^j& pdZddlZddlZddlmZddlmZddlm Z m Z ddl m Z m Z ddlmZmZmZmZmZdd lmZmZdd lmZmZd eeefd eeefd efdZ ddedeeeefzeezeezd eeeefzeezeezdeezdzd ef dZGddedZ eGddeZ!dgZ"y)z"Image processor class for Pixtral.N) functional)TorchvisionBackend) BatchFeature get_size_dict)group_images_by_shapereorder_images)ChannelDimension ImageInputPILImageResamplingSizeDictget_image_size) ImagesKwargsUnpack) TensorTypeauto_docstring image_size patch_sizereturnc~|\}}t|ttfr|n||f\}}|dz |zdz}|dz |zdz}||fS)aP Calculate the number of image tokens given the image size and patch size. Args: image_size (`tuple[int, int]`): The size of the image as `(height, width)`. patch_size (`tuple[int, int]`): The patch size as `(height, width)`. Returns: `int`: The number of image tokens. ) isinstancetuplelist)rrheightwidth patch_height patch_widthnum_width_tokensnum_height_tokenss /var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/transformers/models/pixtral/image_processing_pixtral.py_num_image_tokensr"saMFE.8eT].S ZdfpYqL+ k1A5! 4q8 . .. input_imagesizeinput_data_formatct|ttfr|n||f\}}t|ttfr|n||f\}}t||\}} t ||z | |z } | dkDrBt t j|| z }t t j| | z } t|| f||f\} } | |z| |zfS)aN Find the target (height, width) dimension of the output image after resizing given the input image and the desired size. Args: input_image (`ImageInput`): The image to resize. size (`int` or `tuple[int, int]`): Max image size an input image can be. Must be a dictionary with the key "longest_edge". patch_size (`int` or `tuple[int, int]`): The patch_size as `(height, width)` to use for resizing the image. If patch_size is an integer, `(patch_size, patch_size)` will be used input_data_format (`ChannelDimension`, *optional*): The channel dimension format of the input image. If unset, will use the inferred format from the input. Returns: `tuple`: The target (height, width) dimension of the output image after resizing. r) rrrrmaxintmathfloorr") r$r%rr& max_height max_widthrrrrratior rs r!get_resize_output_image_sizer/1s0%/teT]$CD$PTJ .8eT].S ZdfpYqL+";0ABMFE #UY%6 7E qyTZZ/0DJJuu}-.*;VUOl\gMh*i'' | +-= -K KKr#c,eZdZUdZeeefezed<y)PixtralImageProcessorKwargsz patch_size (`Union[dict[str, int], int]` *optional*, defaults to `{"height": 16, "width": 16}`): Size of the patches in the model, used to calculate the output image size. rN)__name__ __module__ __qualname____doc__dictstrr)__annotations__r#r!r1r1Ys S#X$$r#r1F)totalceZdZejZgdZgdZdddZddiZ dZ dZ dZ dZ dZeZdd gZd eeffd Zed ed eed effd Z d#dddededdd df fd Zdedd eeeefd dfdZ d#d eddededddedededededeeezdzdeeezdzd edzd!e e!zdzde"e efezdzd efd"Z#xZ$S)$PixtralImageProcessor)g3<4'?gwg M?gy{ ?)gB91?gwt.?g U?rr longest_edgeiT pixel_values image_sizeskwargsc $t|di|y)Nr9)super__init__)selfrB __class__s r!rEzPixtralImageProcessor.__init__rs "6"r#imagesrc $t||fi|SN)rD preprocess)rFrHrBrGs r!rKz PixtralImageProcessor.preprocessusw!