ó ‡ˆ\c@sddZddlmZmZddlZddlmZddlm Z m Z ddl m Z m Z ddlmZmZd d lmZmZd d lmZd d lmZd d lmZmZmZd dlmZmZddgZd„Z ddd„Z!d„Z"dej#eeeƒfd„ƒYZ$de$fd„ƒYZ%de$fd„ƒYZ&dS(sOSpectral biclustering algorithms. Authors : Kemal Eren License: BSD 3 clause iÿÿÿÿ(tABCMetatabstractmethodN(tnorm(t dia_matrixtissparse(teigshtsvdsi(tKMeanstMiniBatchKMeansi(t BaseEstimatortBiclusterMixin(tsix(tcheck_random_state(tmake_nonnegativetrandomized_svdtsafe_sparse_dot(tassert_all_finitet check_arraytSpectralCoclusteringtSpectralBiclusteringcCs@t|ƒ}tjdtj|jddƒƒƒjƒ}tjdtj|jddƒƒƒjƒ}tjtj|ƒd|ƒ}tjtj|ƒd|ƒ}t|ƒr|j \}}t |dgfd||fƒ}t |dgfd||fƒ}|||}n!|dd…tj f||}|||fS(sNormalize ``X`` by scaling rows and columns independently. Returns the normalized matrix and the row and column scaling factors. gð?taxisiitshapeN( R tnptasarraytsqrttsumtsqueezetwheretisnanRRRtnewaxis(tXtrow_diagtcol_diagtn_rowstn_colstrtctan((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyt_scale_normalizes .. !!!iègñh㈵øä>cCsŸt|ƒ}|}d}x€t|ƒD]r}t|ƒ\}}}t|ƒret|j|jƒ}nt||ƒ}|}|dk r%||kr%Pq%q%W|S(s’Normalize rows and columns of ``X`` simultaneously so that all rows sum to one constant and all columns sum to a different constant. N(R tNonetrangeR&RRtdata(Rtmax_iterttoltX_scaledtdistt_tX_new((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyt_bistochastic_normalize4s  cCst|ddƒ}t|ƒr-tdƒ‚ntj|ƒ}|jddƒdd…tjf}|jddƒ}|jƒ}||||S(s>Normalize ``X`` according to Kluger's log-interactions scheme.t min_valueis[Cannot compute log of a sparse matrix, because log(x) diverges to -infinity as x goes to 0.RNi(R Rt ValueErrorRtlogtmeanR(RtLtrow_avgtcol_avgtavg((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyt_log_normalizeKs % t BaseSpectralc Bs\eZdZeddd eddd d d„ƒZd„Zd d„Zd„Z d „Z RS( s%Base class for spectral biclustering.it randomizeds k-means++i c CsL||_||_||_||_||_||_||_||_dS(N(t n_clusterst svd_methodt n_svd_vecst mini_batchtinittn_inittn_jobst random_state( tselfR<R=R>R?R@RARBRC((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyt__init__]s       cCs7d}|j|kr3tdj|j|ƒƒ‚ndS(NR;tarpacks9Unknown SVD method: '{0}'. svd_method must be one of {1}.(R;RF(R=R2tformat(RDtlegal_svd_methods((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyt_check_parametersjs  cCs6t|dddtjƒ}|jƒ|j|ƒ|S(s™Creates a biclustering for X. Parameters ---------- X : array-like, shape (n_samples, n_features) y : Ignored t accept_sparsetcsrtdtype(RRtfloat64RIt_fit(RDRty((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pytfitqs   c CsÓ|jdkr[i}|jd k r4|j|dR'RRCRRtanyRRtTR tuniformRRR( RDtarrayt n_componentst n_discardtkwargstuR.tvttARCRTtv((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyt_svd€s2 $! '   c CsŒ|jr3t|d|jd|jd|jƒ}n0t|d|jd|jd|jd|jƒ}|j|ƒ|j}|j }||fS(NR@RARCRB( R?RR@RARCRRBRPtcluster_centers_tlabels_(RDR)R<tmodeltcentroidtlabels((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyt_k_means¦s        N( t__name__t __module__t__doc__RR'tFalseRERIRPR`Rf(((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyR:Ys    &c Bs8eZdZdddeddddd„Zd„ZRS(sˆSpectral Co-Clustering algorithm (Dhillon, 2001). Clusters rows and columns of an array `X` to solve the relaxed normalized cut of the bipartite graph created from `X` as follows: the edge between row vertex `i` and column vertex `j` has weight `X[i, j]`. The resulting bicluster structure is block-diagonal, since