autogluon.tabular 1.2.1b20250430__tar.gz → 1.3.0b20250501__tar.gz

{autogluon.tabular-1.2.1b20250430 → autogluon.tabular-1.3.0b20250501}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: autogluon.tabular
-Version: 1.2.1b20250430
+Version: 1.3.0b20250501
 Summary: Fast and Accurate ML in 3 Lines of Code
 Home-page: https://github.com/autogluon/autogluon
 Author: AutoGluon Community

{autogluon.tabular-1.2.1b20250430 → autogluon.tabular-1.3.0b20250501}/src/autogluon/tabular/configs/config_helper.py

@@ -9,7 +9,7 @@ from autogluon.core.scheduler import scheduler_factory
 from autogluon.features import AutoMLPipelineFeatureGenerator
 from autogluon.tabular.configs.hyperparameter_configs import hyperparameter_config_dict
 from autogluon.tabular.configs.presets_configs import tabular_presets_dict
-from autogluon.tabular.register import ag_model_register
+from autogluon.tabular.registry import ag_model_registry
 
 
 class FeatureGeneratorBuilder:
@@ -108,7 +108,7 @@ class ConfigBuilder:
         self.config = {}
 
     def _valid_keys(self):
-        valid_keys = [m for m in ag_model_register.keys if m not in ["ENS_WEIGHTED", "SIMPLE_ENS_WEIGHTED"]]
+        valid_keys = [m for m in ag_model_registry.keys if m not in ["ENS_WEIGHTED", "SIMPLE_ENS_WEIGHTED"]]
         return valid_keys
 
     def presets(self, presets: Union[str, list, dict]) -> ConfigBuilder:

{autogluon.tabular-1.2.1b20250430 → autogluon.tabular-1.3.0b20250501}/src/autogluon/tabular/models/lgb/callbacks.py

@@ -133,13 +133,23 @@ def early_stopping_custom(
         estimated_model_size_mb = (cur_rss - init_mem_rss[0]) >> 20
         available_mb = available >> 20
 
-        model_size_memory_ratio = estimated_model_size_mb / available_mb
+        if available_mb != 0:
+            model_size_memory_ratio = estimated_model_size_mb / available_mb
+        else:
+            model_size_memory_ratio = 100
+
         if verbose or (model_size_memory_ratio > 0.25):
             logger.debug("Available Memory: " + str(available_mb) + " MB")
             logger.debug("Estimated Model Size: " + str(estimated_model_size_mb) + " MB")
 
         early_stop = False
-        if model_size_memory_ratio > 1.0:
+        # FIXME: during parallel fits, model only knows its own memory usage and the overall system memory.
+        #  Because memory usage can spike during saving, OOM can occur if many models finish at the same time and spike in memory at the same time during save.
+        #  To fix this, we need to provide the per-model memory limit as a constraint passed to this method,
+        #  so that we can ensure a given model isn't exceeding its portion of the memory budget.
+        #  Ditto for XGBoost and CatBoost (ex: "kropt" dataset with 8-fold bagging and 32 GB memory. Fits 10k iterations on all 8 folds, then goes OOM)
+        #  We also need to estimate the peak memory usage given the estimated_model_size_mb if we were to save. Otherwise we will go OOM during save anyways.
+        if model_size_memory_ratio > 0.66:
             logger.warning("Warning: Large GBM model size may cause OOM error if training continues")
             logger.warning("Available Memory: " + str(available_mb) + " MB")
             logger.warning("Estimated GBM model size: " + str(estimated_model_size_mb) + " MB")

{autogluon.tabular-1.2.1b20250430 → autogluon.tabular-1.3.0b20250501}/src/autogluon/tabular/models/lgb/lgb_model.py

@@ -91,12 +91,15 @@ class LGBModel(AbstractModel):
         """
         Returns the expected peak memory usage in bytes of the LightGBM model during fit.
 
-        The memory usage of LightGBM is primarily made up of two sources:
+        The memory usage of LightGBM is primarily made up of three sources:
 
         1. The size of the data
         2. The size of the histogram cache
             Scales roughly by 5100*num_features*num_leaves bytes
             For 10000 features and 128 num_leaves, the histogram would be 6.5 GB.
+        3. The size of the model
+            Scales linearly with the number of estimators, number of classes, and number of leaves.
+            Memory usage peaks during model saving, with the peak consuming approximately 2-4x the size of the model in memory.
         """
         if hyperparameters is None:
             hyperparameters = {}
@@ -104,6 +107,8 @@ class LGBModel(AbstractModel):
         data_mem_usage = get_approximate_df_mem_usage(X).sum()
         data_mem_usage_bytes = data_mem_usage * 5 + data_mem_usage / 4 * num_classes  # TODO: Extremely crude approximation, can be vastly improved
 
