^j ,dZddlmZddlZddlmZddlmZddl m Z ddl m Z dd l mZmZdd lmZmZmZmZmZmZdd lmZmZdd lmZmZmZmZej>e Z!Gd dedZ"eGddeZ#dgZ$y)z$Image processor class for MobileViT.)UnionN) functional)TorchvisionBackend)'SemanticSegmentationPostProcessorOutput) BatchFeature)group_images_by_shapereorder_images)IMAGENET_STANDARD_MEANIMAGENET_STANDARD_STDChannelDimension ImageInputPILImageResamplingSizeDict) ImagesKwargsUnpack) TensorTypeauto_docstringloggingrequires_backendsc&eZdZUdZeed<eed<y)MobileVitImageProcessorKwargsa do_flip_channel_order (`bool`, *optional*, defaults to `self.do_flip_channel_order`): Whether to flip the color channels from RGB to BGR or vice versa. do_reduce_labels (`bool`, *optional*, defaults to `self.do_reduce_labels`): Whether or not to reduce all label values of segmentation maps by 1. Usually used for datasets where 0 is used for background, and background itself is not included in all classes of a dataset (e.g. ADE20k). The background label will be replaced by 255. do_flip_channel_orderdo_reduce_labelsN)__name__ __module__ __qualname____doc__bool__annotations__/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/transformers/models/mobilevit/image_processing_mobilevit.pyrr-s r"rF)totalceZdZdZeZejZe Z e Z ddiZ dZdddZdZdZdZdZdZdZdZd eeffd Ze d)d ed edzd eed effd Z d)d ed edzdededee zdzde!edfdzd efdZ"de#dd e#dfdZ$d*dZ% d+d e#ddede&dddede&ded e'd!edzd"ed#ed e#dfd$Z( d,d%e#e)e*e*fdzd&ed d'fd(Z+xZ,S)-MobileViTImageProcessorzSTorchvision backend for MobileViT with flip_channel_order and reduce_label support. shortest_edgeF)heightwidthTNkwargsc $t|di|y)Nr!)super__init__)selfr, __class__s r#r/z MobileViTImageProcessor.__init__Os "6"r"imagessegmentation_mapsreturnc &t|||fi|S)zp segmentation_maps (`ImageInput`, *optional*): The segmentation maps to preprocess. )r. preprocess)r0r2r3r,r1s r#r6z"MobileViTImageProcessor.preprocessRsw!&*;FvFFr"do_convert_rgbinput_data_formatreturn_tensorsdevicez torch.devicec |j||||}|j}d|d<i} |j|fi|| d<||j|ddtj} |j} | j ddt jd|jd d| i| } | D cgc]0} | jd jtj2} } | | d <t| | Scc} w) z"Handle extra inputs beyond images.)r2r7r8r:Fr pixel_values)r2expected_ndimsr7r8) do_rescalerresampler2rlabels)data tensor_typer!) _prepare_image_like_inputscopy _preprocessr FIRSTupdaterNEARESTsqueezetotorchint64r) r0r2r3r7r8r9r:r, images_kwargsrBprocessed_segmentation_mapssegmentation_maps_kwargsprocessed_segmentation_maps r#_preprocess_image_like_inputsz5MobileViTImageProcessor._preprocess_image_like_inputs_s900.L]fl1  ,1 ()/t//H-H^  (*.*I*I( $"2"8"8 +J+ '(.{{} $ $ + +"'-2 2 : :  +;$*:*:+2+6N+ ' 3N+.+221588E+ '+9DN>BB +s35C<rA torch.Tensorc ftt|D]}||}tj|dk(tjd|j |j |}|dz }tj|dk(tjd|j |j |}|||<|S)z/Reduce label values by 1, replacing 0 with 255.r)dtyper:)rangelenrLwheretensorrVr:)r0rAidxlabels r# reduce_labelz$MobileViTImageProcessor.reduce_labelsV% C3KEKK ELLEKKX]XdXd,eglmEAIEKK ell3ekkZ_ZfZf.ginoEF3K   r"c|jdk(r|j}|gd|dd|S|jdk(r&|j}|ddgdf|ddddf<|S|S)zFlip RGB to BGR or vice versa.r)r=rWrrN)ndimclone)r0r2flippeds r#flip_channel_orderz*MobileViTImageProcessor.flip_channel_ordersi ;;! llnG!),GAaLN [[A llnG$Q \2GAqsFON r" do_resizesizer@z7PILImageResampling | tvF.InterpolationMode | int | Nonedo_center_crop crop_sizer?rescale_factordisable_groupingrrc  | r|j|}t|| \} }i}| jD]\}}|r|j|||}|||<!t ||}t|| \} }i}| jD]E\}}|r|j ||}|r|j ||}| r|j|}|||<Gt ||}|S)z#Custom preprocessing for MobileViT.)rk)r_r itemsresizer center_croprescalere)r0r2rfrgr@rhrir?rjrkrrr,grouped_imagesgrouped_images_indexresized_images_groupedshapestacked_imagesresized_imagesprocessed_images_groupedprocessed_imagess r#rFz#MobileViTImageProcessor._preprocesss &&v.F/DV^n/o,,!#%3%9%9%; ; !E>!%^T8!L,: "5 ) ;((>@TU/D^fv/w,,#% %3%9%9%; = !E>!%!1!1.)!L!%nn!M$!%!8!8!H.< $U + =**BDXYr" target_sizesreturn_segmentation_scoreszBlist[torch.Tensor] | list[SemanticSegmentationPostProcessorOutput]ct|d|j}t|}|t|t|k7r tdt |t j r|j}g}t|D]y}t jjj||jd||dd}|jt|djd|dd {n;t|D cgc]'} t|| jd|| d )}} |s|D cgc]} | j }} |Scc} wcc} w) a Converts the output of [`MobileViTForSemanticSegmentation`] into semantic segmentation maps. Args: outputs ([`MobileViTForSemanticSegmentation`]): Raw outputs of the model. target_sizes (`list[tuple[int, int]]`, *optional*): List of tuples corresponding to the requested final size (height, width) of each prediction. return_segmentation_scores (`bool`, *optional*, defaults to `False`): Whether to return segmentation scores alongside the segmentation map. When `True`, each element of the returned list is a [`SemanticSegmentationPostProcessorOutput`] with fields `segmentation` (class IDs, shape `(height, width)`) and `segmentation_scores` (shape `(num_classes, height, width)`). Returns: `list[torch.Tensor]` or `list[SemanticSegmentationPostProcessorOutput]`: When `return_segmentation_scores=False` (default), a list of length `batch_size` where each item is a segmentation map of shape `(height, width)` with class IDs. When `return_segmentation_scores=True`, a list of [`SemanticSegmentationPostProcessorOutput`] with fields `segmentation` (class IDs, shape `(height, width)`) and `segmentation_scores` (shape `(num_classes, height, width)`). In both cases, `(height, width)` corresponds to the target size (if `target_sizes` is specified). rLzTMake sure that you pass in as many target sizes as the batch dimension of the logitsr)dimbilinearF)rgmode align_corners) segmentationsegmentation_scores)rB)rlogitsrZ ValueError isinstancerLTensortolistrYnnr interpolate unsqueezeappendrargmaxr) r0outputsryrzr batch_sizesemantic_segmentationr]resized_logitsiitems r#"post_process_semantic_segmentationz:MobileViTImageProcessor.post_process_semantic_segmentationsx6 $([  #6{c,// j, 5+224 $& !Z( !&!4!4!@!@3K))a)0|C7Hzin"A"&,,;,:1,=,D,D,D,K3A!3D "z* %8*0)*:*:q*:*AZ`abZcd% !%*CX$Y4T%6%6$Y !$Y$$%%Zs ,E=E)N)r2rSr4rS)FT)NF)-rrrrr valid_kwargsrBICUBICr@r image_meanr image_stdrgdefault_to_squarerirfrhr? do_normalizer7rrrr/rrrr6rr strrrrRlistr_rerfloatrFtupleintr __classcell__)r1s@r#r&r&;s)]0L!))H'J%I S !D-IINJLN #(E!F#04 G G&, G67 G  G G&59/C/C&,/C /C , /C j(4/ /Cc>)*T1/C /Cb4#7Drs+ 7;O2E5   H % L Q%0Q%Q%h % %r"