yobx.sklearn.gaussian_process.gp#

yobx.sklearn.gaussian_process.gp.sklearn_gaussian_process_classifier(g: GraphBuilderExtendedProtocol, sts: Dict, outputs: List[str], estimator: GaussianProcessClassifier, X: str, name: str = 'gpc') → Tuple[str, str][source]#

Converts a sklearn.gaussian_process.GaussianProcessClassifier into ONNX.

Binary classification (two classes):

K_trans  = kernel(X, X_train)                      # (N, M)
f_mean   = K_trans @ (y_train - π)                 # (N,)   posterior mean
v        = M_pre @ K_trans.T                       # (M, N) M_pre precomputed
f_var    = diag(K(X,X)) - Σ_col(v²)              # (N,)   posterior variance
π*       ≈ Williams-Barber 5-point approximation   # (N,)
proba    = [[1-π*, π*]]                             # (N, 2)

where M_pre = L⁻¹ · diag(W_sr) is precomputed once at conversion time.

Multiclass (multi_class="one_vs_rest"):

Each binary sub-estimator produces π*ₖ; these are stacked and row-normalised:

Y[:, k]  = π*ₖ  for k in 0…C-1
proba    = Y / row_sum(Y)

Supported kernels — see _emit_kernel_matrix() for the full list.

Parameters:

  • g – the graph builder to add nodes to

  • sts – shapes defined by scikit-learn

  • outputs – desired output names (label, probabilities)

  • estimator – a fitted GaussianProcessClassifier

  • X – input tensor name

  • name – prefix for added node names

Returns:

tuple (label_result_name, proba_result_name)

Raises:

NotImplementedError – for multi_class="one_vs_one" with >2 classes

yobx.sklearn.gaussian_process.gp.sklearn_gaussian_process_regressor(g: GraphBuilderExtendedProtocol, sts: Dict, outputs: List[str], estimator: GaussianProcessRegressor, X: str, name: str = 'gpr') → str | Tuple[str, ...][source]#

Converts a sklearn.gaussian_process.GaussianProcessRegressor into ONNX (mean prediction only).

The predictive mean follows Algorithm 2.1 of Gaussian Processes for Machine Learning (Rasmussen & Williams 2006, p. 19):

K_trans = kernel(X, X_train)          # (N, M)
y_mean  = K_trans @ α + y_train_mean  # (N,)  or  (N, n_targets)

where α (estimator.alpha_) and the denormalisation offsets (_y_train_mean, _y_train_std) are stored as ONNX constants.

Supported kernels — see _emit_kernel_matrix() for the full list.

Parameters:

  • g – the graph builder to add nodes to

  • sts – shapes defined by scikit-learn

  • outputs – desired output names

  • estimator – a fitted GaussianProcessRegressor

  • X – input tensor name

  • name – prefix for added node names

Returns:

output tensor name (predictions)