+        n_trees_per_estimator = num_classes if num_classes > 2 else 1
+
         max_bins = hyperparameters.get("max_bins", 255)
         num_leaves = hyperparameters.get("num_leaves", 31)
         # Memory usage of histogram based on https://github.com/microsoft/LightGBM/issues/562#issuecomment-304524592
@@ -115,7 +120,12 @@ class LGBModel(AbstractModel):
                 histogram_mem_usage_bytes = histogram_mem_usage_bytes_max
         histogram_mem_usage_bytes *= 1.2  # Add a 20% buffer
 
-        approx_mem_size_req = data_mem_usage_bytes + histogram_mem_usage_bytes
+        mem_size_per_estimator = n_trees_per_estimator * num_leaves * 100  # very rough estimate
+        n_estimators = hyperparameters.get("num_boost_round", DEFAULT_NUM_BOOST_ROUND)
+        n_estimators_min = min(n_estimators, 1000)
+        mem_size_estimators = n_estimators_min * mem_size_per_estimator  # memory estimate after fitting up to 1000 estimators
+
+        approx_mem_size_req = data_mem_usage_bytes + histogram_mem_usage_bytes + mem_size_estimators
         return approx_mem_size_req
 
     def _fit(self, X, y, X_val=None, y_val=None, time_limit=None, num_gpus=0, num_cpus=0, sample_weight=None, sample_weight_val=None, verbosity=2, **kwargs):

{autogluon.tabular-1.2.1b20250430 → autogluon.tabular-1.3.0b20250501}/src/autogluon/tabular/predictor/predictor.py

@@ -55,7 +55,7 @@ from ..configs.hyperparameter_configs import get_hyperparameter_config
 from ..configs.presets_configs import tabular_presets_alias, tabular_presets_dict
 from ..learner import AbstractTabularLearner, DefaultLearner
 from ..trainer.abstract_trainer import AbstractTabularTrainer
-from ..register import ag_model_register
+from ..registry import ag_model_registry
 from ..version import __version__
 
 logger = logging.getLogger(__name__)  # return autogluon root logger
@@ -5672,7 +5672,7 @@ class TabularPredictor:
             for key in hyperparameters:
                 models_in_hyperparameters.add(key)
         models_in_hyperparameters_raw_text_compatible = []
-        model_key_to_cls_map = ag_model_register.key_to_cls_map()
+        model_key_to_cls_map = ag_model_registry.key_to_cls_map()
         for m in models_in_hyperparameters:
             if isinstance(m, str):
                 # TODO: Technically the use of MODEL_TYPES here is a hack since we should derive valid types from trainer,

autogluon.tabular-1.3.0b20250501/src/autogluon/tabular/registry/__init__.py

@@ -0,0 +1,2 @@
+from ._model_registry import ModelRegistry
+from ._ag_model_registry import ag_model_registry

autogluon.tabular-1.2.1b20250430/src/autogluon/tabular/register/_ag_model_register.py → autogluon.tabular-1.3.0b20250501/src/autogluon/tabular/registry/_ag_model_registry.py

@@ -4,7 +4,7 @@ from autogluon.core.models import (
     SimpleWeightedEnsembleModel,
 )
 
-from . import ModelRegister
+from . import ModelRegistry
 from ..models import (
     BoostedRulesModel,
     CatBoostModel,
@@ -58,5 +58,5 @@ REGISTERED_MODEL_CLS_LST = [
     DummyModel,
 ]
 
-# TODO: Replace logic in `autogluon.tabular.trainer.model_presets.presets` with `ag_model_register`
-ag_model_register = ModelRegister(model_cls_list=REGISTERED_MODEL_CLS_LST)
+# TODO: Replace logic in `autogluon.tabular.trainer.model_presets.presets` with `ag_model_registry`
+ag_model_registry = ModelRegistry(model_cls_list=REGISTERED_MODEL_CLS_LST)

autogluon.tabular-1.2.1b20250430/src/autogluon/tabular/register/_model_register.py → autogluon.tabular-1.3.0b20250501/src/autogluon/tabular/registry/_model_registry.py

@@ -11,9 +11,9 @@ from autogluon.core.models import AbstractModel
 # TODO: Use this / refer to this in the custom model tutorial
 # TODO: Add to documentation website
 # TODO: Test register logic in AG
-class ModelRegister:
+class ModelRegistry:
     """
-    ModelRegister keeps track of all known model classes to AutoGluon.
+    ModelRegistry keeps track of all known model classes to AutoGluon.
     It can provide information such as:
         What model classes and keys are valid to specify in an AutoGluon predictor fit call.
         What a model's name is.
@@ -21,7 +21,7 @@ class ModelRegister:
         What a model's priority is (aka which order to fit a list of models).
 
