^jRddZddlZddlmcmZddlmZddlmZGddejZ y)zGPost-processing module for converting model outputs to COCO API format.N)nn)box_opsceZdZdZ d,dedeedzdeddffd Zejde e ejfd ejdee e ejffd Z ed ejd ejdzd ejdzd ejddf dZd ejdeejejejffdZedejdejd ejdejfdZed ejdejdejdejdejd ejdee e ejffdZd ejdejdejdejdejd ejdee e ejffdZedejdejdejfdZedejdedeejejffdZdejdejdejd ejd!ejd"ejd#ededeejejejffd$Zdejd%ejd&ejd'ejdejf d(Zed)ejdejfd*Zedejdejdejdee e ejffd+ZxZS)- PostProcessabConvert raw RF-DETR model outputs into per-image prediction tensors. The postprocessor is shared by detection, segmentation, and keypoint inference. It selects top scoring query/class pairs, scales boxes back to the requested image sizes, and then delegates to the head-specific private helper for masks, keypoints, or box-only results. N num_selectnum_keypoints_per_class trace_alphareturncTt|||_|xsg|_||_y)N)super__init__rrr )selfrrr __class__s d/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/rfdetr/models/postprocess.pyr zPostProcess.__init__s, $'>'D"$&outputs target_sizesc\|d|d}}|jdd}|jdd}|j|||||j|\}}} |j|| |} ||j |||| | |S||j |||| | |S|j ||| S)a5Convert raw model tensors into per-image detection dictionaries. Args: outputs: Model output dictionary containing ``pred_logits`` and ``pred_boxes`` plus optional ``pred_masks`` or ``pred_keypoints``. target_sizes: Per-image ``(height, width)`` tensor. For inference and evaluation this should be the original image size so normalized boxes and keypoints are returned in source-image pixel coordinates. Returns: One dictionary per image. Every dictionary contains ``scores``, ``labels``, and ``boxes`` in absolute pixel coordinates clamped to the respective image dimensions. Segmentation outputs also contain ``masks``. Keypoint outputs also contain ``keypoints`` and ``keypoint_precision_cholesky``. pred_logits pred_boxes pred_masksNpred_keypoints)get_validate_outputs _select_topk_gather_and_scale_boxes_postprocess_masks_postprocess_keypoints_postprocess_boxes) rrr out_logitsout_bbox out_masks out_keypointsscoreslabels topk_boxesboxess rforwardzPostProcess.forward&s '}5w|7LH KK d3  $4d;  z9m\R%)%6%6z%B" ,,Xz<P  **9ffeZYef f  $..}ffeU_amn n&&vvu==rr r"r#c|| | tdt|t|k(sJ|jddk(sJy)aValidate mutually exclusive output heads and per-image target sizes. Args: out_logits: Classification logits with shape ``(B, Q, C)``. out_masks: Optional mask logits from segmentation models. out_keypoints: Optional keypoint predictions from keypoint models. target_sizes: Per-image ``(height, width)`` tensor with shape ``(B, 2)``. Raises: ValueError: If both masks and keypoints are present in the model outputs at the same time. Mask and keypoint heads are mutually exclusive at inference. AssertionError: If batch dimensions do not match ``target_sizes``. Nz>masks and keypoints cannot be used together in postprocessing.) ValueErrorlenshape)r r"r#rs rrzPostProcess._validate_outputsCsK,  ]%>]^ ^:#l"3333!!!$)))rc2|j}|j|jdd}t|j|jd}t j ||d\}}|}||jdz}||jdz} || |fS)aSelect the highest scoring query/class pairs. Args: out_logits: Classification logits with shape ``(B, Q, C)``. Returns: Tuple containing selected scores, class labels, and query indices. Scores are sigmoid probabilities before any keypoint uncertainty fusion. Labels are class indices and query indices select rows from box/mask/keypoint outputs. rr*dimr+)sigmoidviewr.minrtorchtopk) rr problogits_for_topk num_to_select topk_values topk_indexesr$r&r%s rrzPostProcess._select_topk^s!!