Introduces API to get inference config, removes unused inference config defaults (#890)

Author: bejaeger

changelog/890.added.md

@@ -0,0 +1 @@
+Add `get_inference_config()` method to `TabPFNClassifier` and `TabPFNRegressor`. This method loads the model checkpoint if needed and returns the active `InferenceConfig`, allowing inspection of preprocessing and inference settings before calling `fit()`.

changelog/890.changed.md

@@ -0,0 +1 @@
+Remove unused v2.6 defaults from `InferenceConfig.get_default()`. V2.6 checkpoints always embed their own `InferenceConfig`, so these defaults were never used at inference time. The v2.6 preprocessor config factories are also removed from `tabpfn.preprocessing`.

src/tabpfn/classifier.py

@@ -551,6 +551,20 @@ def model_(self) -> Architecture:
             )
         return self.models_[0]
 
+    def get_inference_config(self) -> InferenceConfig:
+        """Load the model if needed and return the active inference config.
+
+        Loads the model checkpoint without requiring fit data so the config can be
+        inspected before calling `fit()`. Any ``inference_config`` override
+        passed to the constructor is considered.
+
+        Returns:
+            A deep copy of the active inference config.
+        """
+        if not hasattr(self, "inference_config_"):
+            self._initialize_model_variables()
+        return copy.deepcopy(self.inference_config_)
+
     # TODO: We can remove this from scikit-learn lower bound of 1.6
     def _more_tags(self) -> dict[str, Any]:
         return {

src/tabpfn/inference_config.py

@@ -15,8 +15,6 @@
     PreprocessorConfig,
     v2_5_classifier_preprocessor_configs,
     v2_5_regressor_preprocessor_configs,
-    v2_6_classifier_preprocessor_configs,
-    v2_6_regressor_preprocessor_configs,
     v2_classifier_preprocessor_configs,
     v2_regressor_preprocessor_configs,
 )
@@ -260,18 +258,6 @@ def get_default(
                 return _get_v2_5_config(v2_5_classifier_preprocessor_configs())
             if task_type == "regression":
                 return _get_v2_5_config(v2_5_regressor_preprocessor_configs())
-        elif model_version == ModelVersion.V2_6:
-            if task_type == "multiclass":
-                return _get_v2_6_config(
-                    preprocessor_configs=v2_6_classifier_preprocessor_configs(),
-                    task_type=task_type,
-                )
-            if task_type == "regression":
-                return _get_v2_6_config(
-                    preprocessor_configs=v2_6_regressor_preprocessor_configs(),
-                    task_type=task_type,
-                )
-
         raise ValueError(
             f"No inference config is configured for {model_version=}. "
             "Please make sure you are using a correct model checkpoint that contains "
@@ -327,31 +313,3 @@ def _get_v2_5_config(preprocessor_configs: list[PreprocessorConfig]) -> Inferenc
         _REGRESSION_DEFAULT_OUTLIER_REMOVAL_STD=None,
         _CLASSIFICATION_DEFAULT_OUTLIER_REMOVAL_STD=12.0,
     )
-
-
-def _get_v2_6_config(
-    preprocessor_configs: list[PreprocessorConfig],
-    task_type: TaskType,
-) -> InferenceConfig:
-    return InferenceConfig(
-        MAX_UNIQUE_FOR_CATEGORICAL_FEATURES=30,
-        MIN_UNIQUE_FOR_NUMERICAL_FEATURES=4,
-        MIN_NUMBER_SAMPLES_FOR_CATEGORICAL_INFERENCE=100,
-        OUTLIER_REMOVAL_STD=None,
-        FEATURE_SHIFT_METHOD="shuffle",
-        CLASS_SHIFT_METHOD="shuffle",
-        FINGERPRINT_FEATURE=True,
-        POLYNOMIAL_FEATURES="no" if task_type == "multiclass" else 10,
-        SUBSAMPLE_SAMPLES=None,
-        FEATURE_SUBSAMPLING_METHOD="random",
-        CONSTANT_FEATURE_COUNT=50,
-        PREPROCESS_TRANSFORMS=preprocessor_configs,
-        REGRESSION_Y_PREPROCESS_TRANSFORMS=("none",),
-        USE_SKLEARN_16_DECIMAL_PRECISION=False,
-        MAX_NUMBER_OF_CLASSES=10,
-        MAX_NUMBER_OF_FEATURES=2000,
-        MAX_NUMBER_OF_SAMPLES=50_000,
-        FIX_NAN_BORDERS_AFTER_TARGET_TRANSFORM=True,
-        _REGRESSION_DEFAULT_OUTLIER_REMOVAL_STD=None,
-        _CLASSIFICATION_DEFAULT_OUTLIER_REMOVAL_STD=12.0,
-    )