     Additionally, users can register custom models to AutoGluon so the key is recognized in `hyperparameters` and is treated with the proper priority and name.
-    They can register new models via `ModelRegister.add(model_cls)`.
+    They can register new models via `ModelRegistry.add(model_cls)`.
 
     Therefore, if a user creates a custom model `MyCustomModel` that inherits from `AbstractModel`, they can set the class attributes in `MyCustomModel`:
         ag_key: The string key that can be specified in `hyperparameters`. Example: "GBM" for LGBModel
@@ -29,7 +29,7 @@ class ModelRegister:
         ag_priority: The int priority that is used to order the fitting of models. Higher values will be fit before lower values. Default 0. Example: 90 for LGBModel
         ag_priority_to_problem_type: A dictionary of problem_type to priority that overrides `ag_priority` if specified for a given problem_type. Optional.
 
-    Then they can say `ag_model_register.add(MyCustomModel)`.
+    Then they can say `ag_model_registry.add(MyCustomModel)`.
     Assuming MyCustomModel.ag_key = "MY_MODEL", they can now do:
     ```
     predictor.fit(..., hyperparameters={"MY_MODEL": ...})
@@ -49,7 +49,7 @@ class ModelRegister:
 
     def add(self, model_cls: Type[AbstractModel]):
         """
-        Adds `model_cls` to the model register
+        Adds `model_cls` to the model registry
         """
         assert not self.exists(model_cls), f"Cannot add model_cls that is already registered: {model_cls}"
         if model_cls.ag_key is None:
@@ -80,7 +80,7 @@ class ModelRegister:
 
     def remove(self, model_cls: Type[AbstractModel]):
         """
-        Removes `model_cls` from the model register
+        Removes `model_cls` from the model registry
         """
         assert self.exists(model_cls), f"Cannot remove model_cls that isn't registered: {model_cls}"
         self._model_cls_list = [m for m in self._model_cls_list if m != model_cls]

{autogluon.tabular-1.2.1b20250430 → autogluon.tabular-1.3.0b20250501}/src/autogluon/tabular/trainer/auto_trainer.py

@@ -7,7 +7,7 @@ from ..models.lgb.lgb_model import LGBModel
 from .abstract_trainer import AbstractTabularTrainer
 from .model_presets.presets import get_preset_models
 from .model_presets.presets_distill import get_preset_models_distillation
-from ..register import ag_model_register
+from ..registry import ag_model_registry
 
 logger = logging.getLogger(__name__)
 
@@ -187,4 +187,4 @@ class AutoTrainer(AbstractTabularTrainer):
         return super().compile(model_names=model_names, with_ancestors=with_ancestors, compiler_configs=compiler_configs)
 
     def _get_model_types_map(self) -> dict[str, AbstractModel]:
-        return ag_model_register.key_to_cls_map()
+        return ag_model_registry.key_to_cls_map()

{autogluon.tabular-1.2.1b20250430 → autogluon.tabular-1.3.0b20250501}/src/autogluon/tabular/trainer/model_presets/presets.py

@@ -28,7 +28,7 @@ from autogluon.core.models import (
 )
 from autogluon.core.trainer.utils import process_hyperparameters
 
-from ...register import ag_model_register
+from ...registry import ag_model_registry
 from ...version import __version__
 
 logger = logging.getLogger(__name__)
@@ -99,9 +99,9 @@ def get_preset_models(
         invalid_name_set.update(invalid_model_names)
 
     if default_priorities is None:
-        priority_cls_map = ag_model_register.priority_map(problem_type=problem_type)
+        priority_cls_map = ag_model_registry.priority_map(problem_type=problem_type)
         default_priorities = {
-            ag_model_register.key(model_cls): priority for model_cls, priority in priority_cls_map.items()
+            ag_model_registry.key(model_cls): priority for model_cls, priority in priority_cls_map.items()
         }
 
