nirs4all.operators.models.sklearn.tabpfn_nirs module

TabPFN NIRS regressor with a fixed, empirically-validated preprocessing.

This module provides TabPFNNIRSRegressor, a sklearn-compatible NIRS regressor that pairs a fixed preprocessing chain with the TabPFN tabular foundation model. The recipe avoids per-dataset HPO entirely.

Recipe

SG(window_length=11, polyorder=2, deriv=1) ▶ OSC() ▶ centering (with_std=False) ▶ TabPFNRegressor(n_estimators=16).

If the post-OSC matrix has more than max_features columns, an evenly spaced linspace subsample brings it back below the TabPFN v3 hard limit of 2000 features.

Empirical results on a 57-dataset NIRS cohort (regression, paper splits) versus the TabPFN paper’s 72-chain per-dataset HPO (TabPFN_opt):

  • median RMSE delta: ~+2.5%

  • worst case: +47% on one dataset (Brix_spxy70 at n=35)

  • ~21 of 57 datasets beat TabPFN_opt outright

Cost is a single TabPFN training per dataset (no HPO, no stacking, no CV selector). The recipe was chosen by extracting the most-frequent winning preprocessing motif from the TabPFN paper logs (deriv-1 SG with OSC correction) and a small n_estimators sweep (4/8/16/32) that pinned n_estimators=16 as the sweet spot.

Requirements

  • tabpfn (pip install tabpfn). Imported lazily inside fit(), so importing this module does not require the optional dependency.

class nirs4all.operators.models.sklearn.tabpfn_nirs.TabPFNNIRSRegressor(n_estimators: int = 16, max_features: int = 2000, sg_window_length: int = 11, sg_polyorder: int = 2, sg_deriv: int = 1, osc_n_components: int = 1, random_state: int = 0, device: str = 'auto', model_path: str = 'auto')[source]

Bases: RegressorMixin, BaseEstimator

NIRS-tuned TabPFN regressor with a fixed AGG preprocessing pipeline.

Parameters:

  • n_estimators (int, default=16) – TabPFN ensemble size at final fit. Selected empirically.

  • max_features (int, default=2000) – TabPFN v3 hard limit on input features. If the preprocessed matrix exceeds it, columns are uniformly subsampled via np.linspace.

  • sg_window_length (int, default=11)

  • sg_polyorder (int, default=2)

  • sg_deriv (int, default=1) – Savitzky-Golay smoothing parameters.

  • osc_n_components (int, default=1) – Number of orthogonal-signal-correction components to remove.

  • random_state (int, default=0) – Random seed for TabPFN.

  • device (str, default="auto") – TabPFN device. "auto", "cpu", or "cuda".

  • model_path (str, default="auto") – TabPFN checkpoint path. "auto" loads the default v3 checkpoint.

Variables:

  • sg (SavitzkyGolay) – Fitted SG transformer.

  • osc (OSC) – Fitted OSC transformer.

  • scaler (StandardScaler) – Fitted centering scaler.

  • model (TabPFNRegressor) – Fitted TabPFN model.

  • n_features_in (int) – Number of features seen at fit time.

  • n_features_used (int) – Number of features actually fed to TabPFN (after subsample cap).

  • subsample_idx (ndarray or None) – Column indices used. None when no subsample was applied.

Notes

On a 57-dataset NIRS cohort (paper splits) the fixed recipe lands at median RMSE ~+2.5% versus TabPFN_opt’s 72-chain HPO and beats it on ~37% of datasets, at a single TabPFN training per dataset.

The recipe is intentionally fixed. For per-dataset preprocessing selection, run an external CV loop over candidate chains and feed the chosen ones into this class.

Examples

>>> from nirs4all.operators.models import TabPFNNIRSRegressor
>>> est = TabPFNNIRSRegressor(n_estimators=16)
>>> est.fit(X_train, y_train)
>>> y_pred = est.predict(X_test)
fit(X: ndarray, y: ndarray) → TabPFNNIRSRegressor[source]
predict(X: ndarray) → ndarray[source]
set_score_request(*, sample_weight: bool | None | str = '$UNCHANGED$') → TabPFNNIRSRegressor

Configure whether metadata should be requested to be passed to the score method.

Note that this method is only relevant when this estimator is used as a sub-estimator within a meta-estimator and metadata routing is enabled with enable_metadata_routing=True (see sklearn.set_config()). Please check the User Guide on how the routing mechanism works.

The options for each parameter are:

  • True: metadata is requested, and passed to score if provided. The request is ignored if metadata is not provided.

  • False: metadata is not requested and the meta-estimator will not pass it to score.

  • None: metadata is not requested, and the meta-estimator will raise an error if the user provides it.

  • str: metadata should be passed to the meta-estimator with this given alias instead of the original name.

The default (sklearn.utils.metadata_routing.UNCHANGED) retains the existing request. This allows you to change the request for some parameters and not others.

Added in version 1.3.

Parameters:

sample_weight (str, True, False, or None, default=sklearn.utils.metadata_routing.UNCHANGED) – Metadata routing for sample_weight parameter in score.

Returns:

self – The updated object.

Return type:

object