o ei<@sddlZddlZddlmZddlmZmZmZddl m Z m Z m Z m Z mZmZmZer3ddlmZer9z2LwDetrHungarianMatcher.forward..cSrboxesrrrrrr:rg?g@g:0yE>)pcSg|]}t|dqSr lenrrrrrQdimcSsg|] \}}t||qSrr)ricrrrrWscsDg|]\}}t|d|dgt|d|dgfqS)rr)np concatenate)rindice1indice2group_idgroup_num_queriesrrr[s cSs0g|]\}}tj|tjdtj|tjdfqS))dtype)torch as_tensorint64)rr*jrrrrbs0)shapeflattensigmoidr4catlogr3tofloat32cdistr r bbox_cost class_cost giou_costviewcpusplitrange enumeratezip)selfoutputstargets group_detr batch_size num_queriesout_probout_bbox target_ids target_bboxalphagammaneg_cost_classpos_cost_classrAr3r@rB cost_matrixsizesindicescost_matrix_listgroup_cost_matrix group_indicesrr0rforward+s>"      zLwDetrHungarianMatcher.forwardN)__name__ __module__ __qualname__r4no_gradr]rrrrr*srcsdeZdZfddZddZeddZddZd d Z d d Z d dZ ddZ ddZ ZS)LwDetrImageLosscs,t||_||_||_||_||_dSN)super__init__matcher num_classes focal_alphalossesrL)rIrfrgrhrirL __class__rrrefs  zLwDetrImageLoss.__init__cCsJd|vrtd|d}||}tddt||D}|j}d} |d|} tjddt||Ddd } ttt| t| d} | |j } | } | }t|}|| }t|}|||||| d |}t|d  }|||<d |||<| ||d |}||}d |i}|S) Nrz#No logits were found in the outputscS g|] \}\}}|d|qSrr)rt_Jrrrru z/LwDetrImageLoss.loss_labels..rrcSrlr rrrmrnr*rrrryrprr(rg{Gz?loss_ce)KeyError_get_source_permutation_idxr4r;rHrhdiagr rdetachr=r3cloner: zeros_likelistappendpowclampr<sum)rIrJrKrY num_boxes source_logitsidxtarget_classes_orSrT src_boxes target_boxes iou_targetspos_iousprob pos_weights neg_weightspos_indrmrrrirrr loss_labelsos6        zLwDetrImageLoss.loss_labelsc Csb|d}|j}tjdd|D|d}|djdkd}tj | | } d| i} | S) z Compute the cardinality error, i.e. the absolute error in the number of predicted non-empty boxes. This is not really a loss, it is intended for logging purposes only. It doesn't propagate gradients. rcSr$rr%rrrrrr'z4LwDetrImageLoss.loss_cardinality..)devicer#g?rcardinality_error) rr4r5r:maxvaluesr}nn functionall1_lossfloat) rIrJrKrYr~rrtarget_lengths card_predcard_errrirrrloss_cardinalitysz LwDetrImageLoss.loss_cardinalityc Csd|vrtd||}|d|}tjddt||Ddd}tjj||dd}i} ||| d <d t t t |t |} | || d <| S) a< Compute the losses related to the bounding boxes, the L1 regression loss and the GIoU loss. Targets dicts must contain the key "boxes" containing a tensor of dim [nb_target_boxes, 4]. The target boxes are expected in format (center_x, center_y, w, h), normalized by the image size. rz#No predicted boxes found in outputscSrlr rrqrrrrrpz.LwDetrImageLoss.loss_boxes..rr(none) reduction loss_bboxr loss_giou) rsrtr4r;rHrrrr}rur r) rIrJrKrYr~r source_boxesrrrirrrr loss_boxess  zLwDetrImageLoss.loss_boxesc Csd|vrtd||}||}|d}||}dd|D}t|\} } | |} | |} tjj|dddf| j ddddd }|ddd f d }| d } | |j } t || |t || |d } | S) z Compute the losses related to the masks: the focal