yobx.sklearn.multioutput.regressor_chain

yobx.sklearn.multioutput.regressor_chain.sklearn_regressor_chain(g: GraphBuilderExtendedProtocol, sts: Dict, outputs: List[str], estimator: RegressorChain, X: str, name: str = 'regressor_chain') → str

Converts a sklearn.multioutput.RegressorChain into ONNX.

The converter iterates over the fitted sub-estimators in chain order. Each sub-estimator i receives an augmented feature matrix formed by concatenating the original input X with all previous predictions along axis 1. The prediction of every sub-estimator is reshaped from (N,) to (N, 1) before being appended to the running feature matrix.

After all sub-estimators have been applied the per-chain predictions are concatenated into a (N, n_targets) tensor in chain order. When estimator.order_ is not the identity permutation a final Gather node reorders the columns to match the original target order.

Graph structure (default order, 3 targets):

X ───────────────────── [est 0] ──► pred_0 (N,) ──Reshape(N,1)──► p0
Concat(X, p0) ────────── [est 1] ──► pred_1 (N,) ──Reshape(N,1)──► p1
Concat(X, p0, p1) ────── [est 2] ──► pred_2 (N,) ──Reshape(N,1)──► p2

            Concat(p0, p1, p2, axis=1) ──► chain_preds (N, n_targets)
                                                      │
              Gather(chain_preds, inv_order, axis=1) ──► predictions (N, n_targets)

Parameters:

• g – the graph builder to add nodes to

• sts – shapes and types defined by scikit-learn

• outputs – desired output tensor names

• estimator – a fitted RegressorChain

• X – name of the input tensor

• name – prefix used for names of nodes added by this converter

Returns:

name of the output tensor of shape (N, n_targets)