yobx.sklearn.isotonic.isotonic_regression#

yobx.sklearn.isotonic.isotonic_regression.sklearn_isotonic_regression(g: GraphBuilderExtendedProtocol, sts: Dict, outputs: List[str], estimator: IsotonicRegression, X: str, name: str = 'isotonic_regression') → str[source]#

Converts a sklearn.isotonic.IsotonicRegression into ONNX.

The prediction follows sklearn.isotonic.IsotonicRegression.predict(), which uses piecewise-linear interpolation via numpy.interp():

predict(X) = np.interp(X.ravel(), X_thresholds_, y_thresholds_)

Values below X_min_ are clamped to y_thresholds_[0] and values above X_max_ are clamped to y_thresholds_[-1].

ONNX graph structure (K breakpoints, K ≥ 2):

X (N, 1) or (N,)
  │
  Reshape(-1) ──► x_flat (N,)
  │
  Clip(X_min_, X_max_) ──► x_clipped (N,)
  │
  Unsqueeze(-1) ──► x_exp (N, 1)
  │
xp (K,) ──── GreaterOrEqual ──► cmp (N, K)  [bool]
  │               │
  │          Cast(INT64) ──► cmp_int (N, K)
  │               │
  │          ReduceSum(axis=1) ──► seg_count (N,)
  │               │
  │         Sub(1) ──► seg_lo_raw (N,)
  │               │
  │         Clip(0, K-2) ──► seg_lo (N,)  int64
  │               │
  │         Add(1) ──► seg_hi (N,)  int64
  │
Gather(xp, seg_lo) ──► xp_lo (N,)
Gather(xp, seg_hi) ──► xp_hi (N,)
Gather(fp, seg_lo) ──► fp_lo (N,)
Gather(fp, seg_hi) ──► fp_hi (N,)
  │
dx  = xp_hi - xp_lo
t   = (x_clipped - xp_lo) / dx
out = fp_lo + t * (fp_hi - fp_lo)  ──► predictions (N,)

When all training samples collapse to a single breakpoint (K = 1), the graph simply broadcasts the constant y_thresholds_[0] to all rows.

Parameters:

  • g – the graph builder to add nodes to

  • sts – shapes defined by scikit-learn

  • outputs – desired output names

  • estimator – a fitted IsotonicRegression

  • X – input tensor name (shape (N,) or (N, 1))

  • name – prefix for added node names

Returns:

output tensor name (shape (N,))