SelfTuningAffinity

class torchdr.SelfTuningAffinity(K: int = 7, normalization_dim: int | Tuple[int] = (0, 1), metric: str = 'sqeuclidean', zero_diag: bool = True, device: str | None = None, backend: str | None = None, verbose: bool = False)

Bases: LogAffinity

Self-tuning affinity introduced in [Zelnik-Manor and Perona, 2004].

The affinity has a sample-wise bandwidth \(\mathbf{\sigma} \in \mathbb{R}^n\).

\[\exp \left( - \frac{C_{ij}}{\sigma_i \sigma_j} \right)\]

In the above, \(\mathbf{C}\) is the pairwise distance matrix and \(\sigma_i\) is the distance from the K’th nearest neighbor of data point \(\mathbf{x}_i\).

Parameters:

• K (int, optional) – K-th neirest neighbor .

• normalization_dim (int or Tuple[int], optional) – Dimension along which to normalize the affinity matrix.

• metric (str, optional) – Metric to use for pairwise distances computation.

• zero_diag (bool, optional) – Whether to set the diagonal of the affinity matrix to zero.

• device (str, optional) – Device to use for computations.

• backend ({"keops", "faiss", None}, optional) – Which backend to use for handling sparsity and memory efficiency. Default is None.

• verbose (bool, optional) – Verbosity. Default is False.