:jmLdZddlmZddlZddlmZddlmZddl m Z ddl m Z ddl mZdd lmZdd lmZmZdd lmZdd lmZmZmZdd lmZmZddlmZddlm Z ddl!m"Z"GddeZ#Gdde#Z$Gdde#Z%Gdde#Z&Gdde#Z'GddeZ(y)zClustering pipelines)EnumN) rearrange) AudioFile)PLDA)oracle_segmentation) permutate) cluster_vbx) SlidingWindowSlidingWindowFeature)Pipeline) CategoricalIntegerUniform)fclusterlinkage)linear_sum_assignment)cdist)KMeansc eZdZ ddedeffd Z ddededzdedzdedzfd Z dd ejd e dzd e d e ejejejffdZ dejd ejfdZ dd ejdejdejdejdef dZ dd ejd e dzdedzdedzdedzd ejf dZxZS)BaseClusteringmetricconstrained_assignmentc>t|||_||_y)N)super__init__rrselfrr __class__s x/home/mcse/projects/srt_converter/srt-converter-venv/lib/python3.12/site-packages/pyannote/audio/pipelines/clustering.pyrzBaseClustering.__init__-s  &<#Nnum_embeddings num_clusters min_clusters max_clustersc|xs|xsd}tdt||}|xs|xs|}tdt||}||kDrtd|dd|dd||k(r|}|||fS)NzQmin_clusters must be smaller than (or equal to) max_clusters (here: min_clusters=gz and max_clusters=z).)maxmin ValueError)rr!r"r#r$s rset_num_clusterszBaseClustering.set_num_clusters6s$8|8q 1c.,?@ #E|E~ 1c.,?@ , &''3A&66HVWHXXZ\  < ''L\<77r embeddings segmentationsmin_active_ratioreturncx|jj\}}}tj|jdddk(}tj|j|zd}|||zk\}tjtj |d} tj || z\} } || | f| | fS)aFilter embeddings before clustering Embeddings that are removed: * NaN embeddings * embeddings speaking less than `min_active_ratio` times the chunk duration Parameters ---------- embeddings : (num_chunks, num_speakers, dimension) array Sequence of embeddings. segmentations : (num_chunks, num_frames, num_speakers) array Binary segmentations. min_active_ratio : float, optional Minimum active ratio for a speaker to be considered active during clustering. Returns ------- filtered_embeddings : (num_embeddings, dimension) array chunk_idx : (num_embeddings, ) array speaker_idx : (num_embeddings, ) array Taxiskeepdimsr&r3)datashapenpsumanyisnanwhere) rr,r-r._ num_framessingle_active_masknum_clean_framesactivevalid chunk_idx speaker_idxs rfilter_embeddingsz BaseClustering.filter_embeddingsMs:)--33:q!ff]%7%7a$OSTT66-"4"47I"IPQR"%5 %BB,155"$&5.!9 ;)[019kIIr soft_clusterscJtj|tj|}|j\}}}dtj||ftj z}t |D]0\}}t|d\}} t|| D] \} } | ||| f<2|S)N)nandtypeT)maximize) r8 nan_to_numnanminr7onesint8 enumeraterzip) rrF num_chunks num_speakersr" hard_clustersccostspeakersclusterssks rconstrained_argmaxz!BaseClustering.constrained_argmaxs m=9QR 1>1D1D. L,RWWj,%?rwwOO  / (GAt!6td!K HhHh/ (1&' ad# ( ( r train_chunk_idxtrain_speaker_idxtrain_clusters constrainedc tj|dz}|j\}}} |||f} tjt |D cgc]} tj | || k(d!c} } t tt |d| |jd||} d| z }|r|j|}ntj|d}||| fScc} w) aAssign embeddings to the closest centroid Cluster centroids are computed as the average of the train embeddings previously assigned to them. Parameters ---------- embeddings : (num_chunks, num_speakers, dimension)-shaped array Complete set of embeddings. train_chunk_idx : (num_embeddings,)-shaped array train_speaker_idx : (num_embeddings,)-shaped array Indices of subset of embeddings used for "training". train_clusters : (num_embedding,)-shaped array Clusters of the above subset constrained : bool, optional Use constrained_argmax, instead of (default) argmax. Returns ------- soft_clusters : (num_chunks, num_speakers, num_clusters)-shaped array hard_clusters : (num_chunks, num_speakers)-shaped array centroids : (num_clusters, dimension)-shaped array Clusters centroids r&rr5c s d -> (c s) dr(c s) k -> c s krVrZr1) r8r(r7vstackrangemeanrrrr\argmax)rr,r]r^r_r`r"rSrT dimensiontrain_embeddingsr[ centroids e2k_distancerFrUs rassign_embeddingsz BaseClustering.assign_embeddingssFvvn-1 .8.