"""K-means clustering.""" # Authors: Gael Varoquaux # Thomas Rueckstiess # James Bergstra # Jan Schlueter # Nelle Varoquaux # Peter Prettenhofer # Olivier Grisel # Mathieu Blondel # Robert Layton # License: BSD 3 clause import warnings import numpy as np import scipy.sparse as sp from threadpoolctl import threadpool_limits from threadpoolctl import threadpool_info from ..base import BaseEstimator, ClusterMixin, TransformerMixin from ..metrics.pairwise import euclidean_distances from ..utils.extmath import row_norms, stable_cumsum from ..utils.sparsefuncs_fast import assign_rows_csr from ..utils.sparsefuncs import mean_variance_axis from ..utils.validation import _deprecate_positional_args from ..utils import check_array from ..utils import gen_batches from ..utils import check_random_state from ..utils import deprecated from ..utils.validation import check_is_fitted, _check_sample_weight from ..utils._openmp_helpers import _openmp_effective_n_threads from ..exceptions import ConvergenceWarning from ._k_means_fast import CHUNK_SIZE from ._k_means_fast import _inertia_dense from ._k_means_fast import _inertia_sparse from ._k_means_fast import _mini_batch_update_csr from ._k_means_lloyd import lloyd_iter_chunked_dense from ._k_means_lloyd import lloyd_iter_chunked_sparse from ._k_means_elkan import init_bounds_dense from ._k_means_elkan import init_bounds_sparse from ._k_means_elkan import elkan_iter_chunked_dense from ._k_means_elkan import elkan_iter_chunked_sparse ############################################################################### # Initialization heuristic def _kmeans_plusplus(X, n_clusters, x_squared_norms, random_state, n_local_trials=None): """Computational component for initialization of n_clusters by k-means++. Prior validation of data is assumed. Parameters ---------- X : {ndarray, sparse matrix} of shape (n_samples, n_features) The data to pick seeds for. n_clusters : int The number of seeds to choose. x_squared_norms : ndarray of shape (n_samples,) Squared Euclidean norm of each data point. random_state : RandomState instance The generator used to initialize the centers. See :term:`Glossary `. n_local_trials : int, default=None The number of seeding trials for each center (except the first), of which the one reducing inertia the most is greedily chosen. Set to None to make the number of trials depend logarithmically on the number of seeds (2+log(k)); this is the default. Returns ------- centers : ndarray of shape (n_clusters, n_features) The inital centers for k-means. indices : ndarray of shape (n_clusters,) The index location of the chosen centers in the data array X. For a given index and center, X[index] = center. """ n_samples, n_features = X.shape centers = np.empty((n_clusters, n_features), dtype=X.dtype) # Set the number of local seeding trials if none is given if n_local_trials is None: # This is what Arthur/Vassilvitskii tried, but did not report # specific results for other than mentioning in the conclusion # that it helped. n_local_trials = 2 + int(np.log(n_clusters)) # Pick first center randomly and track index of point center_id = random_state.randint(n_samples) indices = np.full(n_clusters, -1, dtype=int) if sp.issparse(X): centers[0] = X[center_id].toarray() else: centers[0] = X[center_id] indices[0] = center_id # Initialize list of closest distances and calculate current potential closest_dist_sq = euclidean_distances( centers[0, np.newaxis], X, Y_norm_squared=x_squared_norms, squared=True) current_pot = closest_dist_sq.sum() # Pick the remaining n_clusters-1 points for c in range(1, n_clusters): # Choose center candidates by sampling with probability proportional # to the squared distance to the closest existing center rand_vals = random_state.random_sample(n_local_trials) * current_pot candidate_ids = np.searchsorted(stable_cumsum(closest_dist_sq), rand_vals) # XXX: numerical imprecision can result in a candidate_id out of range np.clip(candidate_ids, None, closest_dist_sq.size - 1, out=candidate_ids) # Compute distances to center candidates distance_to_candidates = euclidean_distances( X[candidate_ids], X, Y_norm_squared=x_squared_norms, squared=True) # update closest distances squared and potential for each candidate np.minimum(closest_dist_sq, distance_to_candidates, out=distance_to_candidates) candidates_pot = distance_to_candidates.sum(axis=1) # Decide which candidate is the best best_candidate = np.argmin(candidates_pot) current_pot = candidates_pot[best_candidate] closest_dist_sq = distance_to_candidates[best_candidate] best_candidate = candidate_ids[best_candidate] # Permanently add best center candidate found in local tries if sp.issparse(X): centers[c] = X[best_candidate].toarray() else: centers[c] = X[best_candidate] indices[c] = best_candidate return centers, indices ############################################################################### # K-means batch estimation by EM (expectation maximization) def _tolerance(X, tol): """Return a tolerance which is independent of the dataset.""" if tol == 0: return 0 if sp.issparse(X): variances = mean_variance_axis(X, axis=0)[1] else: variances = np.var(X, axis=0) return np.mean(variances) * tol @_deprecate_positional_args def k_means(X, n_clusters, *, sample_weight=None, init='k-means++', precompute_distances='deprecated', n_init=10, max_iter=300, verbose=False, tol=1e-4, random_state=None, copy_x=True, n_jobs='deprecated', algorithm="auto", return_n_iter=False): """K-means clustering algorithm. Read more in the :ref:`User Guide `. Parameters ---------- X : {array-like, sparse matrix} of shape (n_samples, n_features) The observations to cluster. It must be noted that the data will be converted to C ordering, which will cause a memory copy if the given data is not C-contiguous. n_clusters : int The number of clusters to form as well as the number of centroids to generate. sample_weight : array-like of shape (n_samples,), default=None The weights for each observation in X. If None, all observations are assigned equal weight. init : {'k-means++', 'random'}, callable or array-like of shape \ (n_clusters, n_features), default='k-means++' Method for initialization: 'k-means++' : selects initial cluster centers for k-mean clustering in a smart way to speed up convergence. See section Notes in k_init for more details. 'random': choose `n_clusters` observations (rows) at random from data for the initial centroids. If an array is passed, it should be of shape (n_clusters, n_features) and gives the initial centers. If a callable is passed, it should take arguments X, n_clusters and a random state and return an initialization. precompute_distances : {'auto', True, False} Precompute distances (faster but takes more memory). 