mojolearn.randomized_svd#

mojolearn.randomized_svd(M, n_components, *, n_oversamples=10, n_iter='auto', power_iteration_normalizer='auto', transpose='auto', flip_sign=True, random_state=None, numeric_mode=None)#

Truncated SVD by randomized range finding.

Follows scikit-learn’s utils.extmath.randomized_svd (Halko et al.). Every power iteration re-orthonormalizes with a QR step whatever power_iteration_normalizer says (all choices span the same subspace), and the random test matrix comes from mojolearn’s own generator, so results differ slightly from scikit-learn’s for the same seed.

Parameters:

  • M (array-like or scipy.sparse matrix of shape (n_samples, n_features))

  • n_components (int) – Number of singular values and vectors; between 1 and min(n_samples, n_features).

  • n_oversamples (int, default 10) – Extra random vectors used to sample the range.

  • n_iter (int or “auto”, default “auto”) – Power iterations. "auto" uses 7 when n_components < 0.1 * min(M.shape), else 4.

  • power_iteration_normalizer ({“auto”, “QR”, “LU”, “none”}, default “auto”) – Accepted for scikit-learn compatibility; see above.

  • transpose (bool or “auto”, default “auto”) – Work on M.T; "auto" does so when M has more columns than rows.

  • flip_sign (bool, default True) – Make the largest-magnitude entry of each column of U positive, as scikit-learn’s svd_flip.

  • random_state (int or None, default None) – Seed of the random test matrix; None means 0.

  • numeric_mode ({“identical”, “fast”} or None, default None) – Numeric mode for this call. None uses the package default ("identical" unless changed). In identical mode, mojolearn returns the same bits on Apple, NVIDIA and AMD GPUs. "fast" is available on Apple GPUs only.

Returns:

  • U (ndarray of shape (n_samples, n_components), float32)

  • s (ndarray of shape (n_components,), float32) – Singular values, descending.

  • Vt (ndarray of shape (n_components, n_features), float32)

Raises:

ValueError – If n_components is out of range, power_iteration_normalizer is unknown, or M is numerically zero.

Examples

>>> import numpy as np, mojolearn
>>> X = np.random.default_rng(0).random((1000, 50), dtype=np.float32)
>>> U, s, Vt = mojolearn.randomized_svd(X, 5, random_state=0)
On this page