src/tabpfn/preprocessing/__init__.py

@@ -16,8 +16,6 @@
 from .presets import (
     v2_5_classifier_preprocessor_configs,
     v2_5_regressor_preprocessor_configs,
-    v2_6_classifier_preprocessor_configs,
-    v2_6_regressor_preprocessor_configs,
     v2_classifier_preprocessor_configs,
     v2_regressor_preprocessor_configs,
 )
@@ -36,8 +34,6 @@
     "generate_regression_ensemble_configs",
     "v2_5_classifier_preprocessor_configs",
     "v2_5_regressor_preprocessor_configs",
-    "v2_6_classifier_preprocessor_configs",
-    "v2_6_regressor_preprocessor_configs",
     "v2_classifier_preprocessor_configs",
     "v2_regressor_preprocessor_configs",
 ]

src/tabpfn/preprocessing/presets.py

@@ -80,52 +80,10 @@ def v2_5_regressor_preprocessor_configs() -> list[PreprocessorConfig]:
     ]
 
 
-def v2_6_classifier_preprocessor_configs() -> list[PreprocessorConfig]:
-    """Get the preprocessor configuration for classification in v2.5 of the model."""
-    return [
-        PreprocessorConfig(
-            name="quantile_uni",
-            append_original=False,
-            categorical_name="numeric",
-            global_transformer_name=None,
-            max_features_per_estimator=680,
-        ),
-        PreprocessorConfig(
-            name="quantile_uni",
-            append_original=False,
-            categorical_name="ordinal_very_common_categories_shuffled",
-            global_transformer_name="svd_quarter_components",
-            max_features_per_estimator=500,
-        ),
-    ]
-
-
-def v2_6_regressor_preprocessor_configs() -> list[PreprocessorConfig]:
-    """Get the preprocessor configuration for regression in v2.5 of the model."""
-    return [
-        PreprocessorConfig(
-            name="quantile_uni",
-            append_original=False,
-            categorical_name="numeric",
-            global_transformer_name=None,
-            max_features_per_estimator=680,
-        ),
-        PreprocessorConfig(
-            name="quantile_uni",
-            append_original="auto",
-            categorical_name="ordinal_very_common_categories_shuffled",
-            global_transformer_name="svd_quarter_components",
-            max_features_per_estimator=500,
-        ),
-    ]
-
-
 __all__ = [
     "_V2_FEATURE_SUBSAMPLING_THRESHOLD",
     "v2_5_classifier_preprocessor_configs",
     "v2_5_regressor_preprocessor_configs",
-    "v2_6_classifier_preprocessor_configs",
-    "v2_6_regressor_preprocessor_configs",
     "v2_classifier_preprocessor_configs",
     "v2_regressor_preprocessor_configs",
 ]

src/tabpfn/regressor.py

@@ -17,6 +17,7 @@
 
 from __future__ import annotations
 
+import copy
 import logging
 import typing
 import warnings
@@ -581,6 +582,20 @@ def bardist_(self, value: FullSupportBarDistribution) -> None:
         )
         self.znorm_space_bardist_ = value
 
+    def get_inference_config(self) -> InferenceConfig:
+        """Load the model if needed and return the active inference config.
+
+        Loads the model checkpoint without requiring fit data so the config can be
+        inspected before calling `fit()`. Any ``inference_config`` override
+        passed to the constructor is considered.
+
+        Returns:
+            A deep copy of the active inference config.
+        """
+        if not hasattr(self, "inference_config_"):
+            self._initialize_model_variables()
+        return copy.deepcopy(self.inference_config_)
+
     # TODO: We can remove this from scikit-learn lower bound of 1.6
     def _more_tags(self) -> dict[str, Any]:
         return {