'auto' : do not precompute distances if n_samples * n_clusters > 12 million. This corresponds to about 100MB overhead per job using double precision. True : always precompute distances False : never precompute distances .. deprecated:: 0.23 'precompute_distances' was deprecated in version 0.23 and will be removed in 1.0 (renaming of 0.25). It has no effect. n_init : int, default=10 Number of time the k-means algorithm will be run with different centroid seeds. The final results will be the best output of n_init consecutive runs in terms of inertia. max_iter : int, default=300 Maximum number of iterations of the k-means algorithm to run. verbose : bool, default=False Verbosity mode. tol : float, default=1e-4 Relative tolerance with regards to Frobenius norm of the difference in the cluster centers of two consecutive iterations to declare convergence. random_state : int, RandomState instance or None, default=None Determines random number generation for centroid initialization. Use an int to make the randomness deterministic. See :term:`Glossary `. copy_x : bool, default=True When pre-computing distances it is more numerically accurate to center the data first. If copy_x is True (default), then the original data is not modified. If False, the original data is modified, and put back before the function returns, but small numerical differences may be introduced by subtracting and then adding the data mean. Note that if the original data is not C-contiguous, a copy will be made even if copy_x is False. If the original data is sparse, but not in CSR format, a copy will be made even if copy_x is False. n_jobs : int, default=None The number of OpenMP threads to use for the computation. Parallelism is sample-wise on the main cython loop which assigns each sample to its closest center. ``None`` or ``-1`` means using all processors. .. deprecated:: 0.23 ``n_jobs`` was deprecated in version 0.23 and will be removed in 1.0 (renaming of 0.25). algorithm : {"auto", "full", "elkan"}, default="auto" K-means algorithm to use. The classical EM-style algorithm is "full". The "elkan" variation is more efficient on data with well-defined clusters, by using the triangle inequality. However it's more memory intensive due to the allocation of an extra array of shape (n_samples, n_clusters). For now "auto" (kept for backward compatibility) chooses "elkan" but it might change in the future for a better heuristic. return_n_iter : bool, default=False Whether or not to return the number of iterations. Returns ------- centroid : ndarray of shape (n_clusters, n_features) Centroids found at the last iteration of k-means. label : ndarray of shape (n_samples,) label[i] is the code or index of the centroid the i'th observation is closest to. inertia : float The final value of the inertia criterion (sum of squared distances to the closest centroid for all observations in the training set). best_n_iter : int Number of iterations corresponding to the best results. Returned only if `return_n_iter` is set to True. """ est = KMeans( n_clusters=n_clusters, init=init, n_init=n_init, max_iter=max_iter, verbose=verbose, precompute_distances=precompute_distances, tol=tol, random_state=random_state, copy_x=copy_x, n_jobs=n_jobs, algorithm=algorithm ).fit(X, sample_weight=sample_weight) if return_n_iter: return est.cluster_centers_, est.labels_, est.inertia_, est.n_iter_ else: return est.cluster_centers_, est.labels_, est.inertia_ def _kmeans_single_elkan(X, sample_weight, centers_init, max_iter=300, verbose=False, x_squared_norms=None, tol=1e-4, n_threads=1): """A single run of k-means elkan, assumes preparation completed prior. Parameters ---------- X : {ndarray, sparse matrix} of shape (n_samples, n_features) The observations to cluster. If sparse matrix, must be in CSR format. sample_weight : array-like of shape (n_samples,) The weights for each observation in X. centers_init : ndarray of shape (n_clusters, n_features) The initial centers. max_iter : int, default=300 Maximum number of iterations of the k-means algorithm to run. verbose : bool, default=False Verbosity mode. x_squared_norms : array-like, default=None Precomputed x_squared_norms. tol : float, default=1e-4 Relative tolerance with regards to Frobenius norm of the difference in the cluster centers of two consecutive iterations to declare convergence. It's not advised to set `tol=0` since convergence might never be declared due to rounding errors. Use a very small number instead. n_threads : int, default=1 The number of OpenMP threads to use for the computation. Parallelism is sample-wise on the main cython loop which assigns each sample to its closest center. Returns ------- centroid : ndarray of shape (n_clusters, n_features) Centroids found at the last iteration of k-means. label : ndarray of shape (n_samples,) label[i] is the code or index of the centroid the i'th observation is closest to. inertia : float The final value of the inertia criterion (sum of squared distances to the closest centroid for all observations in the training set). n_iter : int Number of iterations run. """ n_samples = X.shape[0] n_clusters = centers_init.shape[0] # Buffers to avoid new allocations at each iteration. centers = centers_init centers_new = np.zeros_like(centers) weight_in_clusters = np.zeros(n_clusters, dtype=X.dtype) labels = np.full(n_samples, -1, dtype=np.int32) labels_old = labels.copy() center_half_distances = euclidean_distances(centers) / 2 distance_next_center = np.partition(np.asarray(center_half_distances), kth=1, axis=0)[1] upper_bounds = np.zeros(n_samples, dtype=X.dtype) lower_bounds = np.zeros((n_samples, n_clusters), dtype=X.dtype) center_shift = np.zeros(n_clusters, dtype=X.dtype) if sp.issparse(X): init_bounds = init_bounds_sparse elkan_iter = elkan_iter_chunked_sparse _inertia = _inertia_sparse else: init_bounds = init_bounds_dense elkan_iter = elkan_iter_chunked_dense _inertia = _inertia_dense init_bounds(X, centers, center_half_distances, labels, upper_bounds, lower_bounds) strict_convergence = False for i in range(max_iter): elkan_iter(X, sample_weight, centers, centers_new, weight_in_clusters, center_half_distances, distance_next_center, upper_bounds, lower_bounds, labels, center_shift, n_threads) # compute new pairwise distances between centers and closest other # center of each center for next iterations center_half_distances = euclidean_distances(centers_new) / 2 distance_next_center = np.partition( np.asarray(center_half_distances), kth=1, axis=0)[1] if verbose: inertia = _inertia(X, sample_weight, centers, labels) print(f"Iteration {i}, inertia {inertia}") centers, centers_new = centers_new, centers if np.array_equal(labels, labels_old): # First check the labels for strict convergence. if verbose: print(f"Converged at iteration {i}: strict convergence.") strict_convergence = True break else: # No strict convergence, check for tol based convergence. center_shift_tot = (center_shift**2).sum() if center_shift_tot <= tol: if verbose: print(f"Converged at iteration {i}: center shift " f"{center_shift_tot} within tolerance {tol}.") break labels_old[:] = labels if not strict_convergence: # rerun E-step so that predicted labels match cluster centers elkan_iter(X, sample_weight, centers, centers, weight_in_clusters, center_half_distances, distance_next_center, upper_bounds, lower_bounds, labels, center_shift, n_threads, update_centers=False) inertia = _inertia(X, sample_weight, centers, labels) return labels, inertia, centers, i + 1 def _kmeans_single_lloyd(X, sample_weight, centers_init, max_iter=300, verbose=False, x_squared_norms=None, tol=1e-4, n_threads=1): """A single run of k-means lloyd, assumes preparation completed prior. Parameters ---------- X : {ndarray, sparse matrix} of shape (n_samples, n_features) The observations to cluster. If sparse matrix, must be in CSR format. sample_weight : ndarray of shape (n_samples,) The weights for each observation in X. centers_init : ndarray of shape (n_clusters, n_features) The initial centers. max_iter : int, default=300 Maximum number of iterations of the k-means algorithm to run. verbose : bool, default=False Verbosity mode x_squared_norms : ndarray of shape (n_samples,), default=None Precomputed x_squared_norms. tol : float, default=1e-4 Relative tolerance with regards to Frobenius norm of the difference in the cluster centers of two consecutive iterations to declare convergence. It's not advised to set `tol=0` since convergence might never be declared due to rounding errors. Use a very small number instead. n_threads : int, default=1 The number of OpenMP threads to use for the computation. Parallelism is sample-wise on the main cython loop which assigns each sample to its closest center. Returns ------- centroid : ndarray of shape (n_clusters, n_features) Centroids found at the last iteration of k-means. label : ndarray of shape (n_samples,) label[i] is the code or index of the centroid the i'th observation is closest to. inertia : float The final value of the inertia criterion (sum of squared distances