khHC jdZ ddlmZmZmZddlZddlZddlm Z ddl m Z m Z ddl mZmZmZmZmZddlmZdZd Zdd efd Zd ej2d ej2fdZ ddeej2deej2dedefdZdeeefdefdZGdde j<ZGdde Z!y)z RetinaNet / EfficientDet Anchor Gen Adapted for PyTorch from Tensorflow impl at https://github.com/google/automl/blob/6f6694cec1a48cdb33d5d1551a2d5db8ad227798/efficientdet/anchors.py Hacked together by Ross Wightman, original copyright below )OptionalTupleSequenceN) batched_nmsremove_small_boxes) ArgMaxMatcherFasterRcnnBoxCoderBoxList IouSimilarityTargetAssigner)batched_soft_nmsggj output_xyxyc |dddf|dddfzdz }|dddf|dddfzdz }|dddf|dddfz }|dddf|dddfz }|jd\}}} } tj| |z} tj| |z} ||z|z} ||z|z}| | dz z }|| dz z }| | dz z}|| dz z}|rtj||||gd}|Stj||||gd}|S)aTransforms relative regression coordinates to absolute positions. Network predictions are normalized and relative to a given anchor; this reverses the transformation and outputs absolute coordinates for the input image. Args: rel_codes: box regression targets. anchors: anchors on all feature levels. Returns: outputs: bounding boxes. Nrr dim@)unbindtorchexpstack) rel_codesanchorsr ycenter_a xcenter_ahawatytxthtwwhycenterxcenteryminxminymaxxmaxouts J/var/www/teggl/fontify/venv/lib/python3.12/site-packages/effdet/anchors.pydecode_box_outputsr.-sLAA.!3IAA.!3I AA &B AA &B%%!%,NBB " A " A2g !G2g !G QV D QV D QV D QV Dkk4tT2: Jkk4tT2: Jboxessizec~|jd}tj||gd}|j|}|S)Nr)minr)clamprcatr3)r0r1s r-clip_boxes_xyxyr6Rs8 KKAK E 99dD\q )D IIdOE Lr/ img_scaleimg_sizemax_det_per_imagesoft_nmsc |jddk(sJ|jddk(sJ|jddk(sJ||ddf}t|j|d} ||t| ||z } |j j dj} |rt | | |ddd \} } | | | <nt| | |d } | d|} | | } | | df} || dfdz}|| |z} t| } tj| | |jgd }| |krHtj|tj|| z d f|j|jgd }|S)axGenerates detections with RetinaNet model outputs and anchors. Args: cls_outputs: a torch tensor with shape [N, 1], which has the highest class scores on all feature levels. The N is the number of selected top-K total anchors on all levels. box_outputs: a torch tensor with shape [N, 4], which stacks box regression outputs on all feature levels. The N is the number of selected top-k total anchors on all levels. anchor_boxes: a torch tensor with shape [N, 4], which stacks anchors on all feature levels. The N is the number of selected top-k total anchors on all levels. indices: a torch tensor with shape [N], which is the indices from top-k selection. classes: a torch tensor with shape [N], which represents the class prediction on all selected anchors from top-k selection. img_scale: a float tensor representing the scale between original image and input image for the detector. It is used to rescale detections for evaluating with the original groundtruth annotations. max_det_per_image: an int constant, added as argument to make torchscript happy Returns: detections: detection results in a tensor with shape [max_det_per_image, 6], each row representing [x_min, y_min, x_max, y_max, score, class] r NT)rg333333?gMbP?)method_gaussian iou_thresholdscore_threshold?)r?r)devicedtyper)shaper.floatr6sigmoidsqueezerrlenrr5zerosrCrD) cls_outputs box_outputs anchor_boxesindicesclassesr7r8r9r:r0scorestop_detection_idx soft_scoresnum_det detectionss r-generate_detectionsrUYsB   R A %% %   b !Q && &   R A %% % +L {002Ld SE!5x)';<  " * *1 - 3 3 5F)9 67D]a*c&;$/ !'vwcR**<+<= # $E %t+ ,F'-.2G ! #$GE67==?;CJ""YY  KK*W4a8ARARZdZjZj k   r/ image_size max_levelc|}|g}td|dzD]/}|ddz dzdz|ddz dzdzf}|j|1|S)zGet feat widths and heights for all levels. Args: image_size: a tuple (H, W) max_level: maximum feature level. Returns: feat_sizes: a list of tuples (height, width) for each level. r rr)rangeappend)rVrW feat_size feat_sizes_s r-get_feat_sizesr^spIJ 1i!m $%lQ&1,q09Q>4>>A+=> FE$&N5 ! %doo 6 F "00FF"5)00$Q-*U:e3DQ3G-HH$Q-*U:e3DQ3G-HHJ%doo(>>++EDNN,BCEFF F Fr/c