yobx.sklearn.multiclass.output_code#

yobx.sklearn.multiclass.output_code.sklearn_output_code_classifier(g: GraphBuilderExtendedProtocol, sts: Dict, outputs: List[str], estimator: OutputCodeClassifier, X: str, name: str = 'output_code') → str[source]#

Converts a sklearn.multiclass.OutputCodeClassifier into ONNX.

The converter iterates over the fitted binary sub-estimators, calls the registered converter for each one to obtain per-class positive-class probabilities (predict_proba[:, 1]), stacks them into a score matrix Y of shape (N, M) (where M is the number of sub-estimators), and then finds the nearest row in code_book_ using squared Euclidean distance.

Note

sklearn’s predict() uses _predict_binary(), which calls decision_function when available and falls back to predict_proba[:, 1]. This ONNX converter always uses predict_proba[:, 1] for all sub-estimators, matching sklearn exactly for classifiers that expose only predict_proba (such as DecisionTreeClassifier). For classifiers with a decision_function (e.g. LogisticRegression), the ONNX output may differ from sklearn’s prediction in borderline cases.

Two distance-computation paths:

With com.microsoft opset (CDist fast path):

X --[sub-est 0]--> proba_0 (N,2) --Slice[:,1]--> pred_0 (N,1) --+
X --[sub-est 1]--> proba_1 (N,2) --Slice[:,1]--> pred_1 (N,1) --| Concat
...                                                              |  axis=1
X --[sub-est M-1]-> proba_{M-1}  --Slice[:,1]--> pred_{M-1}   --+
                                                                 |
                                                                Y (N,M)
                                                                 |
code_book_ (C,M) --com.microsoft.CDist(sqeuclidean)---------> sq_dists (N,C)
                                                                 |
                                                ArgMin(axis=1) --+-> label_idx (N,)
                                                                 |
                     Gather(classes_, label_idx) -------------> label (N,)

Without com.microsoft opset (standard ONNX path):

X --[sub-est 0]--> proba_0 (N,2) --Slice[:,1]--> pred_0 (N,1) --+
X --[sub-est 1]--> proba_1 (N,2) --Slice[:,1]--> pred_1 (N,1) --| Concat
...                                                              |  axis=1
X --[sub-est M-1]-> proba_{M-1}  --Slice[:,1]--> pred_{M-1}   --+
                                                                 |
                                                                Y (N,M)
                                                                 |
code_book_T (M,C) --MatMul(Y) ---------------------------> cross (N,C)
y_sq (N,M) --ReduceSum(axis=1,keepdims=1) ----------------> y_norms (N,1)
cb_sq (1,C) --------------------------------- Add(y_norms) -> y_plus_cb (N,C)
                        Sub(Mul(2, cross)) -------------> sq_dists (N,C)
                                                                 |
                                                ArgMin(axis=1) --+-> label_idx (N,)
                                                                 |
                     Gather(classes_, label_idx) -------------> label (N,)

Parameters:

g – the graph builder to add nodes to
sts – shapes and types defined by scikit-learn
outputs – desired output tensor names (label only; OutputCodeClassifier has no predict_proba)
estimator – a fitted OutputCodeClassifier
X – name of the input tensor
name – prefix used for names of nodes added by this converter

Returns:

label tensor name

Raises:

NotImplementedError – when a sub-estimator does not expose predict_proba()

yobx.sklearn.multiclass.output_code#

This Page