loss and the dice loss. Targets dicts must contain the key "masks" containing a tensor of dim [nb_target_boxes, h, w]. pred_masksz#No predicted masks found in outputscSr)masksrrrmrrrrrz.LwDetrImageLoss.loss_masks..NbilinearF)sizemode align_cornersrr) loss_mask loss_dice)rsrt_get_target_permutation_idxr decomposer=rr interpolater8r9rCr r ) rIrJrKrYr~ source_idx target_idx source_masksr target_masksvalidrirrr loss_maskss(       zLwDetrImageLoss.loss_maskscC4tddt|D}tdd|D}||fS)NcSs g|] \}\}}t||qSrr4 full_like)rr*sourcernrrrrrpz?LwDetrImageLoss._get_source_permutation_idx..cSsg|]\}}|qSrr)rrrnrrrrrr4r;rG)rIrY batch_idxrrrrrtz+LwDetrImageLoss._get_source_permutation_idxcCr)NcSs g|] \}\}}t||qSrr)rr*rntargetrrrrrpz?LwDetrImageLoss._get_target_permutation_idx..cSsg|]\}}|qSrr)rrnrrrrrrr)rIrYrrrrrrrz+LwDetrImageLoss._get_target_permutation_idxcCs@|j|j|j|jd}||vrtd|d||||||S)N)labels cardinalityr!rzLoss z not supported)rrrr ValueError)rIlossrJrKrYr~loss_maprrrget_lossszLwDetrImageLoss.get_lossc s|jr|jnd}dd|D}||||}tdd|D}||}tj|gtjtt | j d}d}t rHt jikrHt|}t j}tj||dd}i}|jD]} ||| ||||qXd|vrt|dD]/\} || ||}|jD] } | d krq|| | |||} fd d| D} || qqqd |vr|d } |j| ||d }|jD]} || | |||} d d| D} || q|S)a This performs the loss computation. Args: outputs (`dict`, *optional*): Dictionary of tensors, see the output specification of the model for the format. targets (`list[dict]`, *optional*): List of dicts, such that `len(targets) == batch_size`. The expected keys in each dict depends on the losses applied, see each loss' doc. rcSs&i|]\}}|dkr|dkr||qS) enc_outputsauxiliary_outputsrrkrrrr s z+LwDetrImageLoss.forward..css|] }t|dVqdS)rNr%rrrr sz*LwDetrImageLoss.forward..)r3r)minrrc i|] \}}|d|qSrnrrr*rrr#rpr)rLcSi|] \}}|d|qS_encrrrrrr+)trainingrLitemsrfr}r4r5rnextiterrrrr _shared_stater num_processesr|itemriupdaterrG) rIrJrKrLoutputs_without_aux_and_encrYr~ world_sizerirrl_dictrrrrr]sF "      zLwDetrImageLoss.forward)r^r_r`rerr4rarrrrtrrr] __classcell__rrrjrrbes  $ " rbc  st|j|j|jd} gd} t| |j|j| |jd} | |i} d}|| d<|| d<||d| d<|j r>t ||}|| d<| | |d |j d |j d <|j rpi}t |jd D]|fd d DqZ|dd D}|tfddD}||fS)N)rAr@rB)rr!r)rfrgrhrirLrr)rrrrr)rrrrcrrrrrrrr^rpz0LwDetrForObjectDetectionLoss..cSrrrrrrrr`rc3s(|]}|vr||VqdSrcr)rr) loss_dict weight_dictrrrbs&z/LwDetrForObjectDetectionLoss..)rrAr@rBrb num_labelsrhrLr=auxiliary_lossrbbox_loss_coefficientgiou_loss_coefficientrFdecoder_layersrrr})rrrrconfig outputs_class outputs_coordenc_outputs_classenc_outputs_coordkwargsrfri criterion outputs_lossraux_weight_dictenc_weight_dictrr)r*rrrLwDetrForObjectDetectionLoss1sD          r)NNNN)numpyr,r4torch.nnrutilsrrrloss_for_object_detectionrrr r r r r transformers.image_transformsrscipy.optimizer accelerateraccelerate.utilsrrModulerbrrrrrs(  $    ;S