g}|jjD]\}}g}|D]]}|\}}}} | |dzd|zz} | |dzd|zz} t|tr|\} } nt j |} d| z } | | zdz }| | zdz }t j |ddz |jd|d}t j |ddz |jd|d}t j||\}}|jd}|jd}t j||z ||z ||z||zf}t j|dd}|jt j|d`t j|d}|j|jddgt j|}tj |j#}|S) z#Generates multi-scale anchor boxes.r rrg?rr<)axisr=)rkitemsrgrnpsqrtarangerVmeshgridreshapevstackswapaxesrZ expand_dims concatenater from_numpyrF)rn boxes_allr]configs boxes_levelrkstride octave_scaleryrobase_anchor_size_xbase_anchor_size_yaspect_xaspect_yanchor_size_x_2anchor_size_y_2xyxvyvr0rMs r-rmzAnchors._generate_boxess ++++- ;JAwK! B=C: fl%1F1I%=\@Q%Q"%1F1I%=\@Q%Q"fh/)/&Hh!wwvH"X~H"4x"?#"E"4x"?#"EIIfQi!mT__Q-?KIIfQi!mT__Q-?KQ*BZZ^ZZ^ 2#7o9M#%#7o9M#OP E1a0""2>>%a#@A+ B0..1=K   [00"a9 :7 ;:yy+ '' 5;;= r/cF|jt|jzSrr)rerIrf)rns r-get_anchors_per_locationz Anchors.get_anchors_per_locationsT%7%7!888r/) __name__ __module__ __qualname____doc__rintrc classmethodrtrjrmr __classcell__)rps@r-r`r`sD"%>bghkmphpbq%>N44 "H9r/r`c2eZdZdZddedefdZd dZd dZy) AnchorLabelerz)Labeler for multiscale anchor boxes. num_classesmatch_thresholdct}t||dd}t}t||||_||_||_||_i|_y)afConstructs anchor labeler to assign labels to anchors. Args: anchors: an instance of class Anchors. num_classes: integer number representing number of classes in the dataset. match_threshold: float number between 0 and 1 representing the threshold to assign positive labels for anchors. T)unmatched_thresholdnegatives_lower_than_unmatchedforce_match_for_each_rowN) r rr r target_assignerrrr indices_cache)rnrrrsimilarity_calcmatcher box_coders r-rczAnchorLabeler.__init__!s[(/  /+/%) + '( -ow R .&r/cg}g}|r|dkD}||}||}|jjt|jjt||\}}} |dz j } d} t |jj|jjdzD]} |jj| } | d| dz|jjz} |j|| | | zj| d| ddg|j|| | | zj| d| ddg| | z } | jdkDjj}|||fS)aHLabels anchors with ground truth inputs. Args: gt_boxes: A float tensor with shape [N, 4] representing groundtruth boxes. For each row, it stores [y0, x0, y1, x1] for four corners of a box. gt_classes: A integer tensor with shape [N, 1] representing groundtruth classes. filter_valid: Filter out any boxes w/ gt class <= -1 before assigning Returns: cls_targets_dict: ordered dictionary with keys [min_level, min_level+1, ..., max_level]. The values are tensor with shape [height_l, width_l, num_anchors]. The height_l and width_l represent the dimension of class logits at l-th level. box_targets_dict: ordered dictionary with keys [min_level, min_level+1, ..., max_level]. The values are tensor with shape [height_l, width_l, num_anchors * 4]. The height_l and width_l represent the dimension of bounding box regression output at l-th level. num_positives: scalar tensor storing number of positives in an image. r<r r)rassignr rr0longrYrdrWr\rrZr match_resultsrFsum)rngt_boxes gt_classes filter_validcls_targets_outbox_targets_out valid_idx cls_targets box_targetsmatchescountrwr[steps num_positivess r- label_anchorszAnchorLabeler.label_anchors:s, "RI *H#I.J,0,@,@,G,G DLL&& '):J-H) ['#Q,,.  @4<<114<<3I3IA3MN E //6IaL9Q</$,,2W2W2YYE  " ";uUU]#C#K#KYWX\[def[gikLl#m n  " ";uUU]#C#K#KYWX\[def[gikLl#m n UNE  !..3::<@@B >>r/c zt|}|t|k(sJ|jj|jjz dz}t |Dcgc]}g}}t |Dcgc]}g}}g} t |jj } t |D]} | |dz k(} |r"|| dkD} t || | }|| | }nt || }|| }|jj| ||\}}}|dz j} d}t |jj|jjdzD]}||jjz }|jj|}|d|dz|jjz}||j||||zj|d|ddg||j||||zj|d|ddg||z }| stj||||<tj||||<| j|j dkDj#j%| stj| } ||| fScc}wcc}w)Nr r<r)rIrrWrdrYr r0rrrr\rrZrrrrrFr)rnrrr batch_size num_levelsr]rrnum_positives_outanchor_box_listi last_sampler gt_box_list gt_class_irrrrrw level_idxr[rs r-batch_label_anchorsz!AnchorLabeler.batch_label_anchorsls] S_,,,\\++dll.D.DDqH ',Z'89!299',Z'89!299!$,,"4"45z"! CAzA~-K&qMB. %hqk)&<= ']95 %hqk2 '] 040D0D0K0KO]hjt0u -Kg'?002K DEt||55t||7M7MPQ7QR Y!DLL$:$::  LL33E: !! y|3dll6[6[6]] *11eem4<rs* -, Amm&""J5<<u||8=IELL)I5=ell5KII15IX uS#X 3  e9biie9P{CF{Cr/