^jjzVUdZddlZddlZddlZddlmZddlmZddlm cm Z ddl Z ddlmZddlmZddlmZddlmZeZe j.gd e j0 d z Zd ZeZeeed <eddGddZedGddZ dedefdZ!deddfdZ"dededzfdZ#dede$eeffdZ%dede&e'dzde&e'dzfdZ(dede&e'dzde&e'dzfdZ)deddfdZ*dede&e'dzde&efd Z+ded!e&edefd"Z,ded#e&e$e-efdefd$Z.d%e j^de j^fd&Z0d'e&e j^d(e&ede j^fd)Z1d'e j^ddfd*Z2d+ed,e3ddfd-Z4Gd.d/Z5dd0d1ed,e3ddfd2Z6y)3a'COCO evaluator for ONNX/TRT export benchmarking. Provides :class:`CocoEvaluator` used by :mod:`rfdetr.export.benchmark` to compute mAP during ONNX and TensorRT inference benchmarks. Implementation mirrors torchvision's evaluator structure but uses ``faster_coco_eval`` as the runtime backend. N) dataclass)Any)COCO)COCOeval) all_gather) get_logger)gp= ף??r ffffff?r HzG?r ףp= ?r ףp= ?r Q?rףp= ?r{Gz?rdtypeg$@g?"_WARNED_CUSTOM_KEYPOINT_OKS_COUNTST)frozenslotscBeZdZUdZeeed<eed<eedzed<y)_KeypointCategoryGroupzDKeypoint categories sharing one keypoint count and OKS sigma vector. category_idskeypoint_countNkeypoint_oks_sigmas)__name__ __module__ __qualname____doc__listint__annotations__floatf/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/rfdetr/evaluation/coco_eval.pyrrDs$Ns)et++r$r)rcPeZdZUdZeeed<ejed<ee ed<y)_GroupedKeypointCOCOevalzLAggregate COCO keypoint stats for categories with different keypoint counts.groupsstatsevalsN) rrrrrrr!npndarrayrr#r$r%r'r'Ms%V ' (( :: >r$r'coco_gtreturnct|trt|dr|St}tj|j |_|j t|dd}| t|dtj||S)zAReturn a faster-coco-eval COCO object for evaluator construction. cat_img_map label2catN) isinstancerhasattrcopydeepcopydataset createIndexgetattrsetattr)r- faster_cocor1s r%_ensure_faster_cocor;Vsq'4 WWm%D&K--8Kd3I [$-- *BC r$c i}|jjdgD]j}|jd}t|tr|||<|jd}d|vs?t|tsPt d|dddD|d<l|j jD]f\}}|jd}d|vr+t|trt d |dddD|d<|j|}|Zd|vs_|d|d<hy) zEPopulate missing COCO ``num_keypoints`` fields from visibility flags. annotationsid keypoints num_keypointsc3,K|] }|dkDs dywrNr#.0 visibilitys r% z*_backfill_num_keypoints..m-dJU_bcUca-d Nc3,K|] }|dkDs dywrBr#rDs r%rGz*_backfill_num_keypoints..rrHrI)r6getr2r rsumannsitems)r-annotations_by_id annotation annotation_idr?dataset_annotations r%_backfill_num_keypointsrUds35oo))-<e "t, mS )/9 m ,NN;/ * ,It1L*--d)ADqD/-d*dJ ' e&-\\%7%7%9N! zNN;/ * ,It1L*--d)ADqD/-d*dJ '.22=A  )o.K2<_2M  / Nr$ct|jDchc] }|dkDs | }}|syt|dkDrtdt |dt t |Scc}w)zIInfer a single keypoint count from COCO category metadata or annotations.rNrCz_COCO keypoint evaluation requires one keypoint count across evaluated categories; found counts .)"_infer_keypoint_counts_by_categoryvalueslen ValueErrorsortednextiter)r-countcountss r%_infer_keypoint_countraxsw!CG!L!S!S!U cY^abYbe cF c  6{Q "6N+1 .   V ds A*A*ci}|jjD]A\}}|jd}t|ts(|s+t ||t |<C|jjdgD]_}|jd}t|tr|s't |d}t|j|dt |dz||<a|jjD]_}|jd}t|tr|s't |d}t|j|dt |dz||<a|S)zRInfer keypoint count for each category from COCO category metadata or annotations.r?r= category_idrrK) catsrPrMr2rrZr r6maxrOrY)r-r`rccategoryr?rRs r%rXrXsDF!