yobx.sklearn.linear_model.linear_classifier

yobx.sklearn.linear_model.linear_classifier.sklearn_linear_classifier(g: GraphBuilderExtendedProtocol, sts: Dict, outputs: List[str], estimator: RidgeClassifier | SGDClassifier | Perceptron, X: str, name: str = 'linear_classifier') → str | Tuple[str, str]

Converts a sklearn linear classifier into ONNX.

Handles both classifiers that expose predict_proba() (e.g. SGDClassifier with loss='log_loss' or 'modified_huber') and those that do not (e.g. RidgeClassifier).

Binary classification (len(classes_) == 2):

X  ──Gemm(coef, intercept)──►  decision (Nx1)
                                    │
                                Flatten  ──►  decision_1d (N,)
                                    │
                          ┌─────────┴──────────┐
                          │                 (if proba)
                      Greater(0)           Sigmoid ──Sub(1,·)──Concat
                          │                                        │
                         Cast(INT64)                            proba (Nx2)
                          │
                     Gather(classes) ──►  label

Multiclass (len(classes_) > 2):

X  ──Gemm(coef, intercept)──►  decision (NxC)
                                    │
                          ┌─────────┴──────────┐
                          │                 (if proba)
                       ArgMax              Softmax ──►  proba (NxC)
                          │
                      Cast(INT64)
                          │
                     Gather(classes) ──►  label

Parameters:

• g – the graph builder to add nodes to

• sts – shapes defined by scikit-learn

• outputs – desired output names; one name (label only) or two names (label + probabilities) depending on get_output_names()

• estimator – a fitted linear classifier

• X – input tensor name

• name – prefix for added node names

Returns:

label tensor name, or tuple (label, probabilities) when the estimator supports predict_proba()