to the closest centroid for all observations in the training set). n_iter : int Number of iterations run. """ n_clusters = centers_init.shape[0] # Buffers to avoid new allocations at each iteration. centers = centers_init centers_new = np.zeros_like(centers) labels = np.full(X.shape[0], -1, dtype=np.int32) labels_old = labels.copy() weight_in_clusters = np.zeros(n_clusters, dtype=X.dtype) center_shift = np.zeros(n_clusters, dtype=X.dtype) if sp.issparse(X): lloyd_iter = lloyd_iter_chunked_sparse _inertia = _inertia_sparse else: lloyd_iter = lloyd_iter_chunked_dense _inertia = _inertia_dense strict_convergence = False # Threadpoolctl context to limit the number of threads in second level of # nested parallelism (i.e. BLAS) to avoid oversubsciption. with threadpool_limits(limits=1, user_api="blas"): for i in range(max_iter): lloyd_iter(X, sample_weight, x_squared_norms, centers, centers_new, weight_in_clusters, labels, center_shift, n_threads) if verbose: inertia = _inertia(X, sample_weight, centers, labels) print(f"Iteration {i}, inertia {inertia}.") centers, centers_new = centers_new, centers if np.array_equal(labels, labels_old): # First check the labels for strict convergence. if verbose: print(f"Converged at iteration {i}: strict convergence.") strict_convergence = True break else: # No strict convergence, check for tol based convergence. center_shift_tot = (center_shift**2).sum() if center_shift_tot <= tol: if verbose: print(f"Converged at iteration {i}: center shift " f"{center_shift_tot} within tolerance {tol}.") break labels_old[:] = labels if not strict_convergence: # rerun E-step so that predicted labels match cluster centers lloyd_iter(X, sample_weight, x_squared_norms, centers, centers, weight_in_clusters, labels, center_shift, n_threads, update_centers=False) inertia = _inertia(X, sample_weight, centers, labels) return labels, inertia, centers, i + 1 def _labels_inertia(X, sample_weight, x_squared_norms, centers, n_threads=None): """E step of the K-means EM algorithm. Compute the labels and the inertia of the given samples and centers. Parameters ---------- X : {ndarray, sparse matrix} of shape (n_samples, n_features) The input samples to assign to the labels. If sparse matrix, must be in CSR format. sample_weight : ndarray of shape (n_samples,) The weights for each observation in X. x_squared_norms : ndarray of shape (n_samples,) Precomputed squared euclidean norm of each data point, to speed up computations. centers : ndarray of shape (n_clusters, n_features) The cluster centers. n_threads : int, default=None The number of OpenMP threads to use for the computation. Parallelism is sample-wise on the main cython loop which assigns each sample to its closest center. Returns ------- labels : ndarray of shape (n_samples,) The resulting assignment. inertia : float Sum of squared distances of samples to their closest cluster center. """ n_samples = X.shape[0] n_clusters = centers.shape[0] n_threads = _openmp_effective_n_threads(n_threads) labels = np.full(n_samples, -1, dtype=np.int32) weight_in_clusters = np.zeros(n_clusters, dtype=centers.dtype) center_shift = np.zeros_like(weight_in_clusters) if sp.issparse(X): _labels = lloyd_iter_chunked_sparse _inertia = _inertia_sparse else: _labels = lloyd_iter_chunked_dense _inertia = _inertia_dense _labels(X, sample_weight, x_squared_norms, centers, centers, weight_in_clusters, labels, center_shift, n_threads, update_centers=False) inertia = _inertia(X, sample_weight, centers, labels) return labels, inertia class KMeans(TransformerMixin, ClusterMixin, BaseEstimator): """K-Means clustering. Read more in the :ref:`User Guide `. Parameters ---------- n_clusters : int, default=8 The number of clusters to form as well as the number of centroids to generate. init : {'k-means++', 'random'}, callable or array-like of shape \ (n_clusters, n_features), default='k-means++' Method for initialization: 'k-means++' : selects initial cluster centers for k-mean clustering in a smart way to speed up convergence. See section Notes in k_init for more details. 'random': choose `n_clusters` observations (rows) at random from data for the initial centroids. If an array is passed, it should be of shape (n_clusters, n_features) and gives the initial centers. If a callable is passed, it should take arguments X, n_clusters and a random state and return an initialization. n_init : int, default=10 Number of time the k-means algorithm will be run with different centroid seeds. The final results will be the best output of n_init consecutive runs in terms of inertia. max_iter : int, default=300 Maximum number of iterations of the k-means algorithm for a single run. tol : float, default=1e-4 Relative tolerance with regards to Frobenius norm of the difference in the cluster centers of two consecutive iterations to declare convergence. precompute_distances : {'auto', True, False}, default='auto' Precompute distances (faster but takes more memory). 'auto' : do not precompute distances if n_samples * n_clusters > 12 million. This corresponds to about 100MB overhead per job using double precision. True : always precompute distances. False : never precompute distances. .. deprecated:: 0.23 'precompute_distances' was deprecated in version 0.22 and will be removed in 1.0 (renaming of 0.25). It has no effect. verbose : int, default=0 Verbosity mode. random_state : int, RandomState instance or None, default=None Determines random number generation for centroid initialization. Use an int to make the randomness deterministic. See :term:`Glossary `. copy_x : bool, default=True When pre-computing distances it is more numerically accurate to center the data first. If copy_x is True (default), then the original data is not modified. If False, the original data is modified, and put back before the function returns, but small numerical differences may be introduced by subtracting and then adding the data mean. Note that if the original data is not C-contiguous, a copy will be made even if copy_x is False. If the original data is sparse, but not in CSR format, a copy will be made even if copy_x is False. n_jobs : int, default=None The number of OpenMP threads to use for the computation. Parallelism is sample-wise on the main cython loop which assigns each sample to its closest center. ``None`` or ``-1`` means using all processors. .. deprecated:: 0.23 ``n_jobs`` was deprecated in version 0.23 and will be removed in 1.0 (renaming of 0.25). algorithm : {"auto", "full", "elkan"}, default="auto" K-means algorithm to use. The classical EM-style algorithm is "full". The "elkan" variation is more efficient on data with well-defined clusters, by using the triangle inequality. However it's more memory intensive due to the allocation of an extra array of shape (n_samples, n_clusters). For now "auto" (kept for backward compatibiliy) chooses "elkan" but it might change in the future for a better heuristic. .. versionchanged:: 0.18 Added Elkan algorithm Attributes ---------- cluster_centers_ : ndarray of shape (n_clusters, n_features) Coordinates of cluster centers. If the algorithm stops before fully converging (see ``tol`` and ``max_iter``), these will not be consistent with ``labels_``. labels_ : ndarray of shape (n_samples,) Labels of each point inertia_ : float Sum of squared distances of samples to their closest cluster center. n_iter_ : int Number of iterations run. See Also -------- MiniBatchKMeans : Alternative online implementation that does incremental updates of the centers positions using mini-batches. For large scale learning (say n_samples > 10k) MiniBatchKMeans is probably much faster than the default batch implementation. Notes ----- The k-means problem is solved using either Lloyd's or Elkan's algorithm. The average complexity is given by O(k n T), where n is the number of samples and T is the number of iteration. The worst case complexity