(!3!3!56 XLL- i &9'*9~F3{# $6 oo))-<S NN;/ )T*) *]34 !&**[!"QR{ S ll))+S NN;/ )T*) *]34 !&**[!"QR{ S Mr$rcnt|}|tj|tj}|jdk7s|j dk(r t dtj|jrtj|dkr t d|+|j |k7rt d|j d|d |jS||ttk(ryt|tj|ttjjS) z#ERZZ X _ _ aar$rcV|V|ttk(ryt|tj|t tj jStj|tj }|jdk7s|jdk(r tdtj|jrtj|dkr td|j|krtd|jd|d |d|jS) z0Resolve OKS sigmas for one keypoint-count group.NrrCrrhrirjrkrW)rZrtrur+rvrwrmrsrlrnror[rprqrr)rrrxs r%"_resolve_group_keypoint_oks_sigmasr{s " S!=> >,^<ww~'IQSQ[Q[\ccee ZZ+2:: >F {{a6;;!+\]] ;;v  " " $v{(;VWW {{^#)&++6]^l]mmn o   /> " ) ) ++r$cv|tvrytj|tjd|ty)zBWarn once per keypoint count when using uniform custom OKS sigmas.NzCOCO keypoint metadata defines %s keypoints, but no keypoint_oks_sigmas were provided. Using uniform OKS sigma %.3f for custom keypoint evaluation.)raddloggerwarningrw)rs r%rurus4;;&**>: NN G* r$c <t|}i}|jD],\}}|dkr |j|gj|.t |jDcgc]&\}}t t ||t ||(c}}Scc}}w)zPBuild category groups that can each be evaluated by one COCO keypoint evaluator.r)rrr)rXrP setdefaultappendr\rr{)r-rcounts_by_categorygrouped_category_idsrcrrs r%_build_keypoint_category_groupsrs Sf g   s)+Brc *t|}tj|j}|j dgDcgc]}t |d|vs|c}|d<|j dgDcgc]!}t |j dd|vr|#c}|d<t }||_|jt|dd}|4t|d|jD cic] \}} | |vs || c} }|Scc}wcc}wcc} }w)zNReturn a COCO object containing only the requested categories and annotations. categoriesr>r=rcr1N) setr4r5r6rMr rr7r8r9rP) r-rcategory_id_setr6rfrRfilteredr1labelcat_ids r%_filter_coco_by_category_idsrs,'OmmGOO,G!(\2!>#htnBUYhBhGL "++mR8  z~~mR0 1_ D GM vHH d3I  090A _}ufVE^UF] _ O' `sDD3&D ) D6Dresultsc|rtj||St}tj|jj di|jd<tj|jj dg|jd<tj|jj dg|jd<g|jd<|j |S)z@Build a COCO detections object, including the empty-result case.infoimagesrr=)rloadResr4r5r6rMr7)r-rcoco_dts r%_load_coco_resultsrs||GW--fG"mmGOO,?,?,KLGOOF $ goo.A.A(B.O PGOOH$(MM'//2E2ElTV2W$XGOOL!%'GOOM"  Nr$boxesc|j}|dddfxx|dddfzcc<|dddfxx|dddfzcc<|S)z6Convert boxes from [x1, y1, x2, y2] to [x1, y1, w, h].NrJrrKrC)r4)rs r% _xyxy_to_xywhrsF JJLE !Q$K5A;K !Q$K5A;K Lr$r)weightsc|s&tjddtjStd|D}tj|fdtj}t |D]X}d}d}t ||D]5\}}t ||ks||dkr|t|||zz }||z }7|dkDsQ||z ||<Z|S)zNCompute category-weighted mean COCO stats, ignoring unavailable ``-1`` values. rc32K|]}t|yw)N)rZ)rEitems r%rGz,_weighted_mean_coco_stats..!s.#d).sgr)r+rvrmrerangeziprZr") r)rmax_len aggregatedstat_idx numerator denominatorstatweights r%_weighted_mean_coco_statsrs wwud"**55...G'TLog keypoint COCO stats from an already accumulated evaluator.) )Average Precision(AP) 0.50:0.95rqr)rr0.50rqrrC)rr0.75rqrrJ)rrrmediumrrK)rrrlarger)Average Recall(AR)rrqr)rrrrqr)rrrrqr)rrrrr)rrrrr B {:<18} {} @[ IoU={:<9} | area={:>6s} | maxDets={:>3d} ] = {:0.3f}rN)rZr"r~rformat) r)labels log_templatetitle metric_typeiouareamax_detsrvalues r%_log_keypoint_statsr0sl FXL=CY9{Cx*-e*x*?eHo&T L''{CxQVWXYr$ coco_eval log_summaryc,|r|jt|yttjd5}t j |5|jt|dddddddy#1swYxYw#1swYyxYw)zJAccumulate a COCO evaluator and populate ``stats`` regardless of log mode.NwFr) accumulatepatched_pycocotools_summarizeopenosdevnull contextlibredirect_stdout)rrrs r%_accumulate_and_summarizerDs%i0 bjj# H'  ' ' 0 H  " )) G HHH H HHHs#B A>-B >B B  Bc eZdZdZ d dedeededeedzde ddf d Z d ed e dedzfd Z d eede fdZ de eefddfdZd!dZd!dZd!dZdededdfdZdeddfdZedededee eefdeeddf dZde eefdedee eeffdZde eefdee eeffdZde eefdee eeffdZde eefdee eeffdZy)" CocoEvaluatorz.COCO evaluator that works in distributed mode.Nr- iou_typesrrrr.c Jt|ttfsJtjt |}d}g}d|vr1t |t||}t|dkr t||}||_ ||_ t|dd|_ ||_i|_|D]}|dk(rNt|dkDr@t!|t#j$ddt"j&g|j|<V|dk(rd|ini} t)|fd |i| } |dk(rd gndd |g| j*_| |j|<g|_|D cic]} | gc} |_t3|j4j7|_t;||_d |_||_ ycc} w) Nr?rCr1rrr)r(r)r*kpt_oks_sigmasiouTyperrF)!r2rtupler4r5r;rUrrZryr-rr8r1rrr'r+rvrmrparamsmaxDetsimg_ids coco_resultsrrdkeyscat_idsrX_keypoint_counts_by_category_prefer_raw_category_ids _log_summary) selfr-rrrrresolved_keypoint_oks_sigmaskeypoint_category_groupsiou_typekwargsrks r%__init__zCocoEvaluator.__init__Ts)dE]333-- 3G <='+$AC ) # #G ,'FwPc'd $+,1/KGUh/i,   18d0S"IK! 1H;&3/G+H1+L+C3''%RZZ@,x( IQU`I`&(DEfhF E(EfEI/7;/FtQPRT\L]I   $'0DNN8 $ 1#% MV=Wae=W7<<,,./ ,Nw,W)(-%' >Xs F ruse_raw_category_idsc|r||jvr|SdS|j-|jj|}||jvr|SdS||jvr|Sy)z0Resolve a predicted label to a COCO category_id.N)rr1rM)rrrrcs r%_resolve_category_idz"CocoEvaluator._resolve_category_idsc !T\\15 ;t ; >> %..,,U3K"-"=; G4 G DLL Lr$rc|jy|jryt|jjt|jj k(}|rd|_yy)zCDetect whether model predictions are already raw COCO category IDs.TF)r1rrrrY)rr uses_raw_idss r%_should_use_raw_category_idsz*CocoEvaluator._should_use_raw_category_idssZ >> !  ( (DNN//12d4>>;P;P;R6SS ,0D )r$ predictionsc&ttjt|j}|jj ||j D]2}|j||}|j|j |4y)z!Accumulate per-image predictions.N) rr+uniquerrextendrpreparer)rrrrrs r%updatezCocoEvaluator.updatesqryyk&6&6&8!9:; G$ 8Hll;9G   h ' . .w 7 8r$ct|j}i}t|D]\}}|D] }||vs|||<t|j |_|j D]n}t|j |}g}t|D]5\}} | D]+} |j| d|k(s|j| -7||j |<py)uMerge image IDs and COCO result records across distributed processes. Each image ID is assigned to exactly one rank (first rank that reports it), so predictions for images that appear on multiple ranks due to ``DistributedSampler(drop_last=False)`` padding are included only once. Without this deduplication, padded images produce duplicate DT entries that compete for the per-image ``maxDets`` cap and bias mAP upward in early epochs then downward as genuine new detections are displaced — the "peak-then-decrease" pattern. Contract: This method assumes that when the same ``image_id`` appears on multiple ranks its predictions are *identical* (DDP-padding duplicates). It is **not** safe for evaluation strategies where independent predictions for the same image are produced on