:Qg[y dZddlZddlmZddlmZddlZddlZejgdZ dZ ejdddZ eGd d ZeGd d ZGd dZedk(rddlmZededgZededgZeeeD]f\ZZej1eeZej5\ZZedej<deeedej<deehyy)zZ Implements object detection metrics: average precision, precision, recall, and f1 score. N) dataclass)Path) g?g?g333333?g?gffffff?g?g?g333333?g?gffffff?g?g)\(?g{Gz?c:eZdZUdZeed<eed<eed<eed<y)#ObjectDetectionAggregatedEvaluationzGClass representing a gathered class-aggregated object detection metricsf1_score precisionrecallm_apN)__name__ __module__ __qualname____doc__float__annotations__b/var/www/html/answerous/venv/lib/python3.12/site-packages/unstructured/metrics/object_detection.pyrrsQO M KrrcreZdZUdZeeefed<eeefed<eeefed<eeefed<edZ y)!ObjectDetectionPerClassEvaluationz@Class representing a gathered object detection metrics per-classrrr r ctt|Dcic] }|||| }}tt|Dcic] }|||| }}tt|Dcic] }|||| }}tt|Dcic] }|||| }}|||||Scc}wcc}wcc}wcc}w)N)rangelen) clsaprr f1 class_labelsirr s r from_tensorsz.ObjectDetectionPerClassEvaluation.from_tensors&s49#l:K4LMqLORU*MMz>ObjectDetectionEvalProcessor.from_json_files..qs AhKr)keyc |dSr6rr8s rr:z>ObjectDetectionEvalProcessor.from_json_files..rs Qx[rr7zPage numbers in predictions z (z) and ground truth z) do not match.sizezPage sizes in predictions rz x z) do not match for page .T) prediction) openjsonloadsortedrzipname _process_data_parse_page_dimensions)rr/r0fpredictions_dataground_truth_data pred_pagegt_pagepage_numrr$r%r&r's rfrom_json_filesz,ObjectDetectionEvalProcessor.from_json_filesUs & ' ,1#yy|  , ( ) -Q $ !  -&'ABCv 8 9H   C B C #G,- g &2   H G H #& #G,2G H $W-3H I#   IwX&'(*;; ./C/H/H.IJh'((;q-A,BC$$:$?$?#@76?STCUBVVY6?1%&&>xjK 7 &((BC **+;\VZ*[,,-> M$'$>$>?O$P! k>#3\;P\]]G , , - -sF*F7*F47GcFg}t|j|j|j|jD]-\}}}}|j ||||}|j |/d\}}} } t|j} tj| tj} tj| tj} tj| tj}tj| tj}t|r"tt|Dcgc]}tj|d}}|j|\}}}}}|j!|j!|j!} } }|j!}|j!d}|j!d}|j!d}|j!d}t#|D]I\}}t%||| |<t%||| |<t%||||<t%||||<Kt&j)| | |||j}t+t%| t%|t%| t%|}||fScc}w)zGet per document OD metrics. Returns: tuple: Tuple of ObjectDetectionAggregatedEvaluation and ObjectDetectionPerClassEvaluation )predstargetsheightwidth)rUrUrUrr>)rrr rr)rrr r )rEr$r%r&r' _compute_page_detection_matchingappendrrnpfullnanlisttorchcat_compute_detection_metricsmean enumeraterrrr)r+document_matchingsrQrRrSrTpage_matching_tensorsmean_apmean_precision mean_recallmean_f1num_clsmean_ap_per_classmean_precision_per_classmean_recall_per_classmean_f1_per_classxmatching_info_tensorsap_per_present_classesprecision_per_present_classesrecall_per_present_classesf1_per_present_classespresent_classes ap_per_classprecision_per_classrecall_per_class f1_per_classr class_indexod_per_class_evaluation od_evaluations r get_metricsz(ObjectDetectionEvalProcessor.get_metricss -0   !6!68I8I4K[K[. = )E7FE%)$I$I %J% !  % %&; < =9 5gd''(GGGRVV4#%777BFF#; " 8GGGRVV4 ! ">B3HZC[>\$]UYYq!_$] !$]0//&  &-*&.224*//1&++-*1KN-113G266q9L"?"D"DQ"G 9>>qA 166q9L"+O"< H;16|A1G!+.8=>QRS>T8U(55:;KA;N5O%k216|A1G!+.  H#D"P"P .( ** #Q# <7^N+%w  555k%^s4Jdatacg}g}|dD]0}|j|dd|j|dd2||fS)zX Process the page dimensions from the json file to the required format. r4r=rSrT)rW)r{r&r'pages rrHz3ObjectDetectionEvalProcessor._parse_page_dimensionss\   M 6D   V X 6 7   tF|G4 5 6[((rr@c ,g}|dD]}g}|dD]W}|d}|j|}|d\} } } } |r|d} |j| | | | | |gB|j|| | | | gYtj|}|j||S)zQ Process the elements from the json file to the required format. r4elementstypebboxprob)indexrWr\tensor)r{rr@ pages_listr} page_elementselement class_label class_idxx1y1x2y2 confidence page_tensors rrGz*ObjectDetectionEvalProcessor._process_datas  M +DM + F%fo (..{; !(BB!(J!(("b"b*i)PQ!