o 9wi@spddlmZddlmZmZddlmZmZddlZ ddl Z ddl m Z ddl m mZGddde jZdS))cached_property) combinations permutations)DictTupleNcseZdZdZdedeffdd ZedefddZdej fd d Z dej fd d Z ddej de dej fddZ ddej de dej fddZdej dej fddZdeedfdeedffddZedeeedfeedfffddZZS)PowersetzPowerset to multilabel conversion, and back. Parameters ---------- num_classes : int Number of regular classes. max_set_size : int Maximum number of classes in each set. num_classes max_set_sizecsBt||_||_|jd|dd|jd|dddS)NmappingF) persistent cardinality)super__init__rr register_buffer build_mappingbuild_cardinality)selfrr  __class__Z/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/pyannote/audio/utils/powerset.pyr0s zPowerset.__init__returncs&ttfddtdjdDS)Nc3s |] }tjj|VqdS)N)scipyspecialbinomr).0irrr >s  z0Powerset.num_powerset_classes..r)intsumranger rrrrnum_powerset_classes8s  zPowerset.num_powerset_classescCsXt|j|j}d}td|jdD]}tt|j|D] }d|||f<|d7}qq|S)atCompute powerset to regular mapping Returns ------- mapping : (num_powerset_classes, num_classes) torch.Tensor mapping[i, j] == 1 if jth regular class is a member of ith powerset class mapping[i, j] == 0 otherwise Example ------- With num_classes == 3 and max_set_size == 2, returns [0, 0, 0] # none [1, 0, 0] # class #1 [0, 1, 0] # class #2 [0, 0, 1] # class #3 [1, 1, 0] # classes #1 and #2 [1, 0, 1] # classes #1 and #3 [0, 1, 1] # classes #2 and #3 rr)torchzerosr#rr"r r)rr powerset_kset_size current_setrrrrDs  zPowerset.build_mappingcCstj|jddS)z#Compute size of each powerset classrdim)r$r!r rrrrrcszPowerset.build_cardinalityFpowersetsoftcCs>|rt|}ntjjtj|dd|j}t||j S)a/Convert predictions from powerset to multi-label Parameter --------- powerset : (batch_size, num_frames, num_powerset_classes) torch.Tensor Soft predictions in "powerset" space. soft : bool, optional Return soft multi-label predictions. Defaults to False (i.e. hard predictions) Assumes that `powerset` are "log probabilities". Returns ------- multi_label : (batch_size, num_frames, num_classes) torch.Tensor Predictions in "multi-label" space. r)) r$expnn functionalone_hotargmaxr#floatmatmulr )rr+r,powerset_probsrrr to_multilabelgs  zPowerset.to_multilabelcCs|j||dS)zAlias for `to_multilabel`)r,)r6)rr+r,rrrforwardszPowerset.forward multilabelcCs&tjtjt||jjdd|jdS)aConvert (hard) predictions from multi-label to powerset Parameter --------- multi_label : (batch_size, num_frames, num_classes) torch.Tensor Prediction in "multi-label" space. Returns ------- powerset : (batch_size, num_frames, num_powerset_classes) torch.Tensor Hard, one-hot prediction in "powerset" space. Note ---- This method will not complain if `multilabel` is provided a soft predictions (e.g. the output of a sigmoid-ed classifier). However, in that particular case, the resulting powerset output will most likely not make much sense. r-r))r)Fr1r$r2r4r Tr#)rr8rrr to_powersetszPowerset.to_powersetmultilabel_permutation.cCs|jdd|f}tj|j|jjtjd}d||jdf}tj|j|dd}tj||dd}|d|dddfkj dd}t | S)aHelper function for `permutation_mapping` property Takes a (num_classes,)-shaped permutation in multilabel space and returns the corresponding (num_powerset_classes,)-shaped permutation in powerset space. This does not cache anything and only works on one single permutation at a time. Parameters ---------- multilabel_permutation : tuple of int Permutation in multilabel space. Returns ------- powerset_permutation : tuple of int Permutation in powerset space. Example ------- >>> powerset = Powerset(3, 2) >>> powerset._permutation_powerset((1, 0, 2)) # (0, 2, 1, 3, 4, 6, 5) N)devicedtyperr-r)r) r r$arangerr=r tiler#r!r2tupletolist)rr<permutated_mappingr@ powers_of_twobeforeafterpowerset_permutationrrr_permutation_powersets$ zPowerset._permutation_powersetcCs2i}tt|j|jD] }|||t|<q |S)aMapping between multilabel and powerset permutations Example ------- With num_classes == 3 and max_set_size == 2, returns { (0, 1, 2): (0, 1, 2, 3, 4, 5, 6), (0, 2, 1): (0, 1, 3, 2, 5, 4, 6), (1, 0, 2): (0, 2, 1, 3, 4, 6, 5), (1, 2, 0): (0, 2, 3, 1, 6, 4, 5), (2, 0, 1): (0, 3, 1, 2, 5, 6, 4), (2, 1, 0): (0, 3, 2, 1, 6, 5, 4) } )rr"rrIrB)rpermutation_mappingr<rrrrJs zPowerset.permutation_mapping)F)__name__ __module__ __qualname____doc__r rrr#r$Tensorrrboolr6r7r;rrIrrJ __classcell__rrrrr%s"     -0r) functoolsr itertoolsrrtypingrr scipy.specialrr$torch.nnr/torch.nn.functionalr0r9Modulerrrrrs