yobx.sklearn.discriminant_analysis.qda#

yobx.sklearn.discriminant_analysis.qda.sklearn_quadratic_discriminant_analysis(g: GraphBuilderExtendedProtocol, sts: Dict, outputs: List[str], estimator: QuadraticDiscriminantAnalysis, X: str, name: str = 'qda') → Tuple[str, str][source]#

Converts a sklearn.discriminant_analysis.QuadraticDiscriminantAnalysis into ONNX.

For each class k the log-likelihood (decision function) is computed from the per-class SVD stored in rotations_ and scalings_ (eigenvalues of the class covariance matrix):

W_k     = R_k * S_k^(-0.5)                  # scaled rotation (F, r_k) – constant
offset_k = mean_k @ W_k                      # (r_k,) – constant
const_k  = -0.5 * sum(log S_k) + log(prior_k)  # scalar – constant

z_k     = X @ W_k - offset_k                # (N, r_k)
norm2_k = ReduceSum(z_k * z_k, axis=1)      # (N,)
dec_k   = -0.5 * norm2_k + const_k          # (N,)

When all classes share the same SVD rank r, all classes are processed in a single batched MatMul following the same pattern as the sklearn_gaussian_mixture() 'full' covariance path:

W_2d   = hstack(W_0, …, W_{C-1})            # (F, C*r) – constant
b_flat = hstack(offset_0, …, offset_{C-1})  # (1, C*r) – constant
consts = [-0.5*logdet_k + logprior_k …]     # (C,) – constant

XW     = MatMul(X, W_2d)                    # (N, C*r)
diff   = XW - b_flat                         # (N, C*r)
diff3d = Reshape(diff, [-1, C, r])           # (N, C, r)
norm2  = ReduceSum(diff3d * diff3d, axis=2)  # (N, C)
dec    = -0.5 * norm2 + consts               # (N, C)

When classes have different SVD ranks (degenerate covariances), the same computation is performed per class and the resulting (N, 1) columns are concatenated.

In both cases probabilities and labels are obtained as:

proba = Softmax(dec, axis=1)                 # (N, C) – output
label = Gather(classes_, ArgMax(proba, 1))   # (N,)   – output
Parameters:

  • g – the graph builder to add nodes to

  • sts – shapes defined by scikit-learn

  • estimator – a fitted QuadraticDiscriminantAnalysis

  • outputs – desired names (label, probabilities)

  • X – input tensor name

  • name – prefix names for the added nodes

Returns:

tuple (label_result_name, proba_result_name)