is given by O(n^(k+2/p)) with n = n_samples, p = n_features. (D. Arthur and S. Vassilvitskii, 'How slow is the k-means method?' SoCG2006) In practice, the k-means algorithm is very fast (one of the fastest clustering algorithms available), but it falls in local minima. That's why it can be useful to restart it several times. If the algorithm stops before fully converging (because of ``tol`` or ``max_iter``), ``labels_`` and ``cluster_centers_`` will not be consistent, i.e. the ``cluster_centers_`` will not be the means of the points in each cluster. Also, the estimator will reassign ``labels_`` after the last iteration to make ``labels_`` consistent with ``predict`` on the training set. Examples -------- >>> from sklearn.cluster import KMeans >>> import numpy as np >>> X = np.array([[1, 2], [1, 4], [1, 0], ... [10, 2], [10, 4], [10, 0]]) >>> kmeans = KMeans(n_clusters=2, random_state=0).fit(X) >>> kmeans.labels_ array([1, 1, 1, 0, 0, 0], dtype=int32) >>> kmeans.predict([[0, 0], [12, 3]]) array([1, 0], dtype=int32) >>> kmeans.cluster_centers_ array([[10., 2.], [ 1., 2.]]) """ @_deprecate_positional_args def __init__(self, n_clusters=8, *, init='k-means++', n_init=10, max_iter=300, tol=1e-4, precompute_distances='deprecated', verbose=0, random_state=None, copy_x=True, n_jobs='deprecated', algorithm='auto'): self.n_clusters = n_clusters self.init = init self.max_iter = max_iter self.tol = tol self.precompute_distances = precompute_distances self.n_init = n_init self.verbose = verbose self.random_state = random_state self.copy_x = copy_x self.n_jobs = n_jobs self.algorithm = algorithm def _check_params(self, X): # precompute_distances if self.precompute_distances != 'deprecated': warnings.warn("'precompute_distances' was deprecated in version " "0.23 and will be removed in 1.0 (renaming of 0.25)" ". It has no effect", FutureWarning) # n_jobs if self.n_jobs != 'deprecated': warnings.warn("'n_jobs' was deprecated in version 0.23 and will be" " removed in 1.0 (renaming of 0.25).", FutureWarning) self._n_threads = self.n_jobs else: self._n_threads = None self._n_threads = _openmp_effective_n_threads(self._n_threads) # n_init if self.n_init <= 0: raise ValueError( f"n_init should be > 0, got {self.n_init} instead.") self._n_init = self.n_init # max_iter if self.max_iter <= 0: raise ValueError( f"max_iter should be > 0, got {self.max_iter} instead.") # n_clusters if X.shape[0] < self.n_clusters: raise ValueError(f"n_samples={X.shape[0]} should be >= " f"n_clusters={self.n_clusters}.") # tol self._tol = _tolerance(X, self.tol) # algorithm if self.algorithm not in ("auto", "full", "elkan"): raise ValueError(f"Algorithm must be 'auto', 'full' or 'elkan', " f"got {self.algorithm} instead.") self._algorithm = self.algorithm if self._algorithm == "auto": self._algorithm = "full" if self.n_clusters == 1 else "elkan" if self._algorithm == "elkan" and self.n_clusters == 1: warnings.warn("algorithm='elkan' doesn't make sense for a single " "cluster. Using 'full' instead.", RuntimeWarning) self._algorithm = "full" # init if not (hasattr(self.init, '__array__') or callable(self.init) or (isinstance(self.init, str) and self.init in ["k-means++", "random"])): raise ValueError( f"init should be either 'k-means++', 'random', a ndarray or a " f"callable, got '{self.init}' instead.") if hasattr(self.init, '__array__') and self._n_init != 1: warnings.warn( f"Explicit initial center position passed: performing only" f" one init in {self.__class__.__name__} instead of " f"n_init={self._n_init}.", RuntimeWarning, stacklevel=2) self._n_init = 1 def _validate_center_shape(self, X, centers): """Check if centers is compatible with X and n_clusters.""" if centers.shape[0] != self.n_clusters: raise ValueError( f"The shape of the initial centers {centers.shape} does not " f"match the number of clusters {self.n_clusters}.") if centers.shape[1] != X.shape[1]: raise ValueError( f"The shape of the initial centers {centers.shape} does not " f"match the number of features of the data {X.shape[1]}.") def _check_test_data(self, X): X = self._validate_data(X, accept_sparse='csr', reset=False, dtype=[np.float64, np.float32], order='C', accept_large_sparse=False) return X def _check_mkl_vcomp(self, X, n_samples): """Warns when vcomp and mkl are both present""" # The BLAS call inside a prange in lloyd_iter_chunked_dense is known to # cause a small memory leak when there are less chunks than the number # of available threads. It only happens when the OpenMP library is # vcomp (microsoft OpenMP) and the BLAS library is MKL. see #18653 if sp.issparse(X): return active_threads = int(np.ceil(n_samples / CHUNK_SIZE)) if active_threads < self._n_threads: modules = threadpool_info() has_vcomp = "vcomp" in [module["prefix"] for module in modules] has_mkl = ("mkl", "intel") in [ (module["internal_api"], module.get("threading_layer", None)) for module in modules] if has_vcomp and has_mkl: if not hasattr(self, "batch_size"): # KMeans warnings.warn( f"KMeans is known to have a memory leak on Windows " f"with MKL, when there are less chunks than available " f"threads. You can avoid it by setting the environment" f" variable OMP_NUM_THREADS={active_threads}.") else: # MiniBatchKMeans warnings.warn( f"MiniBatchKMeans is known to have a memory leak on " f"Windows with MKL, when there are less chunks than " f"available threads. You can prevent it by setting " f"batch_size >= {self._n_threads * CHUNK_SIZE} or by " f"setting the environment variable " f"OMP_NUM_THREADS={active_threads}") def _init_centroids(self, X, x_squared_norms, init, random_state, init_size=None): """Compute the initial centroids. Parameters ---------- X : {ndarray, sparse matrix} of shape (n_samples, n_features) The input samples. x_squared_norms : ndarray of shape (n_samples,) Squared euclidean norm of each data point. Pass it if you have it at hands already to avoid it being recomputed here. init : {'k-means++', 'random'}, callable or ndarray of shape \ (n_clusters, n_features) Method for initialization. random_state : RandomState instance Determines random number generation for centroid initialization. See :term:`Glossary `. init_size : int, default=None Number of samples to randomly sample for speeding up the initialization (sometimes at the expense of accuracy). Returns ------- centers : ndarray of shape (n_clusters, n_features) """ n_samples = X.shape[0] n_clusters = self.n_clusters if init_size is not None and init_size < n_samples: init_indices = random_state.randint(0, n_samples, init_size) X = X[init_indices] x_squared_norms = x_squared_norms[init_indices] n_samples = X.shape[0] if isinstance(init, str) and init == 'k-means++': centers, _ = _kmeans_plusplus(X, n_clusters, random_state=random_state, x_squared_norms=x_squared_norms) elif isinstance(init, str) and init == 'random': seeds = random_state.permutation(n_samples)[:n_clusters] centers = X[seeds] elif hasattr(init, '__array__'): centers = init elif callable(init): centers = init(X, n_clusters, random_state=random_state) centers = check_array( centers, dtype=X.dtype, copy=False, order='C') self._validate_center_shape(X, centers) if sp.issparse(centers): centers = centers.toarray() return centers def fit(self, X, y=None, sample_weight=None): """Compute k-means clustering. Parameters ---------- X : {array-like, sparse matrix} of shape (n_samples, n_features) Training instances to cluster. It must be noted that the data will be converted to C ordering, which will cause a memory copy if the given data is not C-contiguous. If a sparse matrix is passed, a copy will be made if it's not in CSR format. y : Ignored Not used, present here for API consistency by convention. sample_weight : array-like of shape (n_samples,), default=None The weights for each observation in X. If None, all observations are assigned equal weight. .. versionadded:: 0.20 Returns ------- self Fitted estimator. """ X = self._validate_data(X, accept_sparse='csr', dtype=[np.float64, np.float32], order='C', copy=self.copy_x, accept_large_sparse=False) self._check_params(X) random_state = check_random_state(self.random_state) sample_weight = _check_sample_weight(sample_weight, X, dtype=X.dtype) # Validate init array init = self.init if hasattr(init, '__array__'): init = check_array(init, dtype=X.dtype, copy=True, order='C') self._validate_center_shape(X, init) # subtract of mean of x for more accurate distance computations if not sp.issparse(X): X_mean = X.mean(axis=0) # The copy was already done above X -= X_mean if hasattr(init, '__array__'): init -= X_mean # precompute squared norms of data points x_squared_norms = row_norms(X, squared=True) if self._algorithm == "full": kmeans_single = _kmeans_single_lloyd self._check_mkl_vcomp(X, X.shape[0]) else: kmeans_single = _kmeans_single_elkan best_inertia = None for i in range(self._n_init): # Initialize centers centers_init = self._init_centroids( X, x_squared_norms=x_squared_norms, init=init, random_state=random_state) if self.verbose: print("Initialization complete") # run a k-means once labels, inertia, centers, n_iter_ = kmeans_single( X, sample_weight, centers_init, max_iter=self.max_iter, verbose=self.verbose, tol=self._tol, x_squared_norms=x_squared_norms, n_threads=self._n_threads) # determine if these results are the best so far if best_inertia is None or inertia < best_inertia: best_labels = labels best_centers = centers best_inertia = inertia best_n_iter = n_iter_ if not sp.issparse(X): if not self.copy_x: X += X_mean best_centers += X_mean distinct_clusters = len(set(best_labels)) if distinct_clusters < self.n_clusters: warnings.warn( "Number of distinct clusters ({}) found smaller than " "n_clusters ({}). Possibly due to duplicate points " "in X.".format(distinct_clusters, self.n_clusters), ConvergenceWarning, stacklevel=2) self.cluster_centers_ = best_centers self.labels_ = best_labels self.inertia_ = best_inertia self.n_iter_ = best_n_iter return self def fit_predict(self, X, y=None, sample_weight=None): """Compute cluster centers and predict cluster index for each sample. Convenience method; equivalent to calling fit(X) followed by predict(X). Parameters ---------- X : {array-like, sparse matrix} of shape (n_samples, n_features) New data to transform. y : Ignored Not used, present here for API consistency by convention. sample_weight : array-like of shape (n_samples,), default=None The weights for each observation in X. If None, all observations are assigned equal weight. Returns ------- labels : ndarray of shape (n_samples,) Index of the cluster each sample belongs to. """ return self.fit(X, sample_weight=sample_weight).labels_ def fit_transform(self, X, y=None, sample_weight=None): """Compute clustering and transform X to cluster-distance space. Equivalent to fit(X).transform(X), but more efficiently implemented. Parameters ---------- X : {array-like, sparse matrix} of shape (n_samples, n_features) New data to transform. y : Ignored Not used, present here for API consistency by convention. sample_weight : array-like of shape (n_samples,), default=None The weights for each observation in X. If None, all observations are assigned equal weight. Returns ------- X_new : ndarray of shape (n_samples, n_clusters) X transformed in the new space. """ # Currently, this just skips a copy of the data if it is not in # np.array or CSR format already. # XXX This skips _check_test_data, which may change the dtype; # we should refactor the input validation. return self.fit(X, sample_weight=sample_weight)._transform(X) def transform(self, X): """Transform X to a cluster-distance space. In the new space, each dimension is the distance to the cluster centers. Note that even if X is sparse, the array returned by `transform` will typically be dense. Parameters ---------- X : {array-like, sparse matrix} of shape (n_samples, n_features) New data to transform. Returns ------- X_new : ndarray of shape (n_samples, n_clusters) X transformed in the new space. """ check_is_fitted(self) X = self._check_test_data(X) return self._transform(X) def _transform(self, X): """Guts of transform method; no input validation.""" return euclidean_distances(X, self.cluster_centers_) def predict(self, X, sample_weight=None): """Predict the closest cluster each sample in X belongs to. In the vector quantization literature, `cluster_centers_` is called the code book and each value returned by `predict` is the index of the closest code in the code book. Parameters ---------- X : {array-like, sparse matrix} of shape (n_samples, n_features) New data to predict. sample_weight : array-like of shape (n_samples,), default=None The weights for each observation in X. If None, all observations are assigned equal weight. Returns ------- labels : ndarray of shape (n_samples,) Index of the cluster each sample belongs to. """ check_is_fitted(self) X = self._check_test_data(X) x_squared_norms = row_norms(X, squared=True) sample_weight = _check_sample_weight(sample_weight, X, dtype=X.dtype) return _labels_inertia(X, sample_weight, x_squared_norms, self.cluster_centers_, self._n_threads)[0] def score(self, X, y=None, sample_weight=None): """Opposite of the value of X on the K-means objective. Parameters ---------- X : {array-like, sparse matrix} of shape (n_samples, n_features) New data. y : Ignored Not used, present here for API consistency by convention. sample_weight : array-like of shape (n_samples,), default=None The weights for each observation in X. If None, all observations are assigned equal weight. Returns ------- score : float Opposite of the value of X on the K-means objective. """ check_is_fitted(self) X = self._check_test_data(X) x_squared_norms = row_norms(X, squared=True) sample_weight = _check_sample_weight(sample_weight, X, dtype=X.dtype) return -_labels_inertia(X, sample_weight, x_squared_norms, self.cluster_centers_)[1] def _more_tags(self): return { '_xfail_checks': { 'check_sample_weights_invariance': 'zero sample_weight is not equivalent to removing samples', }, } def _mini_batch_step(X, sample_weight, x_squared_norms, centers, weight_sums, old_center_buffer, compute_squared_diff, distances, random_reassign=False, random_state=None, reassignment_ratio=.01, verbose=False): """Incremental update of the centers for the Minibatch K-Means algorithm. Parameters ---------- X : ndarray of shape (n_samples, n_features) The original data array. sample_weight : array-like of shape (n_samples,) The weights for each observation in X. x_squared_norms : ndarray of shape (n_samples,) Squared euclidean norm of each data point. centers : ndarray of shape (k, n_features) The cluster centers. This array is MODIFIED IN PLACE old_center_buffer : int Copy of old centers for monitoring convergence. compute_squared_diff : bool If set to False, the squared diff computation is skipped. distances : ndarray of shape (n_samples,), dtype=float, default=None If not None, should be a pre-allocated array that will be used to store the distances of each sample to its closest center. May not be None when random_reassign is True. random_reassign : bool, default=False If True, centers with very low counts are randomly reassigned to observations. random_state : int, RandomState instance or None, default=None Determines random number generation for centroid initialization and to pick new clusters amongst observations with uniform probability. Use an int to make the randomness deterministic. See :term:`Glossary `. reassignment_ratio : float, default=.01 Control the fraction of the maximum number of counts for a center to be reassigned. A higher value means that low count centers are more likely to be reassigned, which means that the model will take longer to converge, but should converge in a better clustering. verbose : bool, default=False Controls the verbosity. Returns ------- inertia : float Sum of squared distances of samples to their closest cluster center. squared_diff : ndarray of shape (n_clusters,) Squared distances between previous and updated cluster centers. """ # Perform label assignment to nearest centers nearest_center, inertia = _labels_inertia(X, sample_weight, x_squared_norms, centers) if random_reassign and reassignment_ratio > 0: random_state = check_random_state(random_state) # Reassign clusters that have very low weight to_reassign = weight_sums < reassignment_ratio * weight_sums.max() # pick at most .5 * batch_size samples as new centers if to_reassign.sum() > .5 * X.shape[0]: indices_dont_reassign = \ np.argsort(weight_sums)[int(.5 * X.shape[0]):] to_reassign[indices_dont_reassign] = False n_reassigns = to_reassign.sum() if n_reassigns: # Pick new clusters amongst observations with uniform probability new_centers = random_state.choice(X.shape[0], replace=False, size=n_reassigns) if verbose: print("[MiniBatchKMeans] Reassigning %i cluster centers." % n_reassigns) if sp.issparse(X) and not sp.issparse(centers): assign_rows_csr( X, new_centers.astype(np.intp, copy=False), np.where(to_reassign)[0].astype(np.intp, copy=False), centers) else: centers[to_reassign] = X[new_centers] # reset counts of reassigned centers, but don't reset them too small # to avoid instant reassignment. This is a pretty dirty hack as it # also modifies the learning rates. weight_sums[to_reassign] = np.min(weight_sums[~to_reassign]) # implementation for the sparse CSR representation completely written in # cython if sp.issparse(X): return inertia, _mini_batch_update_csr( X, sample_weight, x_squared_norms, centers, weight_sums, nearest_center, old_center_buffer, compute_squared_diff) # dense variant in mostly numpy (not as memory efficient though) k = centers.shape[0] squared_diff = 0.0 for center_idx in range(k): # find points from minibatch that are assigned to this center center_mask = nearest_center == center_idx wsum = sample_weight[center_mask].sum() if wsum > 0: if compute_squared_diff: old_center_buffer[:] = centers[center_idx] # inplace remove previous count scaling centers[center_idx] *= weight_sums[center_idx] # inplace sum with new points members of this cluster centers[center_idx] += \ np.sum(X[center_mask] * sample_weight[center_mask, np.newaxis], axis=0) # update the count statistics for this center weight_sums[center_idx] += wsum # inplace rescale to compute mean of all points (old and new) # Note: numpy >= 1.10 does not support '/=' for the following # expression for a mixture of int and float (see numpy issue #6464) centers[center_idx] = centers[center_idx] / weight_sums[center_idx] # update the squared diff if necessary if compute_squared_diff: diff = centers[center_idx].ravel() - old_center_buffer.ravel() squared_diff += np.dot(diff, diff) return inertia, squared_diff def _mini_batch_convergence(model, iteration_idx, n_iter, tol, n_samples, centers_squared_diff, batch_inertia, context, verbose=0): """Helper function to encapsulate the early stopping logic.""" # Normalize inertia to be able to compare values when # batch_size changes batch_inertia /= model.batch_size centers_squared_diff /= model.batch_size # Compute an Exponentially Weighted Average of the squared # diff to monitor the convergence while discarding # minibatch-local stochastic variability: # https://en.wikipedia.org/wiki/Moving_average ewa_diff = context.get('ewa_diff') ewa_inertia = context.get('ewa_inertia') if ewa_diff is None: ewa_diff = centers_squared_diff ewa_inertia = batch_inertia else: alpha = float(model.batch_size) * 2.0 / (n_samples + 1) alpha = 1.0 if alpha > 1.0 else alpha ewa_diff = ewa_diff * (1 - alpha) + centers_squared_diff * alpha ewa_inertia = ewa_inertia * (1 - alpha) + batch_inertia * alpha # Log progress to be able to monitor convergence if verbose: progress_msg = ( 'Minibatch iteration %d/%d:' ' mean batch inertia: %f, ewa inertia: %f ' % ( iteration_idx + 1, n_iter, batch_inertia, ewa_inertia)) print(progress_msg) # Early stopping based on absolute tolerance on squared change of # centers position (using EWA smoothing) if tol > 0.0 and ewa_diff <= tol: if verbose: print('Converged (small centers change) at iteration %d/%d' % (iteration_idx + 1, n_iter)) return True # Early stopping heuristic due to lack of improvement on smoothed inertia ewa_inertia_min = context.get('ewa_inertia_min') no_improvement = context.get('no_improvement', 0) if ewa_inertia_min is None or ewa_inertia < ewa_inertia_min: no_improvement = 0 ewa_inertia_min = ewa_inertia else: no_improvement += 1 if (model.max_no_improvement is not None and no_improvement >= model.max_no_improvement): if verbose: print('Converged (lack of improvement in inertia)' ' at iteration %d/%d' % (iteration_idx + 1, n_iter)) return True # update the convergence context to maintain state across successive calls: context['ewa_diff'] = ewa_diff context['ewa_inertia'] = ewa_inertia context['ewa_inertia_min'] = ewa_inertia_min context['no_improvement'] = no_improvement return False class MiniBatchKMeans(KMeans): """ Mini-Batch K-Means clustering. Read more in the :ref:`User Guide `. Parameters ---------- n_clusters : int, default=8 The number of clusters to form as well as the number of centroids to generate. init : {'k-means++', 'random'}, callable or array-like of shape \ (n_clusters, n_features), default='k-means++' Method for initialization: 'k-means++' : selects initial cluster centers for k-mean clustering in a smart way to speed up convergence. See section Notes in k_init for more details. 'random': choose `n_clusters` observations (rows) at random from data for the initial centroids. If an array is passed, it should be of shape (n_clusters, n_features) and gives the initial centers. If a callable is passed, it should take arguments X, n_clusters and a random state and return an initialization. max_iter : int, default=100 Maximum number of iterations over the complete dataset before stopping independently of any early stopping criterion heuristics. batch_size : int, default=100 Size of the mini batches. verbose : int, default=0 Verbosity mode. compute_labels : bool, default=True Compute label assignment and inertia for the complete dataset once the minibatch optimization has converged in fit. random_state : int, RandomState instance or None, default=None Determines random number generation for centroid initialization and random reassignment. Use an int to make the randomness deterministic. See :term:`Glossary `. tol : float, default=0.0 Control early stopping based on the relative center changes as measured by a smoothed, variance-normalized of the mean center squared position changes. This early stopping heuristics is closer to the one used for the batch variant of the algorithms but induces a slight computational and memory overhead over the inertia heuristic. To disable convergence detection based on normalized center change, set tol to 0.0 (default). max_no_improvement : int, default=10 Control early stopping based on the consecutive number of mini batches that does not yield an improvement on the smoothed inertia. To disable convergence detection based on inertia, set max_no_improvement to None. init_size : int, default=None Number of samples to randomly sample for speeding up the initialization (sometimes at the expense of accuracy): the only algorithm is initialized by running a batch KMeans on a random subset of the data. This needs to be larger than n_clusters. If `None`, `init_size= 3 * batch_size`. n_init : int, default=3 Number of random initializations that are tried. In contrast to KMeans, the algorithm is only run once, using the best of the ``n_init`` initializations as measured by inertia. reassignment_ratio : float, default=0.01 Control the fraction of the maximum number of counts for a center to be reassigned. A higher value means that low