(()RR)DE F ,,}5K   k * +r preds_scores preds_clstop_kctj|}|jddtj|dz|jjddk(}|jdd|z}|j dd\}}|d|ddfj d }||jdd } | jdS) a Get the indexes of all the top k predictions for every class Args: preds_scores: The confidence scores, vector of shape (n_pred) preds_cls: The predicted class, vector of shape (n_pred) top_k: Number of predictions to keep per class, ordered by confidence score Returns: top_k_idx: Indexes of the top k predictions. length <= (k * n_unique_class) r>r(rT descendingNFas_tupledim)r\maxviewaranger(sortnonzerosplit) rrr n_unique_clsmaskpreds_scores_per_clssorted_scores_per_cls sorting_idxidx_with_satisfying_scores top_k_idxs r_get_top_k_idx_per_clsz3ObjectDetectionEvalProcessor._get_top_k_idx_per_cls s yy+ ~~b!$ 1 \%8%8) $q"+ ,00Q7$>-A-F-FqUY-F-Z*{%:6E619%E%M%MW\%M%]" : @ @ @ JK ~~b!!rboxes img_shapec|dddgfjd|d|dddgf<|dddgfjd|d|dddgf<|S)a* Clips bboxes to image boundaries. Args: bboxes: Input bounding boxes in XYXY format of [..., 4] shape img_shape: Image shape (height, width). Returns: clipped_boxes: Clipped bboxes in XYXY format of [..., 4] shape .rr>)minr)clip)rrs r"_change_bbox_bounds_for_image_sizez?ObjectDetectionEvalProcessor._change_bbox_bounds_for_image_size$so#3A;/44 ! 4McAq6k"3A;/44 ! 4McAq6k rbox1box2cVd}||j}||j}tj|dddddf|ddddftj|dddddf|ddddfz j dj d}||dddf|z|z z S)a Return intersection-over-union (Jaccard index) of boxes. Both sets of boxes are expected to be in (x1, y1, x2, y2) format. Args: box1: Tensor of shape [N, 4] box2: Tensor of shape [M, 4] Returns: iou: Tensor of shape [N, M]: the NxM matrix containing the pairwise IoU values for every element in boxes1 and boxes2 c0|d|dz |d|dz zS)Nrrrr>r)boxs rbox_areaz7ObjectDetectionEvalProcessor._box_iou..box_areaFs%FSVOAQ8 8rNrr)Tr\rrclampprod)rrrarea1area2inters r_box_iouz%ObjectDetectionEvalProcessor._box_iou6s  9  YYtAtQRK($q!"u+ 644QSRSQS CTVZ[\^`_`^`[`Va9b b U1X T!W  ag.677rpreds_box_xyxytargets_box_xyxy targets_cls preds_matchedtargets_matchedpreds_idx_to_useiou_thresholdsc |j|||} ||jdd|jddk7} d| | <| jdd\} } | |dkDjdD]x\} }|| }| | |f}| | |f|kD}t j ||ddf||ddf}t j ||}d|||f<d|||f<|j sw|S|S) a Computes the matching targets based on IoU for regular scenarios. Args: preds_box_xyxy: (torch.Tensor) Predicted bounding boxes in XYXY format. preds_cls: (torch.Tensor) Predicted classes. targets_box_xyxy: (torch.Tensor) Target bounding boxes in XYXY format. targets_cls: (torch.Tensor) Target classes. preds_matched: (torch.Tensor) Tensor indicating which predictions are matched. targets_matched: (torch.Tensor) Tensor indicating which targets are matched. preds_idx_to_use: (torch.Tensor) Indices of predictions to use. Returns: targets: Computed matching targets. rr>rT)rstableFrN)rrrrr\ logical_andall)r+rrrrrrrriou cls_mismatch sorted_iou target_sortedpred_selected_itarget_sorted_ipred_itarget_iis_iou_above_thresholdare_candidates_freeare_candidates_goods r_compute_targetsz-ObjectDetectionEvalProcessor._compute_targetsUs\8mmN+;<>NO !!1277A>+BRBRSTVXBYY L %(HHTH$J! M2O1O0X0X1Y1   ,O_ &o6F$_o%EFH&00P%QTb%b "#("3"3vqy))OHaK,H+H#  #("3"34JL_"`  >BOH&99 :9=M&"55 6""$= <rrQrRrSrT return_on_cpuc |jj|j}t|}|t|dk(rt j d|ftj |j} t j d|ftj |j} t jgtj|j} t jgtj|j} |dddfj|j} | | | | | fSt j t||tj |j} t j t||tj |j}t j t||tj |j} |dddf|ddddf|dddf} }} |dddf|ddddf}} |j| | |}d | ddddf<d | |<t|dkDr,|j|||f|j|| || | |||} | | | | | fS) aC Match predictions (NMS output) and the targets (ground truth) with respect to metric and confidence score for a given image. Args: preds: Tensor of shape (num_img_predictions, 6) format: (x1, y1, x2, y2, confidence, class_label) where x1,y1,x2,y2 are according to image size targets: targets for this image of shape (num_img_targets, 5) format: (label, x1, y1, x2, y2) where x1,y1,x2,y2 are according to image size height: dimensions of the image width: dimensions of the image top_k: Number of predictions to keep per class, ordered by confidence score return_on_cpu: If True, the output will be returned on "CPU", otherwise it will be returned on "device" Returns: preds_matched: Tensor of shape (num_img_predictions, n_thresholds) True when prediction (i) is matched with a target with respect to the (j)th threshold preds_to_ignore: Tensor of shape (num_img_predictions, n_thresholds) True when prediction (i) is matched with a crowd target with respect to the (j)th threshold preds_scores: Tensor of shape (num_img_predictions), confidence score for every prediction preds_cls: Tensor of shape (num_img_predictions), predicted class for every prediction targets_cls: Tensor of shape (num_img_targets), ground truth class for every target rNr)dtyper(rr>TF) rr*r(rr\zerosboolrfloat32rrr)r+rQrRrSrTrr thresholdsnum_thresholdsrpreds_to_ignorerrrr preds_box targets_boxrs rrVz=ObjectDetectionEvalProcessor._compute_page_detection_matchings&R((++4;;+? Z =CJ!O!KKN(;5::VZVaVabM#kk1n*=UZZX\XcXcdO <<%-- TL Ru}}T[[QI!!Q$-**$++*>K /<KW W Jejj  ++ L. 4;;  ++ Jejj .31b5\5AaC=%PQSTPT+l9 #*1a4='!QqS&/[  66|YPUV $1,1() w*7)9)9$++)FHZHZ KKI   OODKK ( LL % NN4;; '"-i !2255dkkB..k2779">2M4G4G4K [[';/ DKK+ 6t{{K g{3DKKH"+N"; * E;.7;.F{*+M77+M: / >)-8)--/"3 / 8  "BuaxL"/IeQh )F5!8 % *(]V #y6'9E'A B9fb.88rrrrc|jd}|}tjtj|tj|} t |dk(rbtj ||j tj ||j tj ||j fS|jr,|jtjurtjn |j} tj|j| d} || ddf}| | ddf} || j}tj|dtj} tj| dtj} | |z }| | | ztj tj"j$zz }|j'dj)dj*j'd}tj,| | d}|dk(rCtj ||j }tj ||j }n||d z }||d z }|j/d dj1|d }tj,|j2j|d j2}tj4|tj d ||j fd }tj6||d }|j9d}|||fS) a Compute the list of precision, recall and MaP of a given class for every recall threshold. Args: preds_matched: Tensor of shape (num_predictions, n_thresholds) True when prediction (i) is matched with a target with respect to the(j)th threshold preds_to_ignore Tensor of shape (num_predictions, n_thresholds) True when prediction (i) is matched with a crowd target with respect to the (j)th threshold preds_scores: Tensor of shape (num_predictions), confidence score for every prediction n_targets: Number of target boxes of this class recall_thresholds: Tensor of shape (max_n_rec_thresh) list of recall thresholds used to compute MaP score_threshold: Minimum confidence score to consider a prediction for the computation of precision and recall (not MaP) Returns: ap, precision, recall: Tensors of shape (nb_thrs) rrrTrN)axisr)rightr>Fr)inputrr)rr\r logical_notrrr(is_cudarruint8argsortr* contiguouscumsumrfinfofloat64epsflipcummaxvalues searchsortedrrepeatrr]gatherr_)r+rrrrrrrtpsfpsrsort_ind rolling_tps rolling_fpsrolling_recallsrolling_precisionslowest_score_above_thresholdr rrecall_threshold_idxsampled_precision_pointsrs rrz?ObjectDetectionEvalProcessor._compute_detection_metrics_per_cls@s@$))"-    m ,e.?.?.P  s8q= K < K < K < ## (:(:ejj(H KK##  ==!7DI(A+(A+#H-88: ll3QekkB ll3QekkB % 1( + % EMM(B(F(F F  044Q7>>qAHHMMaP(-'9'9 MO+4( $ )A -[[T[[AF DKKI%%AA%EFF*+G!+KLI .221b9@@aP %11    ( ( *,=U !  #YY Q DKK!P QWX  $)<<$,@a$  & * *1 -9f$$rN)cpu)F)dT)%r r r IOU_THRESHOLDSrSCORE_THRESHOLDrRECALL_THRESHOLDSrr[r\Tensorintr r.r!rrOtuplerrrz staticmethodrrHrrGrrXndarrayrrrrVr^rrrrrr#r#0sx#N%O))U\\*)u||,)3i ) #Y ) 3i ))@2^"2^!%2^ ( 2^2^hZ6 24UU VZ6x )T )eD$J.? ) )DDTRVZ,"ll"/4||"DG" ""4zz&+CHo "8u||85<<8ELL88r(sh ! T ELLD$/   66 6$I %I %X z"""<=tD_?`a () )* 9<69 d446EE "8 &4%?%?%A"" 3889VG_