o pi!@sdddlZddlmZddlmZmZ            ddd Zdd d ZddZ ddZ dS)N)eigh) logsumexpsoftmax? -C6?FcCsZ |jd}t|turt||}|dur/tjj||jdt|fd}||jddd}|jdt|krB|jd|jdksDJdtj|dddd |t dtj }t |}||}g}d}t |D]}|dksy| dusy| durd d|||jddd j |} ||| |j |} ||| j d | | d||}d }t ||}t||d d}tj|dd}t|||dddf}tj|dd}||}||d tt | | | dd}||g|dkr||dd|kr||dddkrtdnqk|||f| r*| | fSdS)a Inputs: X - T x D array, where columns are D dimensional feature vectors (e.g. x-vectors) for T frames Phi - D array with across-class covariance matrix diagonal. The model assumes zero mean, diagonal across-class and identity within-class covariance matrix. Fa - Scale sufficient statistics Fb - Speaker regularization coefficient Fb controls the final number of speakers pi - If integer value, it sets the maximum number of speakers that can be found in the utterance. If vector, it is the initialization for speaker priors (see Outputs: pi) gamma - An initialization for the matrix of responsibilities (see Outputs: gamma) maxIters - The maximum number of VB iterations epsilon - Stop iterating, if the obj. fun. improvement is less than epsilon alphaQInit - Dirichlet concentraion parameter for initializing gamma ref - T dim. integer vector with per frame reference speaker IDs (0:maxSpeakers) plot - If set to True, plot per-frame marginal speaker posteriors 'gamma' return_model- Return also speaker model parameter alpha, invL - If provided, these are speaker model parameters used in the first iteration Outputs: gamma - S x T matrix of responsibilities (marginal posteriors) attributing each frame to one of S possible speakers (S is defined by input parameter pi) pi - S dimensional column vector of ML learned speaker priors. This allows us to estimate the number of speaker in the utterance as the probabilities of the redundant speaker converge to zero. Li - Values of auxiliary function (and DER and frame cross-entropy between gamma and reference, if 'ref' is provided) over iterations. alpha, invL - Speaker model parameters returned only if return_model=True Reference: Landini F., Profant J., Diez M., Burget L.: Bayesian HMM clustering of x-vector sequences (VBx) in speaker diarization: theory, implementation and analysis on standard tasks Nr)sizeT)keepdimsg)axisr rg?g:0yE>r z3WARNING: Value of auxiliary function has decreased!)shapetypeintnponesrandomgammalensumlogpisqrtrangeTdotrexpappendprint)XPhiFaFbrrmaxItersepsilon alphaQInitrefplot return_modelalphainvLDGVrhoLiELBOiilog_p_epslpilog_p_xlog_pX_rrL/home/ubuntu/.local/lib/python3.10/site-packages/pyannote/audio/utils/vbx.pyVBxsR6  *&  $ ,  r<@c Cstt||df}d|tt||tf<|dkr |nt||dd}t|||||j d||dd\}} } } } || fS)zahc_init (T x N_clusters)rrrrT)r%r&rrr'r,) rzerosrmaxrastyperrr<r) ahc_initfear$r%r&r'init_smoothingqinitrr_rrr; cluster_vbxs rGcCsbt|jdkr|tj|St|jdkr(|tjj|dddddtjfStdt|j)zL2 normalization of vector array. Args: vec_or_matrix (np.array): one vector or array of vectors Returns: np.array: normalized vector or array of normalized vectors rr )r ordNz8Wrong number of dimensions, 1 or 2 is supported, not %i.)rrrlinalgnormnewaxis ValueError) vec_or_matrixrrr;l2_norms "rNc st|}|d|d|dt|}|d|d|d}tjj}tjj|}t||\}}|ddd}|jdddfd d } jd ffd d } | | |fS) a' Loads the transformation pipeline for x-vectors into the PLDA space for VBx. Args: transform_npz (str | Path): Path to the x-vector transformation file. plda_npz (str | Path): Path to the PLDA file. Returns: xvec_tf (function): Transformation function to preprocess x-vectors (centering, whitening, LDA). plda_tf (function): Transformation function to map x-vectors into the PLDA latent space. plda_psi (np.ndarray): Eigenvalues of the between-class covariance in the PLDA space (reordered). mean1mean2ldamutrpsiNr cs@tjdtjtjdt|jjS)Nrr)rrrrNrr)x)rQrOrPrr;s*zvbx_setup..rcs |jddd|fS)N)rr)x0lda_dim)plda_muplda_trrr;rVs)rloadrIinvrrrr) transform_npzplda_npzrUpplda_psiWBacvarwccnxvec_tfplda_tfr)rQrOrPrYrZr; vbx_setups   rg) rrrNrrrNFFNN)r=r>) numpyr scipy.linalgr scipy.specialrrr<rGrNrgrrrr;s&  q