#))J$4$4Q$7<DOO_-B-B1-EF $)JJ ST$U! \!Z%5%5a%88  0 0 33vz))rr!r&c tj|}tj|d|j dj ddd}|j d\}}tj||||gdj|j}||dddddfz}|jdj|dddddf}|S)aGather selected boxes and scale normalized coordinates to pixels. Args: out_bbox: Normalized ``cxcywh`` boxes with shape ``(B, Q, 4)``. topk_boxes: Query indices selected by :meth:`_select_topk`. target_sizes: Per-image ``(height, width)`` tensor. Returns: Absolute ``xyxy`` boxes with shape ``(B, K, 4)`` in pixel units, clamped to ``[0, width]`` for x-coordinates and ``[0, height]`` for y-coordinates. r*r0r1N)max) rbox_cxcywh_to_xyxyr6gather unsqueezerepeatunbindstacktodtype clamp_minclamp)r!r&rr'img_himg_w scale_fcts rrz#PostProcess._gather_and_scale_boxesss"**84 UAz';';B'?'F'Fq!Q'OP#**1- uKKue 0``, scores for valid keypoint classes are multiplied by an uncertainty penalty derived from the active keypoints of the predicted class. r)default) keypoints_ilabels_iscores_i target_sizeoutput_keypointsoutput_keypoint_precisionnum_keypoint_classesmax_num_keypoints)r$r%r' keypointskeypoint_precision_cholesky) r@rr-rVr._gather_keypoints_for_queries_empty_keypoint_outputs_decode_keypoints_for_imager[)rr#r$r%r'r&rr\rlrkr]rfboxes_irgrerirjs rrz"PostProcess._postprocess_keypointss8 < :V:V!; 7 7$a',=,AHLHhHh +%% ,Q%5.G)=&7Ii IE*,E NN&&$!13L  + <rout_keypoints_i query_indicesc tj|d|jdjdjd|jd|jdS)ajGather keypoint predictions for the selected query rows of one image. Args: out_keypoints_i: Keypoint predictions for one image with shape ``(Q, C * max(K_c), D)``. query_indices: Top-k query indices for that image. Returns: Keypoint predictions aligned with the selected detections. rr0r*rP)r6rBrCrDr.)rsrts rroz)PostProcess._gather_keypoints_for_queriess\||   # #B ' 1 1" 5 < ) r.r,r4anynonzeror _apply_keypoint_trace_fusionrVrr3)rrerfrgrhrirjrkrltotal_padded_keypoint_slotsreshapedvalid_class_mask valid_indicesselected_labelsselected_keypointsrKrL has_precision class_idx class_masknum_active_keypointsout_idxactive_keypointss rrqz'PostProcess._decode_keypoints_for_images[B'2&7&7&:# &*>AR*R R56Q5RS))=(>>UVgUhhjl ##   a "68I;K\K\]_K` $&::##%-/HH H(00$0?B "=1%m_&DE   a $6$<$>##'#?#? #J #q(#J/G1*>S?S>S2ST BRSYBZ]bBb W&;';&;Q> ?BRSYBZ]bBb W&;';&;Q> ?BRSYBZBbBbBd W&;';&;Q> ?L\]`bcdebe]eLf)'3H4H3H*HI g )+DDDrrrrct|j}|j|jd}t |D]J}||k(}|j s|j|} | dkr.||d| f} |j | ||<L|j}||tj|j |zz||<|S)aFuse object scores with keypoint localization uncertainty. Args: scores_i: Object scores for one image before keypoint uncertainty fusion. valid_indices: Detection indices whose predicted class has a keypoint schema slot. selected_labels: Class labels corresponding to ``valid_indices``. selected_keypoints: Class-selected keypoint predictions for those detections. Returns: A score tensor where valid keypoint detections are multiplied by ``exp(-trace_alpha * log_mean_trace)``. The trace is the findability-weighted mean expected squared localization error implied by the predicted precision-Cholesky parameters. rN) r-rryr.rVr_keypoint_log_mean_tracecloner6expr ) rrgrrrrklog_mean_tracesrrrrs rrz(PostProcess._apply_keypoint_trace_fusiongs0 #4#?#?@,667L7LQ7OP 34 ZI(I5J>>##'#?#? #J #q(1*>S?S>S2ST *.*G*GHX*YOJ ' Z>>#"*="9EIItGWGWFWZiFi).).rgg@g-q=)r5r0r1)r3r6logabsrJ logsumexprF) rlog_l11l21log_l22w_findlog_t1log_t2log_t3log_trace_sigma log_w_finds rrz$PostProcess._keypoint_log_mean_traces(#6*v&"6*!&)113uyyU!;<tC$56>ell>W[\`adfkfrfrar\sWt>>8*LL*<<$&*||d**ll *  **4*u||*ellELLZ_ZfZf>f8g**,,LLll  2&<<& & &|| & LL & ll & d3 $% &&&P=||= = = || = LL =ll= d3 $% &=~ u|| TYT`T` ejeqeq  ";\\;; u||U\\) *;;(IE\\IE,, IE ,, IE \\ IE ,,IE$)<<IE"IEIE u||U\\5<<7 8IEV.,,.||. . "LL .  .`k5<<kELLkk<v v v||v d3 $% & vvrr) rr6torch.nn.functionalr functionalrXrfdetr.utilitiesrModulerrrrs.N $sv"))svr