qyiYBSrSSKrSSKrSSKJr SSKrSSKJs Js J r SSK J r SSK JrJrJr SSKJr SSKJrJr SS KJrJr SS KJrJrJr SS KJrJr SS K J!r!J"r" SS K#J$r$ Sr%"SS\SS9r&S"S\'S\(S\'S\'S\(S\(S\(S\(S\)S-S\'S-S\R4Sjjr*\RVRXS5r-\""S S!\55r.S!/r/g)#z%Image processor class for Pix2Struct.N)Union)hf_hub_download)Image ImageDraw ImageFont)TorchvisionBackend) BatchFeature get_size_dict)group_images_by_shapereorder_images)ChannelDimension ImageInputSizeDict) ImagesKwargsUnpack) TensorTypeauto_docstring)requires_backendszybelkada/fontsc^\rSrSr%Sr\\S'\\\4\S'\ \S'\ \\-S-\S'Sr g) Pix2StructImageProcessorKwargs%a] max_patches (`int`, *optional*): Maximum number of patches to extract. patch_size (`dict[str, int]`, *optional*, defaults to `{"height": 16, "width": 16}`): The patch size to use for the image. According to Pix2Struct paper and code, the patch size is 16x16. is_vqa (`bool`, *optional*, defaults to `False`): Whether or not the image processor is for the VQA task. If `True` and `header_text` is passed in, text is rendered onto the input images. header_text (`Union[list[str], str]`, *optional*): Text to render as a header. Only has an effect if `image_processor.is_vqa` is `True`. max_patches patch_sizeis_vqaN header_text) __name__ __module__ __qualname____firstlineno____doc__int__annotations__dictstrboollist__static_attributes__rw/home/mande/repo/quber/.venv/lib/python3.13/site-packages/transformers/models/pix2struct/image_processing_pix2struct.pyrr%s5 S#X LcS4''r*rF)totaltext text_size text_colorbackground_color left_padding right_padding top_paddingbottom_padding font_bytes font_pathreturnc H[[S5 [R"SS9n U R US9n SR U 5n UbU c[ R"U5n OU bU n O[[S5n [R"U SUS9n [R"S S 5n[R"U5nUR!S XS 9u nnnUU-U-nUU-U-n[R"S UU4U5n[R"U5nUR#XF4XU S 9 U$)aC Render text. This script is entirely adapted from the original script that can be found here: https://github.com/google-research/pix2struct/blob/main/pix2struct/preprocessing/preprocessing_utils.py Args: text (`str`, *optional*, defaults to ): Text to render. text_size (`int`, *optional*, defaults to 36): Size of the text. text_color (`str`, *optional*, defaults to `"black"`): Color of the text. background_color (`str`, *optional*, defaults to `"white"`): Color of the background. left_padding (`int`, *optional*, defaults to 5): Padding on the left. right_padding (`int`, *optional*, defaults to 5): Padding on the right. top_padding (`int`, *optional*, defaults to 5): Padding on the top. bottom_padding (`int`, *optional*, defaults to 5): Padding on the bottom. font_bytes (`bytes`, *optional*): Bytes of the font to use. If `None`, the default font will be used. font_path (`str`, *optional*): Path to the font to use. If `None`, the default font will be used. visionP)width)r- z Arial.TTFzUTF-8)encodingsizeRGB)r@rr)font)fillrB)r render_texttextwrap TextWrapperwrapjoinioBytesIOrDEFAULT_FONT_PATHrtruetypernewrDrawtextbboxr-)r-r.r/r0r1r2r3r4r5r6wrapperlines wrapped_textrBtemp_img temp_draw_wh text_width text_heightimgdraws r+rDrD9s Lk8,"",G LLdL #E99U#L)"3zz*%  0+>   dW9 EDyy'Hx(I##FL#DJAq!Q\!M1Jk/N2K ))EJ 46F GC >># DII|)[TU]"S0UD6nUb-[U[5(d[S0[ USS9D6US'U$XS'U$)zS Process custom Pix2Struct kwargs, specifically converting patch_size to SizeDict. r)r> param_namer)super_standardize_kwargs isinstancerr )selfrkwargs __class__s r+rz,Pix2StructImageProcessor._standardize_kwargssY,6v6  !*Z*J*J#+#fmXd.e#fF<  $.<  r*c g)zC Skip standard validation as Pix2Struct uses custom preprocessing. Nr)r|r}s r+_validate_preprocess_kwargs4Pix2StructImageProcessor._validate_preprocess_kwargss r*image torch.Tensorheaderr5r6cpURnURnUR[R:Xa[R "U5nOHUS-R SS5R[R5n[R "U5n[X#US9n [U RUR5n [URXR- -5n [U RXR- -5n [R"SXU -4S5n U RU R!X45S5 U RUR!X45SU 45 [R""U 5RU5nU[R:waUR%5S- nU$)a Render header text on image using torch tensors. Args: image (`torch.Tensor`): Image tensor in channel-first format (C, H, W). header (`str`): Header text to render. font_bytes (`bytes`, *optional*): Font bytes to use for rendering. font_path (`str`, *optional*): Path to font file to use for rendering. Returns: `torch.Tensor`: Image with header in channel-first format (C, H, W). rr5r6r?whiterAgo@)devicedtyperbuint8tvF to_pil_imageclamptorDmaxr;r#rmrrMpasteresize pil_to_tensorfloat)r|rrr5r6rr image_pil image_uint8 header_image new_width new_heightnew_header_height new_imageresults r+ render_header&Pix2StructImageProcessor.render_headersl.  ;;%++ %((/I!3;--a588EK((5I#6IV  **IOO< ))Y-HIJ  3 3yCUCU7U VWIIei>O1O%PRYZ  ++Y,JKVT (())@AAGXCYZ""9-008 EKK \\^e+F r*imagesc$URSSS9nURSSS9nURSURS-URS-nSUS-- n[R"U[R "XSR S 95nX- U- $) z Normalize batched images using per-image mean and standard deviation. Args: images (`torch.Tensor`): Batched float image tensor of shape (B, C, H, W). Returns: `torch.Tensor`: Normalized images of shape (B, C, H, W). )r@r`rT)dimkeepdimr@r`rg??r)meanstdrarbmaximumtensorr)r|rrrnum_elements_per_imagemin_stdadjusted_stddevs r+ normalize"Pix2StructImageProcessor.normalizes{{y${7jjYj5!'a6<<?!BV\\RS_!T.33--U\\'**-UV 00r*rc URURpTURupgpX$U- -XY- -S-n [[ [ X-U- 5U5S5n [[ [ X-U- 5U5S5n [X-S5n [X-S5n[ XS9nURX[RRSS9n[XU5nURunnnnURUUU-U5n[R"UURS9RUS5R!SU5RSUU-S5n[R"UURS9RSU5R!US5RSUU-S5nUR#USS5nUR#USS5nUS-R%5nUS-R%5n[R&"UUU/SS9n[R(R*R-US S S UUU-- /5R%5nU$) a Extract flattened patches from a batch of images. Args: images (`torch.Tensor`): Batched images tensor of shape (batch, channels, height, width). max_patches (`int`): Maximum number of patches to extract. patch_size (`SizeDict`): Dictionary containing patch height and width. Returns: `torch.Tensor`: Batched flattened patches with row/column IDs of shape (batch, max_patches, patch_dim). rr@rtT)rr>resample antialiasrr^rr)rmr;rarminr#rrrInterpolationModeBILINEARrorfrbarangerrepeatexpandrcatrcrdpad)r|rrrrirjrkrl image_height image_widthscalenum_feasible_rowsnum_feasible_colsresized_height resized_width resize_sizernrowscolumnsdepthrow_idscol_idsrs r+extract_flattened_patches2Pix2StructImageProcessor.extract_flattened_patchess=(%/$5$5z7G7Gk:@,,7 l| ;< @YZ_bbC(<|(K$Lk Z\]^C(;k(I$JK XZ[\.=qA-;Q? nJ 6cF[F[FdFdptu(kJ+2==( D'5//*dWneD LLfmm 4 < [TU]"U4SU0UD6$)z header_text (`Union[str, list[str]]`, *optional*): Text to render as a header. Only has an effect if `image_processor.is_vqa` is `True`. r)ry preprocess)r|rrr}r~s r+r#Pix2StructImageProcessor.preprocessKsw!&LkLVLLr*do_convert_rgbinput_data_formatrz torch.devicec URUUUUS9nURSUR5nU(aUc [S5eUR SS5nUR SS5n [ U[ 5(aU/[U5-n[U5V V s/sHupURXU XS9PM nn n UR"U40UD6$s sn n f)z# Preprocess images for Pix2Struct. )rrrrrNz.A header text must be provided for VQA models.r5r6r) _prepare_image_like_inputsgetr ValueErrorpopr{r&len enumerater _preprocess) r|rrrrrr}rr5r6irs r+_preprocess_image_like_inputs6Pix2StructImageProcessor._preprocess_image_like_inputsXs00)/ 1 Hdkk2 " !QRRL$7J ;5I+s++*mc&k9 !*& 1 1HA""5a.Z"e 1  1&11 sC do_normalizereturn_tensorsdisable_groupingc [XS9up0n 0n UR5HupU R[R:XaU R 5n U(aUR U 5n URXUS9nURSS9S:gR 5nXU 'XU 'M [X5n[X5nU(a*[R"USS9n[R"USS9n[UUS.US9$)z1 Preprocess images to extract flattened patches. )r)rrrr^rr)rurv)data tensor_type) r itemsrrbrrrrsumr stackr )r|rrrrrrr}grouped_imagesgrouped_images_indexflattened_patches_groupedattention_masks_groupedrastacked_imagesrnmasksruattention_maskss r+r$Pix2StructImageProcessor._preprocesss 0EV/o,$&!"$%3%9%9%; !E##u{{2!/!5!5!7!%!?44%:5G[[R[(A-446E/6e ,-2E *&<"++D[()@W  % ,=1 E #kk/qAO'8O\&  r*r)N)NN)rrr7r)&rrr r!rescale_factorrrrrrr valid_kwargsmodel_input_namesr%r&r#rrzrbytesrrrrrr(rr rrFIRSTr'rrrrr) __classcell__)r~s@r+rqrqs NLN,JK F1L,.>?8<cNX-4    $( $ 666DL 6 : 6  6p1*CCC C  CJ/3 M M49_t+ M78 M  M M/3#.>.D.D48 &2&249_t+&2 &2 , &2 c>)*T1 &278&2 &2P/ ^$/ /  /  / j(4/ / /  / / r*rq) $blackrrrrNN)0r"rIrEtypingrrb$torchvision.transforms.v2.functional transformsv2rdrhuggingface_hubrPILrrrimage_processing_backendsr image_processing_utilsr r image_transformsr r image_utilsrrrprocessing_utilsrrutilsrrutils.import_utilsrrKrr&r#rrDcompilerdisablerorq__all__rr*r+rs@,  22+++;AEAA4/3%(\(,## C CCC C  C  CCC CTzC [[CN*X 1X X v & &r*