different ranks, because only the first rank's predictions are kept and the rest are silently discarded. Single-process path: When ``all_gather`` is called in a non-distributed context it returns a single-element list ``[x]``, so ``rank_idx`` is always 0 and every result passes the ownership check unchanged — the dedup loop is a no-op and all results are preserved. This makes the method safe for single-GPU and ONNX/TRT benchmark use via :mod:`rfdetr.export.benchmark`. image_idN) rr enumerater\rrrrMr) rgathered_img_idsimg_id_to_rankrank_idx rank_img_idsimg_idrgathered_resultsdeduped rank_resultsresults r%synchronize_between_processesz+CocoEvaluator.synchronize_between_processess0&dll3*,&/0@&A 6 "Hl& 6/-5N6* 6 6n1134  2H)$*;*;H*EF ,.G*34D*E /&,*/F%))&*<=Iv./ /+2D  h ' 2r$c |jjD]q\}}t|tr3|j ||j rt |jI|j||t||j sy)z:Accumulate per-image evaluation results into mean metrics.rN) rrPr2r'_evaluate_grouped_keypointsrrr) _evaluaterrrrs r%rzCocoEvaluator.accumulatest#'>>#7#7#9 P Hi)%=>00;$$' 8 NN8Y / %iT=N=N O Pr$c|jjD]Z\}}tjdj |t |t rt|jPt|\y)z&Print and log COCO summary statistics.zIoU metric: {}N) rrPr~rrr2r'rr)rrs r% summarizezCocoEvaluator.summarizesZ#'>>#7#7#9 9 Hi KK(//9 :)%=>#IOO4-i8  9r$rrc|j|}ttjd5}t j |5t |j|}||_ttj|j|j_|jddddddy#1swYxYw#1swYyxYw)zDRun faster-coco-eval evaluation for accumulated COCO result records.rN)rrrrrrrr-cocoDtrr+rrrimgIdsevaluate)rrrrrrs r%rzCocoEvaluator._evaluates##H- "**c " %g++G4 %,T\\7C#*  *.ryy/F*G   '""$  % % % % % % %s$CA%B6%C6B? ;CC  grouped_evalcg|_g}g}|jd}ttj|j }|j D] }t|j}t|j|j}|D cgc]} t| d|vs| } } t|d|j} dg| j_|j!| || |t#| d|jj%| |j%tj&| j(tj*|j%t-|j#t/|||_ycc} w) zERun keypoint evaluation per keypoint-count group and aggregate stats.r?rc)rrrFrrN)r*rrr+rrr(rrrr-r rrrr"_evaluate_grouped_keypoint_resultsrrrlr)rmrZr) rr  group_stats group_weights all_resultsrgrouprgroup_gtr group_resultsgroup_coco_evals r%rz)CocoEvaluator._evaluate_grouped_keypointssL (* #% '' 4 ryy./!(( :E!%"4"45O3DLL%BTBTUH2=oVMEZA[_nAnVoMo&#$88O /1TO " " *  3 3OX}^e f %o5 I    % %o 6   rzz/*?*?rzzR S  U%7%7!8 9 : 7{MR ps F"Frrcttjd5}tj|5t ||}||_||j_|jddddddy#1swYxYw#1swYyxYw)z)Evaluate one grouped keypoint result set.rN) rrrrrrr rr r )rr-rrrrs r%rz0CocoEvaluator._evaluate_grouped_keypoint_resultss"**c " %g++G4 %,Wg>#*  *1   '""$  % % % % % % %s"B5A7&B7B <BB c|dk(r|j|S|dk(r|j|S|dk(r|j|Stdj |)z:Convert predictions to COCO format for the given iou_type.bboxsegmr?zUnknown iou type {})prepare_for_coco_detectionprepare_for_coco_segmentationprepare_for_coco_keypointr[r)rrrs r%rzCocoEvaluator.preparesd v 22;? ?  55kB B  $11+> >299(CD Dr$c g}|jD]\}}t|dk(r|d}t|jj j }|dj }|dj }|j |}t|D]6\} } |j|| |} | |j|| | || d8|S)z5Format bounding-box predictions as COCO result dicts.rrscoresr)rrcrscore) rPrZrcpunumpyrsrrrr) rrr original_id predictionrrrrrboxrcs r%rz(CocoEvaluator.prepare_for_coco_detections '2'8'8':  #K:!