o wi@sTddlZddlmZmZmZddlmZmZmZm Z m Z dgZ GdddeeZ dS)N)LossTyping typecheck) LabelsType LengthsTypeLossType NeuralType ProbsTypeBCELossc szeZdZdZeddZeddZddedd gd d d fd e d e d ej de de de f fdd Z eddZZS)r zk Computes Binary Cross Entropy (BCE) loss. The BCELoss class expects output from Sigmoid function. cCs$tdttdttdtdS)z'Input types definitions for AnguarLoss.)BTCr )probslabels target_lens)rr rrselfra/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/nemo/collections/asr/losses/bce_loss.py input_typess   zBCELoss.input_typescCsdttdiS)z Output types definitions for binary cross entropy loss. Weights for labels can be set using weight variables. loss) elements_type)rrrrrr output_types&szBCELoss.output_typesmean?g?g?F reductionalphaweight sorted_preds sorted_lossclass_normalizationcsPt||_|rd|_nd|_||_tjj|jd|_||_ ||_ d|_ dS)a! A custom loss function that supports class normalization, weighted binary cross-entropy, and optional sorting. Args: reduction (str): Specifies the reduction to apply to the output, options are 'mean', 'sum', or 'none'. Default is 'mean'. alpha (float): Scaling factor for loss (unused in this implementation). Default is 1.0. weight (torch.Tensor): Class weights for the binary cross-entropy loss. Default is [0.1, 0.9]. sorted_preds (bool): If True, assumes predictions are sorted. Default is False. sorted_loss (bool): If True, sorts the loss before reduction. Default is False. class_normalization (bool): If True, uses 'none' reduction for per-class loss. Default is False. noner)rgư>N) super__init__r r loss_weighttorchnnr loss_frreps)rrrrrrr  __class__rrr#-s  zBCELoss.__init__c s"fddtjdD}fddtjdD}tj|ddtj|ddt}tdj }|j dvr|j dvrwd j dd|j }|| }tj |d d d }||}|d d d f}n t}|j dvrt} jdjd j } | | dk<d | | d k<n$t} n|j dks|j st} t}|jdkr|}|S|jdkr|}|S|jdkr|j dvr | || }|S|}|S)a) Calculate binary cross entropy loss based on probs, labels and target_lens variables. Args: probs (torch.tensor) Predicted probability value which ranges from 0 to 1. Sigmoid output is expected. labels (torch.tensor) Groundtruth label for the predicted samples. target_lens (torch.tensor): The actual length of the sequence without zero-padding. Returns: loss (NeuralType) Binary cross entropy loss value. c&g|]}|d|ddfqSNr.0k)rrrr `&z#BCELoss.forward..rcr+r,rr-)rrrrr0ar1)dimg)class class_binarybinary)r3r4g?r)minmaxN)r5r4r!sumr)rangeshaper%cat zeros_likedetachclonetensortodevicer r9r(clampr8 expand_as ones_likerr'r) rrrr probs_list targets_list norm_weightrnorm_weight_norm binary_weight one_weightr)rrrrforwardOsF      $      zBCELoss.forward)__name__ __module__ __qualname____doc__propertyrrr%r@strfloatTensorboolr#rrL __classcell__rrr)rr s6   ") r%nemo.core.classesrrrnemo.core.neural_typesrrrrr __all__r rrrrs