yobx.sklearn.kernel_approximation.nystroem

yobx.sklearn.kernel_approximation.nystroem.sklearn_nystroem(g: GraphBuilderExtendedProtocol, sts: Dict, outputs: List[str], estimator: Nystroem, X: str, name: str = 'nystroem') → str

Converts a sklearn.kernel_approximation.Nystroem into ONNX.

The out-of-sample transform replicates sklearn.kernel_approximation.Nystroem.transform():

1. Compute the pairwise kernel matrix between X (shape (N, F)) and the stored components components_ (shape (C, F)):

   • linear — K = X @ components_.T

   • rbf — K[i,j] = exp(-gamma * ||X[i] - components_[j]||^2)

   • poly — K = (gamma * X @ components_.T + coef0) ^ degree

   • sigmoid — K = tanh(gamma * X @ components_.T + coef0)

   • cosine — K[i,j] = (X[i]/||X[i]||) * (components_[j]/||components_[j]||)

   Callable and precomputed kernels are not supported.

2. Project using the normalization matrix stored in normalization_:

   result = K @ normalization_.T     # (N, C)
   

Parameters:

• g – the graph builder to add nodes to

• sts – shapes defined by scikit-learn

• estimator – a fitted Nystroem

• outputs – desired output names (transformed inputs)

• X – input tensor name

• name – prefix name for the added nodes

Returns:

output tensor name

Raises:

NotImplementedError – for callable kernels, kernel='precomputed', or unsupported kernel names