#w'E!%))+"3"3"56==?E)002F)002F#'#D#DV#L #E* 3"77q CWX &##$/'2 #!'   *r$c g}|jD]2\}}t|dk(r|d}|d}|d}|dkD}|dj}|dj}|j|}|D cgc]e} t j t j| jdddddt jft jddg} } | D]} | d jd | d <t| D]6\} } |j|| |} | |j|| | || d 85|Scc} w) z:Format segmentation mask predictions as COCO result dicts.rrrmasks?NF)rorderr`zutf-8)rrc segmentationr )rPrZrsr mask_utilencoder+arrayr!newaxisuint8decoderrr)rrrr#r$rrr'rmaskrlesrlerrcs r%rz+CocoEvaluator.prepare_for_coco_segmentation7s '2'8'8':  #K:!#)F)Fw'ECKE)002F)002F#'#D#DV#L "  $((*Q1bjj5H*IQSQYQYad!efghiD > #H 4 4W =H  >$D/ 3"77q CWX &##$/'2(+!'  + B's8A*E c Zg}|jD]\}}t|dk(r|d}t|jj j }|dj }|dj }|d}|j dj }|j|} t|D][\} } |j|| | } | |jj| } | | d| d z} |j|| | || d ]|S) z1Format keypoint predictions as COCO result dicts.rrrrr?rC) start_dimNrK)rrcr?r ) rPrZrr!r"rsflattenrrrrrMr)rrrr#r$rrrr?rrkeypointrcrs r%rz'CocoEvaluator.prepare_for_coco_keypoint]sM '2'8'8':  #K:!#w'E!%))+"3"3"56==?E)002F)002F";/I!))A)6==?I#'#D#DV#L (3  8"77q CWX &!%!B!B!F!F{!S!-'(<.1*<=H##$/'2%-!'   4r$)dNT)r.N)rrrrrrstrr r"boolrrrdictrrrrrrr'r staticmethodrrrrrrr#r$r%rrQs8 26 +(+(9+( +( "%[4/ +(  +( +(Z # T cTXj  49  8$sCx.8T8(2T P9%#%(%t%S8PSUYS4 % % %d38n% %c %  % % E4S> ES ET$sTWx.EY Ed38ndSVX[S[nI]4$c3h$DQUVY[^V^Q_L`$LT#s(^TRUWZRZ^H\r$rrrc Pddtdtdzdtdtdtf fd dtjffd }dtjffd }j s t d jj}|d k(s|d k(r|}n|d k(r|}_ y)z;Compute and display summary metrics for evaluation results.Napiou_thrarea_rngrr.c j}d}|dk(rdnd}|dk(rdnd}|,dj|jd|jd nd j|}t|jD cgc] \} } | |k(s | } } } t|j D cgc] \} } | |k(s | } } } |dk(rLj d }|*tj||jk(d}||}|dddddd| | f}nHj d }|*tj||jk(d}||}|dddd| | f}t||d kDdk(rd n#ttj||d kD}r)tj|j|||||||Scc} } wcc} } w) NrrCrrrrz{:0.2f}:{:0.2f}rrz{:0.2f} precisionrecall)rriouThrsr areaRngLblrevalr+whererZr"meanr~r)r?r@rArpr title_strtype_striou_striaRngaindmDetmindstmean_srrs r% _summarizez1patched_pycocotools_summarize.._summarizes KK[ +-7'8H 1W6&EL_  $ $QYYq\199R= AZcZjZjkrZs "+1<._summarizeDetssa$++*=*=a*@Aaat{{7J7J17MNaa 8K8KA8NOaa'DKKQ>QRS>TUb q7T[[=P=PQR=STb  r$cFtjd}dd|d<ddd|d<ddd|d <ddd |d <ddd |d<dd|d<ddd|d<ddd|d<ddd |d<ddd |d<|S)NrrCrrYrr()rr@rZrJr)rrArKrrrrrrr)r+r])r)rVs r% _summarizeKpsz4patched_pycocotools_summarize.._summarizeKpssa"-aa"c:aa"d;aa"x@aa"w?aa"-aa"c:aa"d;aa"x@aa"w?a r$zPlease run accumulate() firstrrr?)rCNrqr9) r r"r:r+r,rG Exceptionrrr))rrr^r`rrrVs`` @r%rrss_bmr8BJJ 2::  99788{{""H6X/" [ ! DJr$)7rrr4r dataclassesrtypingrfaster_coco_eval.core.maskcorer2r,r"r+faster_coco_evalr%faster_coco_eval.core.faster_eval_apirrfdetr.utilities.distributedrrfdetr.utilities.loggerrr~rlrmrtrwrrr r!rr'r;rUrar<rXrr"ryr{rurrr:rr,rrrr;rrrr#r$r%rjs !..!:3. BJJ &jj), - 2&*"/2u"CH4 $d#,,$,    NTNdN( 4 C$J c3h.b$bT%[SWEWb\`af\gjn\nb,,,et+, %[4,.    et+ !*$d3iD4  tDcN/C   T"**%5S bjj(YrzzYdY( H H4 HD Hii` JNDD$DRVDr$