     level_key = level if level in hyperparameters.keys() else "default"
@@ -175,7 +175,7 @@ def clean_model_cfg(model_cfg: dict, model_type=None, ag_args=None, ag_args_ense
     if model_cfg[AG_ARGS]["model_type"] is None:
         raise AssertionError(f"model_type was not specified for model! Model: {model_cfg}")
     model_type = model_cfg[AG_ARGS]["model_type"]
-    model_types = ag_model_register.key_to_cls_map()
+    model_types = ag_model_registry.key_to_cls_map()
     if not inspect.isclass(model_type):
         if model_type not in model_types:
             raise AssertionError(f"Unknown model type specified in hyperparameters: '{model_type}'. Valid model types: {list(model_types.keys())}")
@@ -185,7 +185,7 @@ def clean_model_cfg(model_cfg: dict, model_type=None, ag_args=None, ag_args_ense
             f"Warning: Custom model type {model_type} does not inherit from {AbstractModel}. This may lead to instability. Consider wrapping {model_type} with an implementation of {AbstractModel}!"
         )
     else:
-        if not ag_model_register.exists(model_type):
+        if not ag_model_registry.exists(model_type):
             logger.log(20, f"Custom Model Type Detected: {model_type}")
     model_cfg[AG_ARGS]["model_type"] = model_type
     model_type_real = model_cfg[AG_ARGS]["model_type"]
@@ -280,7 +280,7 @@ def model_factory(
         invalid_name_set = set()
     model_type = model[AG_ARGS]["model_type"]
     if not inspect.isclass(model_type):
-        model_type = ag_model_register.key_to_cls(model_type)
+        model_type = ag_model_registry.key_to_cls(model_type)
     name_orig = model[AG_ARGS].get("name", None)
     if name_orig is None:
         ag_name = model_type.ag_name

{autogluon.tabular-1.2.1b20250430 → autogluon.tabular-1.3.0b20250501}/src/autogluon/tabular/version.py

@@ -1,4 +1,4 @@
 """This is the autogluon version file."""
 
-__version__ = "1.2.1b20250430"
+__version__ = "1.3.0b20250501"
 __lite__ = False

{autogluon.tabular-1.2.1b20250430 → autogluon.tabular-1.3.0b20250501}/src/autogluon.tabular.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: autogluon.tabular
-Version: 1.2.1b20250430
+Version: 1.3.0b20250501
 Summary: Fast and Accurate ML in 3 Lines of Code
 Home-page: https://github.com/autogluon/autogluon
 Author: AutoGluon Community

{autogluon.tabular-1.2.1b20250430 → autogluon.tabular-1.3.0b20250501}/src/autogluon.tabular.egg-info/SOURCES.txt

@@ -140,9 +140,9 @@ src/autogluon/tabular/models/xt/xt_model.py
 src/autogluon/tabular/predictor/__init__.py
 src/autogluon/tabular/predictor/interpretable_predictor.py
 src/autogluon/tabular/predictor/predictor.py
-src/autogluon/tabular/register/__init__.py
-src/autogluon/tabular/register/_ag_model_register.py
-src/autogluon/tabular/register/_model_register.py
+src/autogluon/tabular/registry/__init__.py
+src/autogluon/tabular/registry/_ag_model_registry.py
+src/autogluon/tabular/registry/_model_registry.py
 src/autogluon/tabular/testing/__init__.py
 src/autogluon/tabular/testing/fit_helper.py
 src/autogluon/tabular/testing/generate_datasets.py

{autogluon.tabular-1.2.1b20250430 → autogluon.tabular-1.3.0b20250501}/src/autogluon.tabular.egg-info/requires.txt

@@ -3,20 +3,20 @@ scipy<1.16,>=1.5.4
 pandas<2.3.0,>=2.0.0
 scikit-learn<1.7.0,>=1.4.0
 networkx<4,>=3.0
-autogluon.core==1.2.1b20250430
-autogluon.features==1.2.1b20250430
+autogluon.core==1.3.0b20250501
+autogluon.features==1.3.0b20250501
 
 [all]
-numpy<2.3.0,>=1.25
-fastai<2.9,>=2.3.1
-lightgbm<4.7,>=4.0
-torch<2.7,>=2.2
-einops<0.9,>=0.7
-spacy<3.9
 huggingface_hub[torch]
-autogluon.core[all]==1.2.1b20250430
+einops<0.9,>=0.7
+numpy<2.3.0,>=1.25
+autogluon.core[all]==1.3.0b20250501
 xgboost<3.1,>=2.0
+spacy<3.9
+lightgbm<4.7,>=4.0
 catboost<1.3,>=1.2
+fastai<2.9,>=2.3.1
+torch<2.7,>=2.2
 
 [catboost]
 numpy<2.3.0,>=1.25
@@ -34,7 +34,7 @@ imodels<2.1.0,>=1.3.10
 lightgbm<4.7,>=4.0
 
 [ray]
-autogluon.core[all]==1.2.1b20250430
+autogluon.core[all]==1.3.0b20250501
 
 [skex]
 scikit-learn-intelex<2025.5,>=2024.0

autogluon.tabular-1.2.1b20250430/src/autogluon/tabular/register/__init__.py

@@ -1,2 +0,0 @@
-from ._model_register import ModelRegister
-from ._ag_model_register import ag_model_register