/RECORD,,
+autogluon.core-1.2.1b20250114.dist-info/LICENSE,sha256=CeipvOyAZxBGUsFoaFqwkx54aPnIKEtm9a5u2uXxEws,10142
+autogluon.core-1.2.1b20250114.dist-info/METADATA,sha256=qZ8UNbAq_mHz_lqcg-_xSY0z9dXa8fNDuHOncGl7ICg,12328
+autogluon.core-1.2.1b20250114.dist-info/NOTICE,sha256=7nPQuj8Kp-uXsU0S5so3-2dNU5EctS5hDXvvzzehd7E,114
+autogluon.core-1.2.1b20250114.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
+autogluon.core-1.2.1b20250114.dist-info/namespace_packages.txt,sha256=giERA4R78OkJf2ijn5slgjURlhRPzfLr7waIcGkzYAo,10
+autogluon.core-1.2.1b20250114.dist-info/top_level.txt,sha256=giERA4R78OkJf2ijn5slgjURlhRPzfLr7waIcGkzYAo,10
+autogluon.core-1.2.1b20250114.dist-info/zip-safe,sha256=AbpHGcgLb-kRsJGnwFEktk7uzpZOCcBY74-YBdrKVGs,1
+autogluon.core-1.2.1b20250114.dist-info/RECORD,,