+ jVRt^RIt^RIt^RIHtHt^RIt^RIt^RI H t ^RI H t ^RIHt^RIHtHt^RIHtHtHtHtHtHt^RIHtHt^R IHtHtH t H!t!H"t"] !4'd^R I#H$t$^R I%H&t&H't']'d^R I(H)t)]!PT!]+4t,!R R]RR7t-R RRllt.RRlt/Rt0RRlt1R!RRllt2RRlt3RRlt4]!RR] 44t5R.t6R#)"z"Image processor class for VitPose.N) TYPE_CHECKINGUnion) functional)TorchvisionBackend) BatchFeature)group_images_by_shapereorder_images)IMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STDChannelDimension ImageInputPILImageResamplingSizeDict) ImagesKwargsUnpack) TensorTypeauto_docstringis_scipy_availableloggingrequires_backends)inv)affine_transformgaussian_filter)VitPoseEstimatorOutputc0a]tRt^7toRtV3RltRtVtR#)VitPoseImageProcessorKwargsa5 do_affine_transform (`bool`, *optional*): Whether to apply an affine transformation to the input images based on the bounding boxes. normalize_factor (`float`, *optional*, defaults to `200.0`): Width and height scale factor used for normalization when computing center and scale from bounding boxes. cN<V^8dQh/S[R,;R&S[R,;R&#)Ndo_affine_transformnormalize_factor)boolfloat)format __classdict__s"u/home/luca/screen-checking/.venv/lib/python3.14/site-packages/transformers/models/vitpose/image_processing_vitpose.py __annotate__(VitPoseImageProcessorKwargs.__annotate__7s'$dl"N)__name__ __module__ __qualname____firstlineno____doc____annotate_func____static_attributes____classdictcell__)r#s@r$rr7sr'rF)totalc V^8dQhR\\,\P,R\R\R\ R\ /#)rbox image_width image_heightrpadding_factor)tuplelistnpndarrayintr!)r"s"r$r%r%DsG))  "))) )  )r'cVR,wrVrxW, p \P!WWR,,WhR,,.\PR7p WyV,8dVR,V , pMWyV,8d W,p\P!Ws, W, .\PR7p W,p W3#)aa Encodes a bounding box in COCO format into (center, scale). Args: box (`Tuple`, `List`, or `np.ndarray`): Bounding box in COCO format (top_left_x, top_left_y, width, height). image_width (`int`): Image width. image_height (`int`): Image height. normalize_factor (`float`): Width and height scale factor. padding_factor (`float`): Bounding box padding factor. Returns: tuple: A tuple containing center and scale. - `np.ndarray` [float32](2,): Center of the bbox (x, y). - `np.ndarray` [float32](2,): Scale of the bbox width & height. :NN?dtype?)r9arrayfloat32) r3r4r5rr6 top_left_x top_left_ywidthheight aspect_ratiocenterscales &&&&& r$box_to_center_and_scalerKDs:-0G)JE-L XXzCK/sl1JKSUS]S] ^F f$$|+ & &% HHe.0IJRTR\R\ ]E  "E =r'cV^8dQhR\R\PR\PR\P/#)rtheta size_inputsize_dst size_target)r!r9r:)r"s"r$r%r%ps7""5"bjj"BJJ"]_]g]g"r'c\P!V4p\P!R\PR7pV^,V^,, pV^,V^,, p\P !V4V,VR&\P !V4)V,VR&VRV^,,\P !V4,RV^,,\P !V4,,RV^,,,,VR&\P !V4V,VR&\P !V4V,VR &VRV^,,\P !V4,RV^,,\P !V4,, RV^,,,,VR &V#) a Calculate the transformation matrix under the constraint of unbiased. Paper ref: Huang et al. The Devil is in the Details: Delving into Unbiased Data Processing for Human Pose Estimation (CVPR 2020). Source: https://github.com/open-mmlab/mmpose/blob/master/mmpose/core/post_processing/post_transforms.py Args: theta (`float`): Rotation angle in degrees. size_input (`np.ndarray`): Size of input image [width, height]. size_dst (`np.ndarray`): Size of output image [width, height]. size_target (`np.ndarray`): Size of ROI in input plane [w, h]. Returns: `np.ndarray`: A matrix for transformation. r?r>)rrTrTg࿩rTrrVrTrVrVrVr)r9deg2radzerosrCmathcossin)rMrNrOrPmatrixscale_xscale_ys&&&& r$get_warp_matrixrcpsX( JJu E XXfBJJ /FqkKN*GqkKN*G88E?W,F4LHHUO#g-F4L z!