pyi=DSSKrSSKJr SSKJs Jr SSKJr SSKJ r SSK J r J r \"5(aSSK Jr SrSS jrS \S \R$S \S \R$4SjrS\R$S \S \S \R$4SjrSSjr"SS\ 5rSSjrg)N)is_vision_available)box_iou)RTDetrHungarianMatcher RTDetrLoss)center_to_corners_formatcR[X5VVs/sH up#X#S.PM snn$s snnf)N)logits pred_boxeszip) outputs_class outputs_coordabs Z/home/mande/repo/quber/.venv/lib/python3.13/site-packages/transformers/loss/loss_d_fine.py _set_aux_lossrs&7:=7X Y7Xtqq *7X YY Ys#c l[XX#5VVVV s/sHupgpUUUU UUS.PM sn nnn$s sn nnnf)N)r r pred_corners ref_pointsteacher_cornersteacher_logitsr ) rroutputs_corners outputs_refrrrrcds r_set_aux_loss2r!sNmOY ZJA!.,  Z   s. max_num_binsup reg_scalereturnc:[US5[U5-n[US5[U5-S-nUS-SUS- - -n[US-S- SS5Vs/sH oeU-*S-PM nn[SUS-5Vs/sH oeU-S- PM nnU*/U-[R"USS5/-U-U/-n U V s/sH)oR 5S:aU OU R S5PM+ n n [R "U S5n U $s snfs snfs sn f)u Generates the non-uniform Weighting Function W(n) for bounding box regression. Args: max_num_bins (int): Max number of the discrete bins. up (Tensor): Controls upper bounds of the sequence, where maximum offset is ±up * H / W. reg_scale (float): Controls the curvature of the Weighting Function. Larger values result in flatter weights near the central axis W(max_num_bins/2)=0 and steeper weights at both ends. Returns: Tensor: Sequence of Weighting Function. rrrN)absrangetorch zeros_likedim unsqueezecat) rr r! upper_bound1 upper_bound2stepi left_values right_valuesvaluesvs rweighting_functionr41s+r!u:I.Lr!u:I.2L 1 !|a'7"8 9D/4\Q5F5JAr/RS/R!q[>A%/RKS-21la6G-HI-HaK!O-HLIm_{ *e.>.>r!uT{.K-L L| [_k^l lFN+002M"0K"1#45L2- ;K$.k1I#JL "% (D"DK#Q;),L%&(+K$%$'G !"2),L%&(+K$%$03$6G ! + --c6[U5nUSS2S4USS2S4- USU- S-- SU-- nUSS2S4USS2S4- USU- S-- SU-- nUSS2S4USS2S4- USU- S-- SU-- nUSS2S 4USS2S4- USU- S-- SU-- n [R"XgX/S 5n [XX45upn UbU R SX%- S 9n U R S 5R 5U R 5U R 54$) a Converts bounding box coordinates to distances from a reference point. Args: points (Tensor): (n, 4) [x, y, w, h], where (x, y) is the center. bbox (Tensor): (n, 4) bounding boxes in "xyxy" format. max_num_bins (float): Maximum bin value. reg_scale (float): Controlling curvarture of W(n). up (Tensor): Controlling upper bounds of W(n). eps (float): Small value to ensure target < max_num_bins. Returns: Tensor: Decoded distances. Nr).rgؗҜ<?r).rrNr$minmax)r%r'stackrJclampr9detach) pointsbboxrr!r epslefttoprightbottom four_lensrBrCs r bbox2distancer]sd II 1a4L41: %&.9*Du*L MPSV_P_ _D !Q$<$q!t* $))Ce)K LsU^ ^C !Q$Z&A, &6&>I+E+M NQTW`Q` `E1a4j6!Q$<'F6NY,F,N ORUXaRa aF T6;I+7 QZ+_(I[OO|/AOB   R ' ' )<+>+>+@+BTBTBV VVrKcH^\rSrSrSrU4SjrSSjrS SjrSrSr U=r $) DFineLossa= This class computes the losses for D-FINE. The process happens in two steps: 1) we compute hungarian assignment between ground truth boxes and