>.>+ L)%o7H&HIII|, (1)<=AF  ! *&89{{    L(   33MBMIIm! 8Y Q Qv   s&N1$N>$O1N;rurwryrzr{__doc__rr}__annotations__r|rr8rr~rrrrs@rrr$s""'$&',//!%/*$(#'#' VJJVDjVDj V Dj Vr rc eZdZUdZdZeed< d deffd Z d de jde dzd e dzd e dzfd Z xZ S)KMeansClusteringzKMeans clustering Parameters ---------- metric : {"cosine", "euclidean"}, optional Distance metric to use. Defaults to "cosine". Hyper-parameters ---------------- None TrrcL|dvrtd|dt| |y)N)rvrzUnsupported metric: z". Must be 'cosine' or 'euclidean'.rc)r*rr)rrrs rrzKMeansClustering.__init__s: 0 0&vh.PQ  'r Nr,r#r$r"c| td|j\}}||kr%tj|tjS|j dk(rEtj dd5|tjj|dd z}dddt|d d d j|S#1swY'xYw)aYPerform KMeans clustering Parameters ---------- embeddings : (num_embeddings, dimension) array Embeddings num_clusters : int, optional Expected number of clusters. Returns ------- clusters : (num_embeddings, ) array 0-indexed cluster indices. Nz `num_clusters` must be provided.rJrvrrrTr2r*F n_clustersn_init random_statecopy_x) r*r7r8rint32rrrrr fit_predict)rr,r#r$r"r!r=s rrrzKMeansClustering.clusters,  ?@ @&,, L (99^288< < ;;( "Hh? QbiinnZb4nPP  Q#ABu +j ! "  Q Qs -&B99C)rvrwrrs@rrrsm "&$% ( ($(#'#' &"JJ&"Dj&"Dj &" Dj &"r rc eZdZUdZeed< ddededeffd Z dde jd e dzd e dzd e dzd e dzd e jf dZ xZS) VBxClusteringFrpldarrct|||||_tdd|_tdd|_tdd|_y)Nrg?g?g{Gz?g.@)rrrrrFaFb)rrrrrs rrzVBxClustering.__init__+sP #9    c*$$$%r Nr,r-r"r#r$r/c |j}|j||\}} } |jddkrl|j\} } } tj| | ftj } tj | | df} tj|dd}| | |fS|tjj|ddz }t|dd }t||jd dz }tj|d \} }|j|}t|||jj |j"|j$d\}}|j\} } }|dd|dkDf}|j&|j)d|z|j+ddj&z }|j\}} ||kr|}n||kDr|}|rl||k7rgd}t-|dddj/|}tj0t3|Dcgc]}tj|||k(d!c}}t5t7t5|d||j8d| | }d|z } |rF| j;dz }|| |j<j+ddk(<|j?| } ntj@| d} | j)| | } | | |fScc}w)Nrprr1rJr&Tr2rrrrrrr)rrmaxItersgHz>r)r4Frrrr5rbrcrdre?)!rrEr7r8rqrPrOrhrrrrrrrr phirrTreshaper9rrrfrgrrrr)r6r\ri)rr,r-r"r#r$rsrrkr=rSrTrUrFrltrain_embeddings_normedr ahc_clustersfeaqsprjWauto_num_clusterskmeans_clustersr[rmconsts rrtzVBxClustering.__call__<s"&!.>+ L) adlOCC*222yAAAEE!VZED[D]D]] )1 | +'L  -'L L,==&+ "$'5k12  #<0GG,_-ABKI! *&89{{    L(  "!%%'",ErTwindowrr=oracle_num_framesr"rUrFrV segmentationoracle permutationjirkr]r^r_r[rls rrtzOracleClustering.__call__s @0=/A/A/G/G, J --24O(< #$-A-F-F-L-L* l%**1.RJ@Q0R.R+RS 388 3Z!233 3 RWWj,%?rwwOO *lL!IJ )2  3 4*  - %A% f$-VBJJ-?#N A a!+. -19&' ad#), aAg&  -  -   -5 5  " " ' #     '8I'IJII|, (1)<=AF  mY66  s<$F,rx) ryrzr{rrr}rr8rr rr rtr rrrsu!%$%)-59!%'+ O7JJ%O7,d2O7$ O7 $ O7 O7r rceZdZeZeZeZeZy) ClusteringN)ryrzr{rrrrrr rrrs5'!M'r r))renumrnumpyr8einopsrpyannote.audio.core.iorpyannote.audio.core.pldarpyannote.audio.pipelines.utilsr pyannote.audio.utils.permutationrpyannote.audio.utils.vbxr pyannote.corer r pyannote.pipeliner pyannote.pipeline.parameterr rrscipy.cluster.hierarchyrrscipy.optimizerscipy.spatial.distancersklearn.clusterrrrrrrrrr rrs0,)>60=&EE50("u7Xu7p|n|~@"~@"Fw7Nw7tT7~T7n((r