}txx.z!}1DtxxPU1VVY\_jkl_mYmmF4L88E?W,F4L88E?W,F4L z!}txx.z!}1DtxxPU1VVY\_jkl_mYmmF4L Mr'c \VPR,4Uu.uF q0RV3,NK pp\P!V.RO.4p\ V4pVR,VR,VR,VR ,VR ,VR ,3wVR &VR&VR&VR&VR &VR &VUu.uFp\ WvV^R7NK pp\P !VRR7pV#uupiuupi) aO This function implements cv2.warpAffine function using affine_transform in scipy. See https://docs.scipy.org/doc/scipy/reference/generated/scipy.ndimage.affine_transform.html and https://docs.opencv.org/4.x/d4/d61/tutorial_warp_affine.html for more details. Note: the original implementation of cv2.warpAffine uses cv2.INTER_LINEAR. .) output_shapeorderaxis)rTrTrVrYrXrUrSrZrW)rangeshaper9vstackrrstack) srcMsizeichannelsM_scipyM_invchannelnew_srcs &&& r$scipy_warp_affinerws &+399R=%9:%9CF %9H:iiI'G LE d  d  d  d  d  d TPE$KteDk5;d U4[\dd[cPWTK[cGdhhwR(G N#;es C(C cV^8dQhR\PR\\P\P3,/#)rheatmapsreturn)r9r:r7)r"s"r$r%r%s0rzzeBJJ VPwr4rVVP^,pV^8XgW78Xg \R4h\V^, ^,4p\P!VU U u.uF#p V U u.uFp \ V RW3R R7NK up NK% up p 4p\P !VR^24p\P!V4p\P!VR RR7P4p VR ,^,VR ,^,V^,,,p W^,V^,,\P!^W4,4PRV4,, p V P\4PR^4p W,p W^,,pWV,^,,pWV,^,,pWV, ^, ,pW^, ,pW^, V, ,pRVV, ,pRVV, ,p\P!VV.^R7pVPWt^^4pV^V ,, V,pV^V ,, V,pRVV, V, V ,V ,V, V, V,,p\P!VVVV.^R7pVPWt^^4p\PPV\P !\P"4P$\P&!^4,,4pV\P(!R VV4P+4,pV#uup iuup p i)aDARK post-processing. Implemented by unbiased_data_processing. Paper references: - Huang et al. The Devil is in the Details: Delving into Unbiased Data Processing for Human Pose Estimation (CVPR 2020). - Zhang et al. Distribution-Aware Coordinate Representation for Human Pose Estimation (CVPR 2020). Args: coords (`np.ndarray` of shape `(num_persons, num_keypoints, 2)`): Initial coordinates of human pose. batch_heatmaps (`np.ndarray` of shape `(batch_size, num_keypoints, height, width)`): Batched heatmaps as predicted by the model. A batch_size of 1 is used for the bottom up paradigm where all persons share the same heatmap. A batch_size of `num_persons` is used for the top down paradigm where each person has its own heatmaps. kernel (`int`, *optional*, defaults to 3): Gaussian kernel size (K) for modulation. Returns: `np.ndarray` of shape `(num_persons, num_keypoints, 2)` ): Refined coordinates. zQThe batch size of heatmaps should be 1 or equal to the batch size of coordinates.g?)sigmaradiusaxesgMbP?edge)moder>rgzijmn,ijnk->ijmkrU)rSrSrYrY).rT).rVri)rkrr;r9rBrcliplogpadflattenarangerr concatenatelinalgrfinforCepseyeeinsumsqueeze)rrrrrrGrF num_coordsrryheatmapbatch_heatmaps_padindexi_ix1iy1ix1y1ix1_y1_ix1_iy1_dxdy derivativedxxdyydxyhessians&&& r$"post_dark_unbiased_data_processingrs*0>/C/C,JvaJ !Oz7lmm &1*" #FXX+ *go ofn[b_WC8Hv Vfn o* N WW^UB7NVVN+N0PW]^ffh 6NQ &.1"4!C CE aiFQJ '"))Az7Q*R*Z*Z[]_l*m mmE LL  % %b! ,E  "B QY 'C U]Q. /C u}q0 1E !23G ai (D ai%/ 0D d B d BRq1J##Jq!DJ B, C B, C s"R'",t3d:WD ECnnc3S1:GoojA>GiimmGbhhrzz&:&>&>&JJKG