the outputs of the model 2) we supervise each pair of matched ground-truth / prediction (supervise class and box). Args: matcher (`DetrHungarianMatcher`): Module able to compute a matching between targets and proposals. weight_dict (`Dict`): Dictionary relating each loss with its weights. These losses are configured in DFineConf as `weight_loss_vfl`, `weight_loss_bbox`, `weight_loss_giou`, `weight_loss_fgl`, `weight_loss_ddf` losses (`list[str]`): List of all the losses to be applied. See `get_loss` for a list of all available losses. alpha (`float`): Parameter alpha used to compute the focal loss. gamma (`float`): Parameter gamma used to compute the focal loss. eos_coef (`float`): Relative classification weight applied to the no-object category. num_classes (`int`): Number of object categories, omitting the special no-object category. c>[TU]U5 [U5UlURUlUR UR URURURS.Ul /SQUl URUl [R"[R "UR"/5SS9Ulg)N)loss_vfl loss_bbox loss_giouloss_fglloss_ddf)vflboxeslocalF) requires_grad)super__init__rmatcherrweight_loss_vflweight_loss_bboxweight_loss_giouweight_loss_fglweight_loss_ddf weight_dictlossesr!nn Parameterr'tensorr )selfconfig __class__s rrlDFineLoss.__init__s  -f5 "//..0000....  0 )),,u||VYYK8NrKcUR5nUS-n [R"XSS9URS5-[R"XSS9URS5--n UbUR 5nX-n UbU R 5U- n U $US:XaU R 5n U $US:XaU R 5n U $)Nrnone reductionr$meanr:)r<F cross_entropyr9r;r:r) rxpredlabelrBrCweightr avg_factordis_left dis_rightlosss r unimodal_distribution_focal_loss*DFineLoss.unimodal_distribution_focal_losss::<qL t@;CVCVWYCZZ]^]l]l v^   $^%%  \\^F=D  !88: *D  & 99;D % 88:D rKc 0nSU;Ga0URU5n[R"[X#5VV V s/sHunupUSU PM sn n nSS9n USUR SUR S-5n USUR 5n [R"5 [U [U 5UR URUR5Ul SSS5 URupn[R"[[US U5[U 55S5nURS5R!SSS 5R S5R 5nUR#U UUUUUS 9US 'USR SUR S-5n US R SUR S-5n[R$"X5(aU R'5S-US'U$USR)5R+SS9Sn[R,"U[R.S9nSUU'URS5R!SSS 5R S5nUR1UU5R3UR45UU'URS5R!SSS 5R S5R 5nUUS--[6R8"SS9"[:R<"X- SS9[:R>"UR 5U- SS95R'S5-nSUS R@S- nUR'5U-S-U)R'5U-S-sUl!Ul"URG5(aUURI5OSnU)RG5(aUU)RI5OSnUURB-UURD--URBURD-- US'U$s sn n nf!,(df  GNe=f)zYCompute Fine-Grained Localization (FGL) Loss and Decoupled Distillation Focal (DDF) Loss.rrhrr7r$rrNr )rrerrfr)dtypeTrr}r~rM)%_get_source_permutation_idxr'r+rr9rrTno_gradr]r r!r fgl_targetsdiagrr*repeatrequalr:sigmoidrQr(bool reshape_astorru KLDivLossr log_softmaxsoftmaxshapenum_posnum_neganyr)rxoutputstargetsrA num_boxesTrtidxt_r/ target_boxesrrtarget_cornersrBrCiousweight_targetsweight_targets_localr?loss_match_local batch_scaleloss_match_local1loss_match_local2s r loss_localDFineLoss.loss_locals  W $227;C 99SEZ%[EZ 6AajmEZ%[abcL">237??TEVEVYZEZ\L .s3::2::2@Q@QTU@UWL$%67??TEVEVYZEZ\N{{<88%1%5%5%7!%;z"@ =(//?'@'H'H'J'N'NSU'N'VWX'Y$''(rs  '.<FZ fj  Sellsu||2<.U\\<.<.<.%,,<.~W8TD TDz!A.rK