count centers are more easily reassigned, which means that the model will take longer to converge, but should converge in a better clustering. Attributes ---------- cluster_centers_ : ndarray of shape (n_clusters, n_features) Coordinates of cluster centers. labels_ : int Labels of each point (if compute_labels is set to True). inertia_ : float The value of the inertia criterion associated with the chosen partition (if compute_labels is set to True). The inertia is defined as the sum of square distances of samples to their nearest neighbor. n_iter_ : int Number of batches processed. counts_ : ndarray of shape (n_clusters,) Weigth sum of each cluster. .. deprecated:: 0.24 This attribute is deprecated in 0.24 and will be removed in 1.1 (renaming of 0.26). init_size_ : int The effective number of samples used for the initialization. .. deprecated:: 0.24 This attribute is deprecated in 0.24 and will be removed in 1.1 (renaming of 0.26). See Also -------- KMeans : The classic implementation of the clustering method based on the Lloyd's algorithm. It consumes the whole set of input data at each iteration. Notes ----- See https://www.eecs.tufts.edu/~dsculley/papers/fastkmeans.pdf Examples -------- >>> from sklearn.cluster import MiniBatchKMeans >>> import numpy as np >>> X = np.array([[1, 2], [1, 4], [1, 0], ... [4, 2], [4, 0], [4, 4], ... [4, 5], [0, 1], [2, 2], ... [3, 2], [5, 5], [1, -1]]) >>> # manually fit on batches >>> kmeans = MiniBatchKMeans(n_clusters=2, ... random_state=0, ... batch_size=6) >>> kmeans = kmeans.partial_fit(X[0:6,:]) >>> kmeans = kmeans.partial_fit(X[6:12,:]) >>> kmeans.cluster_centers_ array([[2. , 1. ], [3.5, 4.5]]) >>> kmeans.predict([[0, 0], [4, 4]]) array([0, 1], dtype=int32) >>> # fit on the whole data >>> kmeans = MiniBatchKMeans(n_clusters=2, ... random_state=0, ... batch_size=6, ... max_iter=10).fit(X) >>> kmeans.cluster_centers_ array([[3.95918367, 2.40816327], [1.12195122, 1.3902439 ]]) >>> kmeans.predict([[0, 0], [4, 4]]) array([1, 0], dtype=int32) """ @_deprecate_positional_args def __init__(self, n_clusters=8, *, init='k-means++', max_iter=100, batch_size=100, verbose=0, compute_labels=True, random_state=None, tol=0.0, max_no_improvement=10, init_size=None, n_init=3, reassignment_ratio=0.01): super().__init__( n_clusters=n_clusters, init=init, max_iter=max_iter, verbose=verbose, random_state=random_state, tol=tol, n_init=n_init) self.max_no_improvement = max_no_improvement self.batch_size = batch_size self.compute_labels = compute_labels self.init_size = init_size self.reassignment_ratio = reassignment_ratio @deprecated("The attribute 'counts_' is deprecated in 0.24" # type: ignore " and will be removed in 1.1 (renaming of 0.26).") @property def counts_(self): return self._counts @deprecated("The attribute 'init_size_' is deprecated in " # type: ignore "0.24 and will be removed in 1.1 (renaming of 0.26).") @property def init_size_(self): return self._init_size @deprecated("The attribute 'random_state_' is deprecated " # type: ignore "in 0.24 and will be removed in 1.1 (renaming of 0.26).") @property def random_state_(self): return getattr(self, "_random_state", None) def _check_params(self, X): super()._check_params(X) # max_no_improvement if self.max_no_improvement is not None and self.max_no_improvement < 0: raise ValueError( f"max_no_improvement should be >= 0, got " f"{self.max_no_improvement} instead.") # batch_size if self.batch_size <= 0: raise ValueError( f"batch_size should be > 0, got {self.batch_size} instead.") # init_size if self.init_size is not None and self.init_size <= 0: raise ValueError( f"init_size should be > 0, got {self.init_size} instead.") self._init_size = self.init_size if self._init_size is None: self._init_size = 3 * self.batch_size if self._init_size < self.n_clusters: self._init_size = 3 * self.n_clusters elif self._init_size < self.n_clusters: warnings.warn( f"init_size={self._init_size} should be larger than " f"n_clusters={self.n_clusters}. Setting it to " f"min(3*n_clusters, n_samples)", RuntimeWarning, stacklevel=2) self._init_size = 3 * self.n_clusters self._init_size = min(self._init_size, X.shape[0]) # reassignment_ratio if self.reassignment_ratio < 0: raise ValueError( f"reassignment_ratio should be >= 0, got " f"{self.reassignment_ratio} instead.") def fit(self, X, y=None, sample_weight=None): """Compute the centroids on X by chunking it into mini-batches. Parameters ---------- X : {array-like, sparse matrix} of shape (n_samples, n_features) Training instances to cluster. It must be noted that the data will be converted to C ordering, which will cause a memory copy if the given data is not C-contiguous. y : Ignored Not used, present here for API consistency by convention. sample_weight : array-like of shape (n_samples,), default=None The weights for each observation in X. If None, all observations are assigned equal weight (default: None). .. versionadded:: 0.20 Returns ------- self """ X = self._validate_data(X, accept_sparse='csr', dtype=[np.float64, np.float32], order='C', accept_large_sparse=False) self._check_params(X) random_state = check_random_state(self.random_state) sample_weight = _check_sample_weight(sample_weight, X, dtype=X.dtype) # Validate init array init = self.init if hasattr(init, '__array__'): init = check_array(init, dtype=X.dtype, copy=True, order='C') self._validate_center_shape(X, init) n_samples, n_features = X.shape x_squared_norms = row_norms(X, squared=True) if self.tol > 0.0: tol = _tolerance(X, self.tol) # using tol-based early stopping needs the allocation of a # dedicated before which can be expensive for high dim data: # hence we allocate it outside of the main loop old_center_buffer = np.zeros(n_features, dtype=X.dtype) else: tol = 0.0 # no need for the center buffer if tol-based early stopping is # disabled old_center_buffer = np.zeros(0, dtype=X.dtype) distances = np.zeros(self.batch_size, dtype=X.dtype) n_batches = int(np.ceil(float(n_samples) / self.batch_size)) n_iter = int(self.max_iter * n_batches) self._check_mkl_vcomp(X, self.batch_size) validation_indices = random_state.randint(0, n_samples, self._init_size) X_valid = X[validation_indices] sample_weight_valid = sample_weight[validation_indices] x_squared_norms_valid = x_squared_norms[validation_indices] # perform several inits with random sub-sets best_inertia = None for init_idx in range(self._n_init): if self.verbose: print("Init %d/%d with method: %s" % (init_idx + 1, self._n_init, init)) weight_sums = np.zeros(self.n_clusters, dtype=sample_weight.dtype) # TODO: once the `k_means` function works with sparse input we # should refactor the following init to use it instead. # Initialize the centers using only a fraction of the data as we # expect n_samples to be very large when using MiniBatchKMeans cluster_centers = self._init_centroids( X, x_squared_norms=x_squared_norms, init=init, random_state=random_state, init_size=self._init_size) # Compute the label assignment on the init dataset _mini_batch_step( X_valid, sample_weight_valid, x_squared_norms[validation_indices], cluster_centers, weight_sums, old_center_buffer, False, distances=None, verbose=self.verbose) # Keep only the best cluster centers across independent inits on # the common validation set _, inertia = _labels_inertia(X_valid, sample_weight_valid, x_squared_norms_valid, cluster_centers) if self.verbose: print("Inertia for init %d/%d: %f" % (init_idx + 1, self._n_init, inertia)) if best_inertia is None or inertia < best_inertia: self.cluster_centers_ = cluster_centers self._counts = weight_sums best_inertia = inertia # Empty context to be used inplace by the convergence check routine convergence_context = {} # Perform the iterative optimization until the final convergence # criterion for iteration_idx in range(n_iter): # Sample a minibatch from the full dataset minibatch_indices = random_state.randint( 