&3F33r#Nimagez torch.Tensorr%rresamplez7PILImageResampling | tvF.InterpolationMode | int | Nonec |jr|j|jf}n<|jr%|jr|j|jf}n td|jr%|jr|j|jf}n tdt |||}t ||ft|d|d|d|S)a Resize an image. The longest edge of the image is resized to size["longest_edge"], with the aspect ratio preserved. Output dimensions are aligned to patch_size. Args: image (`torch.Tensor`): Image to resize. size (`SizeDict`): Dict containing the longest possible edge of the image. patch_size (`SizeDict`): Patch size used to calculate the size of the output image. resample (`PILImageResampling | tvF.InterpolationMode | int | None`, *optional*): Resampling filter to use when resizing the image. z@size must contain either 'longest_edge' or 'height' and 'width'.z-patch_size must contain 'height' and 'width'.)r%rrrr>)r%rM)r?rr ValueErrorr/rDresizer ) rFrLr%rrMrB size_tuplepatch_size_tuple output_sizerGs r!rPzPixtralImageProcessor.resizeys,   ++T->->?J [[TZZ++tzz2J_` `   !1!1 * 1 1:3C3CD LM M25zVfg w~   Ak!nMX` dj  r#c :td|Dtd|Df}t||Dcgc]F\}}tjjj |d|d|dz d|d|dz fH}}}tj |Scc}}w)a Pads images to form a batch of same shape. Args: pixel_values (`list[torch.Tensor]`): A list of pixel values, each of shape (channels, height, width). image_sizes (`list[tuple[int, int]]`): A list of (height, width) for each image. Returns: `torch.Tensor`: Stacked and padded images. c3&K|] }|d yw)rNr9.0ss r! z:PixtralImageProcessor._pad_for_batching..s3!13c3&K|] }|d yw)rNr9rVs r!rYz:PixtralImageProcessor._pad_for_batching..s8S!18SrZrr)pad)r(ziptorchnnrr\stack)rFr@rA max_shapeimgr%paddeds r!_pad_for_batchingz'PixtralImageProcessor._pad_for_batchings"3{33S8S{8S5ST !{; T HH   # #Ca1Q1GIVWL[_`a[bLb-c # d  {{6""  sA B do_resizedo_center_crop crop_size do_rescalerescale_factor do_normalize image_mean image_stddisable_groupingreturn_tensorsc Zt|xs |jd}tdi|}t|| \}}i}|j D]!\}}|r|j ||||}|||<#t ||}t|| \}}tt|Dcgc] }||d }}i}|j D]4\}}|r|j||}|j|||| | | }|||<6t ||}|j||}t||d| Scc}w) NT)default_to_square)rm)rLr%rrMr)r@rA)data tensor_typer9) rrr ritemsrPr rangelen center_croprescale_and_normalizerdr)rFrHrer%rMrfrgrhrirjrkrlrmrnrrB patch_size_sdgrouped_imagesgrouped_images_indexresized_images_groupedshapestacked_imagesresized_imagesibatch_image_sizesprocessed_images_groupedprocessed_images padded_imagess r! _preprocessz!PixtralImageProcessor._preprocesss$#:#@TXY  .:. /DV^n/o,,!#%3%9%9%; ; !E>!%(t X`"-"-; "5 )  ; ((>@TU/D^fv/w,,AFsK_G`AabA1!4Q7bb#% %3%9%9%; = !E>!%!1!1.)!L!77 NL*V_N/= $U +  =**BDXY..))/ "/@QR`n  #csD(rJ)%r2r3r4r BICUBICrMrkrlrr%rprerhrjdo_convert_rgbr1 valid_kwargsmodel_input_namesrrErr rrKr rPrrr)rdboolfloatr7rr6r __classcell__)rGs@r!r<r<bs!))H4J4I,J D !DIJLN.L'7#(C!D#44v>Y7Z4_k44OS % % %  % L %  % N#>*#%S/*#  #N8<3 ^$3 3  3 L 3  3 3 3 3 3 DK'$.3 4;&-3 +3 j(4/3 cNX-43 " #3 r#r<rJ)#r5r*r^torchvision.transforms.v2rtvFimage_processing_backendsrimage_processing_utilsrrimage_transformsrr image_utilsr r r r rprocessing_utilsrrutilsrrrr)r"rr7r/r1r<__all__r9r#r!rs) 7;AEee4//%S//uS#X/SV/08< %L%L c3h $s) +eCj 8%LeCHo%S 1E#J>%L--4 %L  %LP%,e%H .H H V # #r#