each row and each column belongs to exactly one bicluster. Supports sparse matrices, as long as they are nonnegative. Read more in the :ref:`User Guide `. Parameters ---------- n_clusters : integer, optional, default: 3 The number of biclusters to find. svd_method : string, optional, default: 'randomized' Selects the algorithm for finding singular vectors. May be 'randomized' or 'arpack'. If 'randomized', use :func:`sklearn.utils.extmath.randomized_svd`, which may be faster for large matrices. If 'arpack', use :func:`scipy.sparse.linalg.svds`, which is more accurate, but possibly slower in some cases. n_svd_vecs : int, optional, default: None Number of vectors to use in calculating the SVD. Corresponds to `ncv` when `svd_method=arpack` and `n_oversamples` when `svd_method` is 'randomized`. mini_batch : bool, optional, default: False Whether to use mini-batch k-means, which is faster but may get different results. init : {'k-means++', 'random' or an ndarray} Method for initialization of k-means algorithm; defaults to 'k-means++'. n_init : int, optional, default: 10 Number of random initializations that are tried with the k-means algorithm. If mini-batch k-means is used, the best initialization is chosen and the algorithm runs once. Otherwise, the algorithm is run for each initialization and the best solution chosen. n_jobs : int or None, optional (default=None) The number of jobs to use for the computation. This works by breaking down the pairwise matrix into n_jobs even slices and computing them in parallel. ``None`` means 1 unless in a :obj:`joblib.parallel_backend` context. ``-1`` means using all processors. See :term:`Glossary ` for more details. random_state : int, RandomState instance or None (default) Used for randomizing the singular value decomposition and the k-means initialization. Use an int to make the randomness deterministic. See :term:`Glossary `. Attributes ---------- rows_ : array-like, shape (n_row_clusters, n_rows) Results of the clustering. `rows[i, r]` is True if cluster `i` contains row `r`. Available only after calling ``fit``. columns_ : array-like, shape (n_column_clusters, n_columns) Results of the clustering, like `rows`. row_labels_ : array-like, shape (n_rows,) The bicluster label of each row. column_labels_ : array-like, shape (n_cols,) The bicluster label of each column. Examples -------- >>> from sklearn.cluster import SpectralCoclustering >>> import numpy as np >>> X = np.array([[1, 1], [2, 1], [1, 0], ... [4, 7], [3, 5], [3, 6]]) >>> clustering = SpectralCoclustering(n_clusters=2, random_state=0).fit(X) >>> clustering.row_labels_ array([0, 1, 1, 0, 0, 0], dtype=int32) >>> clustering.column_labels_ array([0, 0], dtype=int32) >>> clustering # doctest: +NORMALIZE_WHITESPACE SpectralCoclustering(init='k-means++', mini_batch=False, n_clusters=2, n_init=10, n_jobs=None, n_svd_vecs=None, random_state=0, svd_method='randomized') References ---------- * Dhillon, Inderjit S, 2001. `Co-clustering documents and words using bipartite spectral graph partitioning `__. iR;s k-means++i c Cs/tt|ƒj||||||||ƒdS(N(tsuperRRE( RDR<R=R>R?R@RARBRC((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyREsc CsOt|ƒ\}}}dttjtj|jƒƒƒ}|j||ddƒ\}}tj|dd…tjf||dd…tjf|fƒ}|j ||jƒ\} } |j d} | | |_ | | |_ tjgt |jƒD]} |j | k^qóƒ|_tjgt |jƒD]} |j | k^q*ƒ|_dS(NiRZi(R&tintRtceiltlog2R<R`tvstackRRfRt row_labels_tcolumn_labels_R(trows_tcolumns_( RDRtnormalized_dataRR tn_svR\R_tzR.ReR!R$((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyRN(s% #    . N(RgRhRiR'RjRERN(((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyR¶s e  c Bs\eZdZdddddd eddd d d„ Zd„Zd „Zd „Zd „Z RS( s_Spectral biclustering (Kluger, 2003). Partitions rows and columns under the assumption that the data has an underlying checkerboard structure. For instance, if there are two row partitions and three column partitions, each row will belong to three biclusters, and each column will belong to two biclusters. The outer product of the corresponding row and column label vectors gives this checkerboard structure. Read more in the :ref:`User Guide `. Parameters ---------- n_clusters : integer or tuple (n_row_clusters, n_column_clusters) The number of row and column clusters in the checkerboard structure. method : string, optional, default: 'bistochastic' Method of normalizing and converting singular vectors into biclusters. May be one of 'scale', 'bistochastic', or 'log'. The authors recommend using 'log'. If the data is sparse, however, log normalization will not work, which is why the default is 'bistochastic'. CAUTION: if `method='log'`, the data must not be sparse. n_components : integer, optional, default: 6 Number of singular vectors to check. n_best : integer, optional, default: 3 Number of best singular vectors to which to project the data for clustering. svd_method : string, optional, default: 'randomized' Selects the algorithm for finding singular vectors. May be 'randomized' or 'arpack'. If 'randomized', uses `sklearn.utils.extmath.randomized_svd`, which may be faster for large matrices. If 'arpack', uses `scipy.sparse.linalg.svds`, which is more accurate, but possibly slower in some cases. n_svd_vecs : int, optional, default: None Number of vectors to use in calculating the SVD. Corresponds to `ncv` when `svd_method=arpack` and `n_oversamples` when `svd_method` is 'randomized`. mini_batch : bool, optional, default: False Whether to use mini-batch k-means, which is faster but may get different results. init : {'k-means++', 'random' or an ndarray} Method for initialization of k-means algorithm; defaults to 'k-means++'. n_init : int, optional, default: 10 Number of random initializations that are tried with the k-means algorithm. If mini-batch k-means is used, the best initialization is chosen and the algorithm runs once. Otherwise, the algorithm is run for each initialization and the best solution chosen. n_jobs : int or None, optional (default=None) The number of jobs to use for the computation. This works by breaking down the pairwise matrix into n_jobs even slices and computing them in parallel. ``None`` means 1 unless in a :obj:`joblib.parallel_backend` context. ``-1`` means using all processors. See :term:`Glossary ` for more details. random_state : int, RandomState instance or None (default) Used for randomizing the singular value decomposition and the k-means initialization. Use an int to make the randomness deterministic. See :term:`Glossary `. Attributes ---------- rows_ : array-like, shape (n_row_clusters, n_rows) Results of the clustering. `rows[i, r]` is True if cluster `i` contains row `r`. Available