cluster.mean_shift

cluster.mean_shift(P, Q=None, kernel='gaussian', bandwidth='scott', batch=25, maxiter=300, atol=1.1920929e-06, blur=False, callback=None)

Mean shift algorithm for clustering or smoothing points via kernel density estimation.

This functions repeatedly mean shifts points P with respect to reference points Q, returning the shifted points. If Q is not supplied the shift points are shifted w.r.t P, or the shift points themselves if blur = True, until either maxiter iterations is reached the distance shifted is less than atol (elementwise), whichever comes first.

Parameters

Name	Type	Description	Default
P	np.ndarray	points to shift, given as a 2d np.array.	required
Q	Optional[np.ndarray]	reference points to shift P with respect to. Defaults to P itself.	None
kernel	Union[str, Callable]	kernel function to use. Only “gaussian” is supported for now.	'gaussian'
bandwidth	Union[str, float, Callable]	smoothing parameter for the kernel.	'scott'
batch	int	number of points to apply the shift to at once. See details.	25
maxiter	Optional[float]	maximum number of times to apply a the shift to a point. Can be None or np.inf to	300
atol	float	absolute tolerance with which to consider a point as converged between iterations.	1.1920929e-06
blur	bool	whether to apply the blurring mean shift operation. See details.	False
callback	Optional[Callable]	callable to execute after each iteration.	None

Returns

Type	Description
np.ndarray	ndarray of shift points with the same dimensions as P.