qyi7SrSSKJr SSKrSSKJr SSKJr SSK J r SSK J r J r SS KJrJrJrJrJr SS KJrJr SS KJrJr \"S S 9S\S\S\\\\44Sj5r\"SS 9S\\\4S\\\4S\S\S\\\44 Sj5rS\S\S\S\S\S\4 Sjr S\S\S\\\4S\\\\\\44Sjr!\"SS 9S*S \\\4S!\S\S"\S\\\44 S#jj5r""S$S%\S&S'9r#\"S(S)\55r$S)/r%g)+z Image processor class for OVIS2.) lru_cacheN) functional)TorchvisionBackend) BatchFeature)group_images_by_shapereorder_images)OPENAI_CLIP_MEANOPENAI_CLIP_STD ImageInputPILImageResamplingSizeDict) ImagesKwargsUnpack) TensorTypeauto_docstring )maxsizemin_image_tilesmax_image_tilesreturnc/n[SUS-5H?n[SUS-5H)nX4-U::dM X4-U:dMURX445 M+ MA [USS9$)zYComputes all allowed aspect ratios for a given minimum and maximum number of input tiles.cUSUS-$Nrrxs m/home/mande/repo/quber/.venv/lib/python3.13/site-packages/transformers/models/ovis2/image_processing_ovis2.py1get_all_supported_aspect_ratios..,sqtad{key)rangeappendsorted)rr aspect_ratioswidthheights rget_all_supported_aspect_ratiosr+$shMq/A-.A23F~0U^5V$$e_54/ -%: ;;r"doriginal_image_sizetarget_tile_sizec[X#5nUupVUupxXe- n Xe-n [S5n Sn UHIn U SU S- n[X- 5nX:aUn U n M&X:XdM-U SU-U-U S-U S-:dMGU n MK U $)zQFind the canvas with the closest aspect ratio to the original image aspect ratio.inf)rrrrg?)r+floatabs)r-r.rrpossible_tile_arrangementsoriginal_heightoriginal_widthtarget_tile_heighttarget_tile_width aspect_ratioareabest_ratio_diff best_gridgridgrid_aspect_ratio ratio_diffs rget_optimal_tiled_canvasr?/s"A!b&9#O,<)!3L  +DElOI* Gd1g-9:  '(OI  *c..1BBT!WLtTUwVV + r"leftupperrightlowersidec`X - nX1- n[XV5[XV5peXT:a Xe- U-nUnXV-$N)maxmin)r@rArBrCrDwhs rcompute_patch_covering_arearKKs; A A q9c!iqx EDL  5Lr"rJrIr<c XS-nXS-n[US5VVs/sHIn[US5H3nXd-XS-XbSS- :XaUOUS-U-XRSS- :XaUOUS-U-4PM5 MK snn$s snnfr)r%)rJrIr< row_height col_widthrowcols rsplit_image_into_gridrQUs1gJ!W Ia> "Ca>C O  Q! #A#'Y)>Q! #A#'Z)?  " "   sAA4 image_sizetarget_patch_sizecovering_thresholdc*^UupEXT-n[SU5n/n/n UHRn [XEU 5n [U4SjU 55U- n URX45 X:dM@U RX45 MT U (a[ U SS9S$[ USS9S$)Nrc3>># UHn[/UQTP76v M g7frF)rK).0regionrSs r -get_min_tile_covering_grid..ts"cVbF+GVG5FGVbsc0USSUSS-US*4$rrrs rr ,get_min_tile_covering_grid..{s(QqT!WqtAw=NQRSTQUPU[TU]"S0UD6 g)Nr)super__init__)selfr __class__s rrOvis2ImageProcessor.__init__s "6"r"imagesrc &>[TU]"U40UD6$rF)r preprocess)rrrrs rrOvis2ImageProcessor.preprocesssw!&3F33r"Nz torch.TensorrmrnrorT patch_sizeresamplez7PILImageResampling | tvF.InterpolationMode | int | NonecURSnURURpURSSupU(a[X4U UUS9upO[ X4X4X#5upX-nX-nX-nUR U[ UUS9US9n/n[U5HNnUU -nUU -nUU -UU -US-U -US-U -4nUSUSUS 2USUS 24nURU5 MP [U5S:wa"UR XUS9nURSU5 [R"USS 9RSS5R5n[U5Vs/sHnX/PM nnUU4$s snf) a[ Crop the images to patches and return a list of cropped images. The number of patches and their grid arrangement are determined by the original image size, the target patch size and the minimum and maximum number of patches. The aspect ratio of the patches grid is chosen to be the closest