a yf=@stddlZddlmZddlmmZddlmZmZddl m Z ddl m Z Gdddej ZGdd d eZdS) N) FocalLoss VarifocalLoss)bbox_iou)HungarianMatchercsheZdZdZdfdd Zdd d Zdd d ZdddZeddZ ddZ dddZ dddZ Z S)DETRLossa+ DETR (DEtection TRansformer) Loss class. This class calculates and returns the different loss components for the DETR object detection model. It computes classification loss, bounding box loss, GIoU loss, and optionally auxiliary losses. Attributes: nc (int): The number of classes. loss_gain (dict): Coefficients for different loss components. aux_loss (bool): Whether to compute auxiliary losses. use_fl (bool): Use FocalLoss or not. use_vfl (bool): Use VarifocalLoss or not. use_uni_match (bool): Whether to use a fixed layer to assign labels for the auxiliary branch. uni_match_ind (int): The fixed indices of a layer to use if `use_uni_match` is True. matcher (HungarianMatcher): Object to compute matching cost and indices. fl (FocalLoss or None): Focal Loss object if `use_fl` is True, otherwise None. vfl (VarifocalLoss or None): Varifocal Loss object if `use_vfl` is True, otherwise None. device (torch.device): Device on which tensors are stored. PNTFrcst|dur$ddddddd}||_tddddd|_||_||_|rTtnd|_|rdt nd|_ ||_ ||_ d|_ dS) a Initialize DETR loss function with customizable components and gains. Uses default loss_gain if not provided. Initializes HungarianMatcher with preset cost gains. Supports auxiliary losses and various loss types. Args: nc (int): Number of classes. loss_gain (dict): Coefficients for different loss components. aux_loss (bool): Use auxiliary losses from each decoder layer. use_fl (bool): Use FocalLoss. use_vfl (bool): Use VarifocalLoss. use_uni_match (bool): Use fixed layer for auxiliary branch label assignment. uni_match_ind (int): Index of fixed layer for uni_match. Nrg?)classbboxgiouZ no_objectmaskZdice)r r r )Z cost_gain)super__init__ncrmatcher loss_gainaux_lossrflrvfl use_uni_match uni_match_inddevice)selfrrrZuse_flZuse_vflrr __class__Y/var/www/html/django/DPS/env/lib/python3.9/site-packages/ultralytics/models/utils/loss.pyr!s zDETRLoss.__init__c Csd|}|jdd\}}tj|||jdftj|jd} | d|dd| dddf} |||d| }|j r|r|j r| ||| } n| || } | t |d|} nt jdd ||d} || |jd iS) z^Computes the classification loss based on predictions, target values, and ground truth scores. loss_classNr r)dtyper.noneZ reductionr )shapetorchzerosrZint64rZscatter_Z unsqueezeviewrrfloatmaxnnZBCEWithLogitsLossmeansumsqueezer) r pred_scorestargets gt_scoresZnum_gtspostfixZ name_classbsnqZone_hotZloss_clsrrr_get_loss_classBs   zDETRLoss._get_loss_classcCsd|}d|}i}t|dkrPtjd|jd||<tjd|jd||<|S|jdtj||ddt|||<d t||d d d ||<||t|||<|jd ||||<d d| DS)zTComputes bounding box and GIoU losses for predicted and ground truth bounding boxes. loss_bbox loss_giourrr r-r$g?T)xywhZGIoUr cSsi|]\}}||qSr)r.).0kvrrr hz+DETRLoss._get_loss_bbox..) lenr&tensorrrFZl1_lossrr-items)r pred_bboxes gt_bboxesr2Z name_bboxZ name_gioulossrrr_get_loss_bboxXs   &zDETRLoss._get_loss_bboxc  Cs,tj|durdnd|jd} |dur`|jr`|j||j||j||||durV||jnd| d}tt||D]\} \} } |dur|| nd}|j| | ||||| ||d }| d|d|7<| d |d |7<| d |d |7<qnd || dd|| d d|| d i} | S)zGet auxiliary losses.Nr r9masksgt_mask)rJrKr2 match_indicesrr rr6r r7Zloss_class_auxZ loss_bbox_auxZ loss_giou_aux) r&r'rrrr enumeratezip _get_loss)rrDr/rEgt_cls gt_groupsrLr2rJrKrFiZ aux_bboxesZ aux_scoresZ aux_masksZloss_rrr _get_loss_auxs@  zDETRLoss._get_loss_auxcCsLtddt|D}tdd|D}tdd|D}||f|fS)z[Returns batch indices, source indices, and destination indices from provided match indices.cSs g|]\}\}}t||qSr)r&Z full_like)r;rRsrc_rrr r?z'DETRLoss._get_index..cSsg|] \}}|qSrr)r;rTrUrrrrVr?cSsg|] \}}|qSrr)r;rUdstrrrrVr?)r&catrM)rLZ batch_idxZsrc_idxZdst_idxrrr _get_indexszDETRLoss._get_indexcsDtfddt||D}tfddt||D}||fS)z[Assigns predicted bounding boxes to ground truth bounding boxes based on the match indices.cs@g|]8\}\}}t|dkr$||ntjd|jdjdqSrr"r9r@r&r'r%r)r;trRrUrrrrVs z1DETRLoss._get_assigned_bboxes..cs@g|]8\}\}}t|dkr$||ntjd|jdjdqSrZr[)r;r\rUjr]rrrVs )r&rXrN)rrDrErLZ pred_assignedZ gt_assignedrr]r_get_assigned_bboxess  zDETRLoss._get_assigned_bboxesc  Cs| dur |j|||||||d} || \} } || || }}|jdd\} } tj| | f|j|j|jd}|| || <tj| | g|jd}t |rt | |dd d|| <i}| ||||t ||| |||||S) z Get losses.NrIr )rr!r9T)r:r")rrYr%r&fullrrr!r'r@rdetachr.updater5rG)rrDr/rErPrQrJrKr2rLidxgt_idxr3r4r0r1rFrrrrOs  zDETRLoss._get_lossc Ks|j|_|dd}|d|d|d}}} |j|d|d||| ||d} |jr| ||dd|dd||| ||| S)a Calculate loss for predicted bounding boxes and scores. Args: pred_bboxes (torch.Tensor): Predicted bounding boxes, shape [l, b, query, 4]. pred_scores (torch.Tensor): Predicted class scores, shape [l, b, query, num_classes]. batch (dict): Batch information containing: cls (torch.Tensor): Ground truth classes, shape [num_gts]. bboxes (torch.Tensor): Ground truth bounding boxes, shape [num_gts, 4]. gt_groups (List[int]): Number of ground truths for each image in the batch. postfix (str): Postfix for loss names. **kwargs (Any): Additional arguments, may include 'match_indices'. Returns: (dict): Computed losses, including main and auxiliary (if enabled). Note: Uses last elements of pred_bboxes and pred_scores for main loss, and the rest for auxiliary losses if self.aux_loss is True. rLNclsZbboxesrQr"r2rL)rgetrOrrbrS) rrDr/batchr2kwargsrLrPrErQ total_lossrrrforwards zDETRLoss.forward)rNTTFFr)r)r)NrNN)NNrN)r)__name__ __module__ __qualname____doc__rr5rGrS staticmethodrYr_rOrk __classcell__rrrrr s&!  8 8  $rcs.eZdZdZdfdd ZeddZZS)RTDETRDetectionLossa# Real-Time DeepTracker (RT-DETR) Detection Loss class that extends the DETRLoss. This class computes the detection loss for the RT-DETR model, which includes the standard detection loss as well as an additional denoising training loss when provided with denoising metadata. Nc s|\}}t|||}|dur~|d|d} } t|dt| ksJJ| | |d} tj|||d| d} || n|fdd|D|S) aF Forward pass to compute the detection loss. Args: preds (tuple): Predicted bounding boxes and scores. batch (dict): Batch data containing ground truth information. dn_bboxes (torch.Tensor, optional): Denoising bounding boxes. Default is None. dn_scores (torch.Tensor, optional): Denoising scores. Default is None. dn_meta (dict, optional): Metadata for denoising. Default is None. Returns: (dict): Dictionary containing the total loss and, if applicable, the denoising loss. N dn_pos_idx dn_num_grouprQ_dnrfcs$i|]}|dtjdjdqS)rur8r9)r&rAr)r;r<r]rrr>Jr?z/RTDETRDetectionLoss.forward..)rrkr@get_dn_match_indicesrbkeys) rpredsrhZ dn_bboxesZ dn_scoresZdn_metarDr/rjrsrtrLZdn_lossrr]rrk,s zRTDETRDetectionLoss.forwardcCsg}tdg|ddd}t|D]\}}|dkrtj|tjd||}||}t||t|ksxJddt||dt|d||||fq*|tj dgtjd tj dgtjd fq*|S) a Get the match indices for denoising. Args: dn_pos_idx (List[torch.Tensor]): List of tensors containing positive indices for denoising. dn_num_group (int): Number of denoising groups. gt_groups (List[int]): List of integers representing the number of ground truths for each image. Returns: (List[tuple]): List of tuples containing matched indices for denoising. rNr")endr!zExpected the same length, zbut got z and z respectively.)r!) r&Z as_tensorZcumsum_rMZarangelongrepeatr@appendr')rsrtrQZdn_match_indicesZ idx_groupsrRZnum_gtrdrrrrvNs  ,z(RTDETRDetectionLoss.get_dn_match_indices)NNN)rlrmrnrorkrprvrqrrrrrr$s"rr)r&Ztorch.nnr+Ztorch.nn.functionalZ functionalrBZultralytics.utils.lossrrZultralytics.utils.metricsropsrModulerrrrrrrs