tests/test_inference_config.py

@@ -7,6 +7,11 @@
 
 import torch
 
+from tabpfn import TabPFNClassifier, TabPFNRegressor
+from tabpfn.architectures import base
+from tabpfn.architectures.base.bar_distribution import FullSupportBarDistribution
+from tabpfn.architectures.base.config import ModelConfig
+from tabpfn.base import ClassifierModelSpecs, RegressorModelSpecs
 from tabpfn.constants import ModelVersion
 from tabpfn.inference_config import InferenceConfig
 from tabpfn.preprocessing import PreprocessorConfig
@@ -69,3 +74,108 @@ def test__override_with_user_input__override_is_None__returns_copy_of_config() -
     new_config = config.override_with_user_input_and_resolve_auto(user_config=None)
     assert new_config is not config
     assert new_config == config
+
+
+def _make_classifier_specs() -> ClassifierModelSpecs:
+    config = ModelConfig(
+        emsize=8,
+        features_per_group=1,
+        max_num_classes=10,
+        nhead=2,
+        nlayers=2,
+        remove_duplicate_features=True,
+        num_buckets=100,
+    )
+    model = base.get_architecture(config=config, cache_trainset_representation=False)
+    inference_config = InferenceConfig.get_default(
+        task_type="multiclass", model_version=ModelVersion.V2_5
+    )
+    return ClassifierModelSpecs(
+        model=model,
+        architecture_config=config,
+        inference_config=inference_config,
+    )
+
+
+def _make_regressor_specs() -> RegressorModelSpecs:
+    config = ModelConfig(
+        emsize=8,
+        features_per_group=1,
+        max_num_classes=10,
+        nhead=2,
+        nlayers=2,
+        remove_duplicate_features=True,
+        num_buckets=100,
+    )
+    model = base.get_architecture(config=config, cache_trainset_representation=False)
+    borders = torch.linspace(-3, 3, config.num_buckets + 1)
+    norm_criterion = FullSupportBarDistribution(borders)
+    inference_config = InferenceConfig.get_default(
+        task_type="regression", model_version=ModelVersion.V2_5
+    )
+    return RegressorModelSpecs(
+        model=model,
+        architecture_config=config,
+        inference_config=inference_config,
+        norm_criterion=norm_criterion,
+    )
+
+
+def test__classifier_get_inference_config__before_fit__returns_config() -> None:
+    specs = _make_classifier_specs()
+    clf = TabPFNClassifier(model_path=specs, device="cpu")
+    assert not hasattr(clf, "inference_config_")
+    config = clf.get_inference_config()
+    assert isinstance(config, InferenceConfig)
+    assert config == specs.inference_config
+
+
+def test__classifier_get_inference_config__returns_deepcopy() -> None:
+    specs = _make_classifier_specs()
+    clf = TabPFNClassifier(model_path=specs, device="cpu")
+    config = clf.get_inference_config()
+    assert config is not clf.inference_config_
+    config.PREPROCESS_TRANSFORMS.clear()
+    assert len(clf.inference_config_.PREPROCESS_TRANSFORMS) > 0
+
+
+def test__classifier_get_inference_config__with_override__applies_override() -> None:
+    specs = _make_classifier_specs()
+    clf = TabPFNClassifier(
+        model_path=specs,
+        device="cpu",
+        inference_config={"POLYNOMIAL_FEATURES": "all"},
+    )
+    config = clf.get_inference_config()
+    assert config.POLYNOMIAL_FEATURES == "all"
+    assert specs.inference_config.POLYNOMIAL_FEATURES == "no"
+
+
+def test__regressor_get_inference_config__before_fit__returns_config() -> None:
+    specs = _make_regressor_specs()
+    reg = TabPFNRegressor(model_path=specs, device="cpu")
+    assert not hasattr(reg, "inference_config_")
+    config = reg.get_inference_config()
+    assert isinstance(config, InferenceConfig)
+    assert config == specs.inference_config
+
+
+def test__regressor_get_inference_config__returns_deepcopy() -> None:
+    specs = _make_regressor_specs()
+    reg = TabPFNRegressor(model_path=specs, device="cpu")
+    config = reg.get_inference_config()
+    assert config is not reg.inference_config_
+    config.PREPROCESS_TRANSFORMS.clear()
+    assert len(reg.inference_config_.PREPROCESS_TRANSFORMS) > 0
+
+
+def test__regressor_get_inference_config__with_override__applies_override() -> None:
+    specs = _make_regressor_specs()
+    reg = TabPFNRegressor(
+        model_path=specs,
+        device="cpu",
+        inference_config={"POLYNOMIAL_FEATURES": "all"},
+    )
+    config = reg.get_inference_config()
+    assert config.POLYNOMIAL_FEATURES == "all"
+    assert specs.inference_config.POLYNOMIAL_FEATURES == "no"