only after calling ``fit``. columns_ : array-like, shape (n_column_clusters, n_columns) Results of the clustering, like `rows`. row_labels_ : array-like, shape (n_rows,) Row partition labels. column_labels_ : array-like, shape (n_cols,) Column partition labels. Examples -------- >>> from sklearn.cluster import SpectralBiclustering >>> import numpy as np >>> X = np.array([[1, 1], [2, 1], [1, 0], ... [4, 7], [3, 5], [3, 6]]) >>> clustering = SpectralBiclustering(n_clusters=2, random_state=0).fit(X) >>> clustering.row_labels_ array([1, 1, 1, 0, 0, 0], dtype=int32) >>> clustering.column_labels_ array([0, 1], dtype=int32) >>> clustering # doctest: +NORMALIZE_WHITESPACE SpectralBiclustering(init='k-means++', method='bistochastic', mini_batch=False, n_best=3, n_clusters=2, n_components=6, n_init=10, n_jobs=None, n_svd_vecs=None, random_state=0, svd_method='randomized') References ---------- * Kluger, Yuval, et. al., 2003. `Spectral biclustering of microarray data: coclustering genes and conditions `__. it bistochasticiR;s k-means++i c CsJtt|ƒj|||||| | | ƒ||_||_||_dS(N(RkRREtmethodRYtn_best( RDR<RxRYRyR=R>R?R@RARBRC((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyRE¯s  cCsCtt|ƒjƒd }|j|krFtdj|j|ƒƒ‚nyt|jƒWn^tk r·y'|j\}}t|ƒt|ƒWq¸ttfk r³tdƒ‚q¸XnX|j dkrâtdj|j ƒƒ‚n|j dkr tdj|j ƒƒ‚n|j |j kr?td j|j |j ƒƒ‚ndS( NRwtscaleR3s1Unknown method: '{0}'. method must be one of {1}.s˜Incorrect parameter n_clusters has value: {}. It should either be a single integer or an iterable with two integers: (n_row_clusters, n_column_clusters)isBParameter n_components must be greater than 0, but its value is {}s<Parameter n_best must be greater than 0, but its value is {}s7n_best cannot be larger than n_components, but {} > {}(RwRzR3( RkRRIRxR2RGRlR<t TypeErrorRYRy(RDt legal_methodsR#R$((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyRI¿s.      cCs÷|j}|jdkr1t|ƒ}|d7}nO|jdkrbt|ƒ\}}}|d7}n|jdkr€t|ƒ}n|jdkr•dnd}|j|||ƒ\}}|j}|j} y|j\} } Wntk rû|j} } nX|j ||j | ƒ} |j | |j | ƒ} |j || j| ƒ|_ |j |j| j| ƒ|_ tjgt| ƒD](}t| ƒD]}|j |k^qˆqxƒ|_tjgt| ƒD](}t| ƒD]}|j |k^qÏq¿ƒ|_dS(NRwiRzR3i(RYRxR0R&R9R`RVR<R{t_fit_best_piecewiseRyt_project_and_clusterRpRqRRoR(RrRs(RDRRuRtR.RZR\R_tutR]tn_row_clusterstn_col_clusterstbest_uttbest_vttlabel((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyRNÜs>            . csg‡‡fd†}tj|ddd|ƒ}tjtddd||ƒ}|tj|ƒ| }|S(sØFind the ``n_best`` vectors that are best approximated by piecewise constant vectors. The piecewise vectors are found by k-means; the best is chosen according to Euclidean distance. cs2ˆj|jddƒˆƒ\}}||jƒS(Niÿÿÿÿi(Rftreshapetravel(R_RdRe(R<RD(s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pytmake_piecewise s$Ritarr(Rtapply_along_axisRtargsort(RDtvectorsRyR<R‡tpiecewise_vectorstdiststresult((R<RDs8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyR}s  cCs+t||ƒ}|j||ƒ\}}|S(s7Project ``data`` to ``vectors`` and cluster the result.(RRf(RDR)R‹R<t projectedR.Re((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyR~sN( RgRhRiR'RjRERIRNR}R~(((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyR;ss    ' ('RitabcRRtnumpyRt scipy.linalgRt scipy.sparseRRtscipy.sparse.linalgRRtRRtbaseR R t externalsR tutilsR t utils.extmathR RRtutils.validationRRt__all__R&R0R9twith_metaclassR:RR(((s8lib/python2.7/site-packages/sklearn/cluster/bicluster.pyts(    \…