bii)7J?GGIIF M; p s) N 2N N N c V^8dQhR\PR\PR\PR\PR\P/#)rrrIrJ output_sizerzr9r:)r"s"r$r%r% sI--BJJ- -2::-\^\f\f-kmkuku-r'c|VP^,R9d \R4h\V4^8wd \R4h\V4^8wd \R4h\V4^8wd \R4hVR,pV^,V^,R, , pV^,V^,R, , p\P!V4pVR ,V,V^,,V^,R,, VR &VR ,V,V^,,V^,R,, VR &V#) a'Get final keypoint predictions from heatmaps and apply scaling and translation to map them back to the image. Note: num_keypoints: K Args: coords (`np.ndarray` of shape `(num_keypoints, ndims)`): * If ndims=2, corrds are predicted keypoint location. * If ndims=4, corrds are composed of (x, y, scores, tags) * If ndims=5, corrds are composed of (x, y, scores, tags, flipped_tags) center (`np.ndarray` of shape `(2,)`): Center of the bounding box (x, y). scale (`np.ndarray` of shape `(2,)`): Scale of the bounding box wrt original image of width and height. output_size (`np.ndarray` of shape `(2,)`): Size of the destination heatmaps in (height, width) format. Returns: np.ndarray: Predicted coordinates in the images. z5Coordinates need to have either 2, 4 or 5 dimensions.z9Center needs to have 2 elements, one for x and one for y.z,Scale needs to consist of a width and heightz2Output size needs to consist of a height and widthi@rAr>)rr=r|rTr|rV)rkrlenr9 ones_like)rrIrJrrbra target_coordss&&&& r$transform_predsr s 2||Ai'PQQ 6{aTUU 5zQGHH ;1MNN EMEAh+a.3./GAh+a.3./GLL(M ,06!9)*T1 2 2r'c VVPWWER7pW&R&VP!V3/VB#)z"Handle extra inputs beyond images.)rrrrr)_prepare_image_like_inputs _preprocess)rrrrrrrs&&&&&&,r$_preprocess_image_like_inputs3VitPoseImageProcessor._preprocess_image_like_inputsps>00L]1  w1&11r'c Z<V^8dQhRRRS[S[,RS[S[,RS[RS[RR/#)rimage torch.TensorrIrJrotationrprz)r7r!r)r"r#s"r$r%rsOe U|      r'c \WBR,\P!VPVP34R, VR,4pVP ^^^4P 4P4p\WvVPVP3R7p\P!V4P ^^^4PVP4p V #)z7Apply an affine transformation to a torch tensor image.g@rAr)rnrorp) rcr9rBrFrGpermutecpunumpyrwtorch from_numpytor) rrrIrJrrptransformationimage_nptransformed_np transformeds &&&&&& r$r&VitPoseImageProcessor.affine_transforms) slBHHdjj$++-F$G#$MuW\} ==Aq)--/557*xQUQ\Q\^b^h^hPij&&~6>>q!QGJJ5<<X r'c&n<V^8dQhRS[R,RS[RS[RRRS[RS[R S[R S[R S[R S[S[S[,,R ,RS[S[S[,,R ,RS[R ,RS[R ,RS[R ,RS[S[,R ,RS[RS[RS[S[P,R ,RS[/#)rrr do_resizerpresamplez7PILImageResampling | tvF.InterpolationMode | int | Nonedo_center_crop crop_size do_rescalerescale_factor do_normalize image_meanN image_stddo_padpad_sizedisable_groupingreturn_tensorsrrrrz) r8r rr!rrr9r:r)r"r#s"r$r%rs$-a-a^$-a-a -a L -a  -a-a-a-a-aDK'$.-a4;&--at -aT/-a+-a j(4/!-a""#-a$ %-a&bjj 4''-a* +-ar'c VeyV'dq.p\VV4F\wppVFPp\VVPVPVR7wppVP VVV^VR7pVP V4KR K^ Tp\ WR7wpp/pVP4FwppVPVWxWV 4pVVV&K! \VV4p \RV /VR7#)z!Custom preprocessing for VitPose.)r4r5r)rrp)rr)data tensor_type) ziprKrFrGrappendritemsrescale_and_normalizerr)!rrrrprrrrrrrrrrrrrrrrtransformed_imagesr image_boxesr3rIrJtransformed_imagegrouped_imagesgrouped_images_indexprocessed_images_groupedrkstacked_imagesprocessed_imagess!&&&&&&&&&&&&&&&&&&&, r$r!VitPoseImageProcessor._preprocesss.  !4!# &)&%&8"{&C$;$(JJ%)[[)9 %MFE )-(=(=eVU]^ei(=(j%&--.?@''9(F/DV/o,,#% %3%9%9%; !E>!77 LV_N/= $U + &< **BDXY.2B!CQ_``r'cn<V^8dQhRS[PRS[PRS[PRS[/#)rryrIrJrr)r"r#s"r$r%rs?  **   zz   r'c VPwrVrx\V4wr\WVR7p \V4F'p \ W,W,,W<,Wx.R7W&K) W3#)zRGet final keypoint predictions from heatmaps and transform them back to the image.r)rIrJr)rkrrrjr) rryrIrJrrrrGrFrrrrqs &&&&& r$keypoints_from_heatmaps-VitPoseImageProcessor.keypoints_from_heatmapss](0~~$ v1(;26FSz"A&ux `f_noEH#}r'c <V^8dQhRRRS[S[S[S[,,,S[P,RS[RS[R,RS[S[S[,,R,/#)routputsrr kernel_size thresholdN target_sizes)r8r!r9r:r;rr7)r"r#s"r$r%rsmFF)FDe%&3F F 4< F !4;.5 Fr'c  \VR4VPPwrgpVeV\V48wd \ R4h\ P !V^3\ PR7p \ P !V^3\ PR7p \\P!V!4p \V4Fp VeGW\,^,W\,^,r\ P!WW.4pW,V,W&VPR,VPR,pp\W,VVR7wppVWR3&VWR3&K VPVPP!4P#4WVR7wpp\ P !V^3\ PR7pV R ,VR &V R ,VR&\$P&!V4p\$P&!V4p\$P(!^V4p\$P&!\+V44p.p\-VVV4pVFp.pVFcp\/V4wppp VP14pTp!Ve!VV8p"VV",pVV",pV!V",p!R VR VR V!R V /p#VP3V#4Ke VP3V4K V#)a^ Transform the heatmaps into keypoint predictions and transform them back to the image. Args: outputs (`VitPoseEstimatorOutput`): VitPoseForPoseEstimation model outputs. boxes (`list[list[list[float]]]` or `np.ndarray`): List or array of bounding boxes for each image. Each box should be a list of 4 floats representing the bounding box coordinates in COCO format (top_left_x, top_left_y, width, height). kernel_size (`int`, *optional*, defaults to 11): Gaussian kernel size (K) for modulation. threshold (`float`, *optional*, defaults to None): Score threshold to keep object detection predictions. target_sizes (`torch.Tensor` or `list[tuple[int, int]]`, *optional*): Tensor of shape `(batch_size, 2)` or list of tuples (`tuple[int, int]`) containing the target size `(height, width)` of each image in the batch. If unset, predictions will be resize with the default value. Returns: `list[list[Dict]]`: A list of dictionaries, each dictionary containing the keypoints and boxes for an image in the batch as predicted by the model. rzTMake sure that you pass in as many target sizes as the batch dimension of the logitsr?rFrG)r4r5r|r keypointsrlabelsbbox)r|:rTrN)r|:rr=N)rryrkrrr9r\rCr8 itertoolschainrjrBrprKrrrrtensorrrrnextrr)$rrrrrrrrrcentersscalesflattened_boxesrqr4r5 scale_factorrFrGrIrJrr all_boxesposesr  bboxes_xyxyresultspose_bbox_pairs image_bboxes image_resultsposescore bbox_xyxykeypoints_labelskeep pose_results$&&&&&& r$post_process_pose_estimation2VitPoseImageProcessor.post_process_pose_estimations8 $(*1*:*:*@*@' 1a  # c,6G(Gst t((J?"**=:q/<y67z"A',8OA,> PQ@R\!xxK(^_ %4%7,%F" IIg. (0C6E3O4FTYhnoMFE"GqDM Fa4L#44    " ( ( *GK5 vHHj!_BJJ? #FO &"6N & U#f%a/ll#5i#@A 79eV[9!L;=M!)-o)>&eY #) ( 9,D:D!$KE'7'=$*D(E8M]_egpq $$[1" NN= )"r'r()N)TrN) )r9NN)r)r*r+r,r-r valid_kwargsmodel_input_namesr rr rrprrrrrrrrrcompilerdisablerrrr7r/r0 __classcell__)rr#s@@r$rrPs@.L'(&J$I c7C (DJL## ; ; ;22" ^^"-a-a^  FFFr'r)rg?)rR)7r-r"r]typingrrrr9rtorchvision.transforms.v2rtvFimage_processing_backendsrimage_processing_utilsrimage_transformsrr image_utilsr r r r r rprocessing_utilsrrutilsrrrrr scipy.linalgr scipy.ndimagerrmodeling_vitposer get_loggerr)loggerrrKrcrwrrrrr__all__r(r'r$rNs) ' 7;2E5 ?8   H % #,e #)X"J4B9x-`(F.FFR # #r'