o 9wiQ"@s^ddlZddlmZmZmZmZddlmZmZddl Z ddl Z ddl mZddlZddlmmZddlmZeddeedefd d Zd d Zd dZej  ddejdejdeededeejeeeff ddZej  dde jde jdeededee jeeeff ddZdefdede deejejgejfde j!fddZ"dS)N)CallableListOptionalTuple)singledispatchpartial)linear_sum_assignment)SlidingWindowFeatureF cost_func return_costcCst)aFind cost-minimizing permutation Parameters ---------- y1 : np.ndarray or torch.Tensor (batch_size, num_samples, num_classes_1) y2 : np.ndarray or torch.Tensor (num_samples, num_classes_2) or (batch_size, num_samples, num_classes_2) cost_func : callable Takes two (num_samples, num_classes) sequences and returns (num_classes, ) pairwise cost. Defaults to computing mean squared error. return_cost : bool, optional Whether to return cost matrix. Defaults to False. Returns ------- permutated_y2 : np.ndarray or torch.Tensor (batch_size, num_samples, num_classes_1) permutations : list of tuple List of permutations so that permutation[i] == j indicates that jth speaker of y2 should be mapped to ith speaker of y1. permutation[i] == None when none of y2 speakers is mapped to ith speaker of y1. cost : np.ndarray or torch.Tensor, optional (batch_size, num_classes_1, num_classes_2) ) TypeError)y1y2r r r]/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/pyannote/audio/utils/permutation.py permutate&srcKstjtj||ddddS)zCompute class-wise mean-squared error Parameters ---------- Y, y : (num_frames, num_classes) torch.tensor Returns ------- mse : (num_classes, ) torch.tensor Mean-squared error none) reductionraxis)torchmeanFmse_lossYykwargsrrr mse_cost_funcDs rcKstjt||ddS)zCompute class-wise mean absolute difference error Parameters ---------- Y, y: (num_frames, num_classes) torch.tensor Returns ------- mae : (num_classes, ) torch.tensor Mean absolute difference error rr)rrabsrrrr mae_cost_funcSs r r rreturnc s|j\}}}t|jdkr||dd}t|jdkr!d}t||j\}} ||ks/|| krBdt|jdt|jd}t|durHtg} g} |rPg} tj|j|j|j d} t t ||D]|\} \t t fd d t|D}Wdn1swY|krt|d d d |fd t|d }n|}dg|}t t|D]\}}||kr|||<dd|f| | dd|f<q| t||r| |qb|r| | t | fS| | fS)NzAIncorrect shape: should be (batch_size, num_frames, num_classes).zShape mismatch: z vs. .)devicedtypec s2g|]}dd||dfdqS)Nr#)expand).0ir num_classes_2y1_y2_rr s$z#permutate_torch..rconstantr()shapelenr) ValueErrortuplerrzerosr&r' enumeratezipno_gradstackrangerpadmaxrcpuappend)r rr r batch_size num_samples num_classes_1msg batch_size_ num_samples_ permutations permutated_y2costsbcost padded_cost permutationk1k2rr,rpermutate_torchbs\      rOcCsPtt|t|||d}|r|\}}}|||fS|\}}||fS)Nr r )rr from_numpynumpy)r rr r outputrGrFrHrrrpermutate_numpys  rTg? segmentationsonsetc Cst||d}|j}|jj\}}}t|j|jd}d|f}t } t |D]\} \} } t t d| |dt || |ddD]} | | krJqCt| | ||j|j}|dkrp| }| |d}|| d||f}n| d||}|| |df}t|tj||dd\}\}\}t |D]D\}}t|dd|f|k}t|dd|f|k}|r| | |f|r| | |f|r|r| j| |f| |f|||fdqqCq(| S) aaBuild permutation graph Parameters ---------- segmentations : (num_chunks, num_frames, local_num_speakers)-shaped SlidingWindowFeature Raw output of segmentation model. onset : float, optionan Threshold above which a speaker is considered active. Defaults to 0.5 cost_func : callable Cost function used to find the optimal bijective mapping between speaker activations of two overlapping chunks. Expects two (num_frames, num_classes) torch.tensor as input and returns cost as a (num_classes, ) torch.tensor. Defaults to mae_cost_func. Returns ------- permutation_graph : nx.Graph Nodes are (chunk_idx, speaker_idx) tuples. An edge between two nodes indicate that those are likely to be the same speaker (the lower the value of "cost" attribute, the more likely). )rVr(r"rNTrP)rJ)rsliding_windowdatar2mathfloordurationstepnxGraphr7r;r=minroundrnpnewaxisanyadd_nodeadd_edge)rUrVr chunks num_chunks num_frames_ max_lookahead lookaheadpermutation_graphCchunk segmentationcshiftthis_segmentationsthat_segmentationsrLrJthisthatthis_is_activethat_is_activerrrbuild_permutation_graphsL  . (rx)NF)#rYtypingrrrr functoolsrrnetworkxr]rRrascipy.optimizerrtorch.nn.functionalnn functionalr pyannote.corer boolrrr registerTensorintrOndarrayrTfloatr^rxrrrrsh   E