to the original image aspect ratio. Args: images (`torch.Tensor`): The images to be cropped. min_patches (`int`): The minimum number of patches to be extracted from the image. max_patches (`int`): The maximum number of patches to be extracted from the image. use_covering_area_grid (`bool`, *optional*, defaults to `True`): Whether to use the original OVIS2 approach: compute the minimal number of tiles that cover at least 90% of the image area. If `False`, the closest aspect ratio to the target is used. covering_threshold (`float`, *optional*, defaults to `0.9`): The threshold for the covering area. Only has an effect if `use_covering_area_grid` is set to `True`. patch_size (`SizeDict`, *optional*): The size of the output patches. resample (`PILImageResampling | tvF.InterpolationMode | int | None`, *optional*): Resampling filter to use if resizing the image. Returns: Tuple[`torch.Tensor`, `list`]: A tuple containing the processed images tensor and the grid information. rN)rSrrTr})rr.r)dim)shaper*r)rhr?resizerr%r&leninserttorchstack transpose contiguous)rrrmrnrorTrr num_imagepatch_size_heightpatch_size_widthr4r5 num_columnsnum_rows target_width target_height num_blocks resized_imageprocessed_imagesicolumnrObox patch_image thumbnail_img_r<s rcrop_image_to_patches)Ovis2ImageProcessor.crop_image_to_patchessHLLO .8.?.?AQAQ+*0,,rs*;' !$> 1"3 +#5 % !K%= 14E3XZe% !K (5 )4  +  FHMQ],^iq r z"A_F{"C))''!//q-- C(SVc!f_c!fs1vo(MNK  # #K 0#  A % KKXKNM  # #A} 5 ;;'7Q?II!QOZZ\16y1AB1AA'1AB%%Cs E( do_resizesizedo_center_crop crop_size do_rescalerescale_factor do_normalize image_mean image_stddisable_groupingreturn_tensorsrlc U(aUS:a[XS9unn0n0nUR5H'unnURUUUUUUS9unnUUU'UUU'M) [UU5nUVVs/sHnUHnUPM M nnn[UU5nO&[ [ U55Vs/sHnSS/PM nn[XS9unn0nUR5H"unnU(aUR UX4S9nUUU'M$ [UU5n[UU S9unn0nUR5H8unnU(aURUU5nURUXxXU 5nUUU'M: [UU5n[UUS.U S9$s snnfs snf)Nr)r)rror)imagerr) pixel_valuesgrids)data tensor_type) ritemsrr r%rr center_croprescale_and_normalizer) rrrrrrrrrrrrrrrlrmrnrorgrouped_imagesgrouped_images_indexprocessed_images_groupedrrstacked_imagesr< images_listrrresized_images_groupedresized_imagesrs r _preprocessOvis2ImageProcessor._preprocesss* {Q3H3s 0N0') $E)7)=)=)?%~'+'A'A"#+A% (B($3A(/#e *@$$<>RSF/5Ov ;%e;evFO"5*>?E%*3v;%78%7aV%7E80EV/o,,!#%3%9%9%; !E>!%>!`,: "5 )&<((>@TU0E^fv/w,,#% %3%9%9%; !E>!%!1!1.)!L!77 LV_N/= $U +&<**BDXY2BU!Saopp9P9s *E:&Fr)T?NN)Frr~T)(rprqrrrsr BICUBICrr rr rrdefault_to_squarerrrdo_convert_rgbrlrmrnrorj valid_kwargsrrrr rrrwrur1rrliststrrrrx __classcell__)rs@rr{r{s!))H!JIC (DIJLNOKK!,L#(A!B#44v>W7X4]i44(,$'&*NRS&S&S& S& !% S& " S&tOS&LS&H!&'+%Aq^$AqAq Aq L Aq  AqAqAqAqAqDK'$.Aq4;&-Aq+Aqj(4/AqAq !Aq"#Aq$!%%Aq( )AqAqr"r{)r)&rt functoolsrrtorchvision.transforms.v2rtvFimage_processing_backendsrimage_processing_utilsrimage_transformsrr image_utilsr r r r rprocessing_utilsrrutilsrrrwrtupler+r?r1rKrQrhrjr{__all__rr"rrs' 7;2E5/ 2