0, n_samples, self.batch_size) # Perform the actual update step on the minibatch data batch_inertia, centers_squared_diff = _mini_batch_step( X[minibatch_indices], sample_weight[minibatch_indices], x_squared_norms[minibatch_indices], self.cluster_centers_, self._counts, old_center_buffer, tol > 0.0, distances=distances, # Here we randomly choose whether to perform # random reassignment: the choice is done as a function # of the iteration index, and the minimum number of # counts, in order to force this reassignment to happen # every once in a while random_reassign=((iteration_idx + 1) % (10 + int(self._counts.min())) == 0), random_state=random_state, reassignment_ratio=self.reassignment_ratio, verbose=self.verbose) # Monitor convergence and do early stopping if necessary if _mini_batch_convergence( self, iteration_idx, n_iter, tol, n_samples, centers_squared_diff, batch_inertia, convergence_context, verbose=self.verbose): break self.n_iter_ = iteration_idx + 1 if self.compute_labels: self.labels_, self.inertia_ = \ self._labels_inertia_minibatch(X, sample_weight) return self def _labels_inertia_minibatch(self, X, sample_weight): """Compute labels and inertia using mini batches. This is slightly slower than doing everything at once but prevents memory errors / segfaults. Parameters ---------- X : array-like of shape (n_samples, n_features) Input data. sample_weight : array-like of shape (n_samples,) The weights for each observation in X. Returns ------- labels : ndarray of shape (n_samples,) Cluster labels for each point. inertia : float Sum of squared distances of points to nearest cluster. """ if self.verbose: print('Computing label assignment and total inertia') sample_weight = _check_sample_weight(sample_weight, X, dtype=X.dtype) x_squared_norms = row_norms(X, squared=True) slices = gen_batches(X.shape[0], self.batch_size) results = [_labels_inertia(X[s], sample_weight[s], x_squared_norms[s], self.cluster_centers_) for s in slices] labels, inertia = zip(*results) return np.hstack(labels), np.sum(inertia) def partial_fit(self, X, y=None, sample_weight=None): """Update k means estimate on a single mini-batch X. Parameters ---------- X : array-like of shape (n_samples, n_features) Coordinates of the data points to cluster. It must be noted that X will be copied if it is not C-contiguous. y : Ignored Not used, present here for API consistency by convention. sample_weight : array-like of shape (n_samples,), default=None The weights for each observation in X. If None, all observations are assigned equal weight (default: None). Returns ------- self """ is_first_call_to_partial_fit = not hasattr(self, 'cluster_centers_') X = self._validate_data(X, accept_sparse='csr', dtype=[np.float64, np.float32], order='C', accept_large_sparse=False, reset=is_first_call_to_partial_fit) self._random_state = getattr(self, "_random_state", check_random_state(self.random_state)) sample_weight = _check_sample_weight(sample_weight, X, dtype=X.dtype) x_squared_norms = row_norms(X, squared=True) if is_first_call_to_partial_fit: # this is the first call to partial_fit on this object self._check_params(X) # Validate init array init = self.init if hasattr(init, '__array__'): init = check_array(init, dtype=X.dtype, copy=True, order='C') self._validate_center_shape(X, init) self._check_mkl_vcomp(X, X.shape[0]) # initialize the cluster centers self.cluster_centers_ = self._init_centroids( X, x_squared_norms=x_squared_norms, init=init, random_state=self._random_state, init_size=self._init_size) self._counts = np.zeros(self.n_clusters, dtype=sample_weight.dtype) random_reassign = False distances = None else: # The lower the minimum count is, the more we do random # reassignment, however, we don't want to do random # reassignment too often, to allow for building up counts random_reassign = self._random_state.randint( 10 * (1 + self._counts.min())) == 0 distances = np.zeros(X.shape[0], dtype=X.dtype) _mini_batch_step(X, sample_weight, x_squared_norms, self.cluster_centers_, self._counts, np.zeros(0, dtype=X.dtype), 0, random_reassign=random_reassign, distances=distances, random_state=self._random_state, reassignment_ratio=self.reassignment_ratio, verbose=self.verbose) if self.compute_labels: self.labels_, self.inertia_ = _labels_inertia( X, sample_weight, x_squared_norms, self.cluster_centers_) return self def predict(self, X, sample_weight=None): """Predict the closest cluster each sample in X belongs to. In the vector quantization literature, `cluster_centers_` is called the code book and each value returned by `predict` is the index of the closest code in the code book. Parameters ---------- X : {array-like, sparse matrix} of shape (n_samples, n_features) New data to predict. sample_weight : array-like of shape (n_samples,), default=None The weights for each observation in X. If None, all observations are assigned equal weight (default: None). Returns ------- labels : ndarray of shape (n_samples,) Index of the cluster each sample belongs to. """ check_is_fitted(self) X = self._check_test_data(X) return self._labels_inertia_minibatch(X, sample_weight)[0] def _more_tags(self): return { '_xfail_checks': { 'check_sample_weights_invariance': 'zero sample_weight is not equivalent to removing samples', } } def kmeans_plusplus(X, n_clusters, *, x_squared_norms=None, random_state=None, n_local_trials=None): """Init n_clusters seeds according to k-means++ .. versionadded:: 0.24 Parameters ---------- X : {array-like, sparse matrix} of shape (n_samples, n_features) The data to pick seeds from. n_clusters : int The number of centroids to initialize x_squared_norms : array-like of shape (n_samples,), default=None Squared Euclidean norm of each data point. random_state : int or RandomState instance, default=None Determines random number generation for centroid initialization. Pass an int for reproducible output across multiple function calls. See :term:`Glossary `. n_local_trials : int, default=None The number of seeding trials for each center (except the first), of which the one reducing inertia the most is greedily chosen. Set to None to make the number of trials depend logarithmically on the number of seeds (2+log(k)). Returns ------- centers : ndarray of shape (n_clusters, n_features) The inital centers for k-means. indices : ndarray of shape (n_clusters,) The index location of the chosen centers in the data array X. For a given index and center, X[index] = center. Notes ----- Selects initial cluster centers for k-mean clustering in a smart way to speed up convergence. see: Arthur, D. and Vassilvitskii, S. "k-means++: the advantages of careful seeding". ACM-SIAM symposium on Discrete algorithms. 2007 Examples -------- >>> from sklearn.cluster import kmeans_plusplus >>> import numpy as np >>> X = np.array([[1, 2], [1, 4], [1, 0], ... [10, 2], [10, 4], [10, 0]]) >>> centers, indices = kmeans_plusplus(X, n_clusters=2, random_state=0) >>> centers array([[10, 4], [ 1, 0]]) >>> indices array([4, 2]) """ # Check data check_array(X, accept_sparse='csr', dtype=[np.float64, np.float32]) if X.shape[0] < n_clusters: raise ValueError(f"n_samples={X.shape[0]} should be >= " f"n_clusters={n_clusters}.") # Check parameters if x_squared_norms is None: x_squared_norms = row_norms(X, squared=True) else: x_squared_norms = check_array(x_squared_norms, dtype=X.dtype, ensure_2d=False) if x_squared_norms.shape[0] != X.shape[0]: raise ValueError( f"The length of x_squared_norms {x_squared_norms.shape[0]} should " f"be equal to the length of n_samples {X.shape[0]}.") if n_local_trials is not None and n_local_trials < 1: raise ValueError( f"n_local_trials is set to {n_local_trials} but should be an " f"integer value greater than zero.") random_state = check_random_state(random_state) # Call private k-means++ centers, indices = _kmeans_plusplus(X, n_clusters, x_squared_norms, random_state, n_local_trials) return centers, indices