yobx.sklearn.preprocessing.polynomial_features#

yobx.sklearn.preprocessing.polynomial_features.sklearn_polynomial_features(g: GraphBuilderExtendedProtocol, sts: Dict, outputs: List[str], estimator: PolynomialFeatures, X: str, name: str = 'polynomial_features') → str[source]#

Converts a sklearn.preprocessing.PolynomialFeatures into ONNX.

The output is the matrix of all polynomial combinations of the input features up to the given degree. Each output feature corresponds to one row of estimator.powers_, where each row defines the exponent for each input feature.

The graph is built as follows:

X  ──Unsqueeze(axis=1)──►  X_3d (N, 1, F)
                               │
                   ┌───────────┤
           powers_3d           │   (1, K, F) constant
           zero_mask_3d        │   (1, K, F) constant bool
                   │           │
                Where(zero_mask_3d, 1.0, X_3d) ──► X_safe (N, K, F)
                               │
                Pow(X_safe, powers_3d) ──► powered (N, K, F)
                               │
                ReduceProd(axis=-1) ──► output (N, K)

where K = n_output_features_ and F = n_features_in_.

The Where guard ensures that 0 ** 0 = 1 (sklearn convention) is satisfied even when the input contains zeros: whenever the exponent is 0, the base is replaced by 1 before the Pow operation.

Parameters:

  • g – the graph builder to add nodes to

  • sts – shapes defined by scikit-learn

  • outputs – desired output names

  • estimator – a fitted PolynomialFeatures

  • X – input tensor name

  • name – prefix for added node names

Returns:

output tensor name