^j)dZddlmZddlZddlmcmZddlmZ ddl m Z ddl m Z ddlmZdd lmZmZdd lmZmZmZmZmZmZdd lmZmZdd lmZmZm Z Gd dedZ!eGdde Z"dgZ#y)zImage processor class for BEiT.)UnionN) functional)TorchvisionBackend)'SemanticSegmentationPostProcessorOutput) BatchFeature)group_images_by_shapereorder_images)IMAGENET_STANDARD_MEANIMAGENET_STANDARD_STDChannelDimension ImageInputPILImageResamplingSizeDict) ImagesKwargsUnpack) TensorTypeauto_docstringis_torch_availableceZdZUdZeed<y)BeitImageProcessorKwargsak 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_reduce_labelsN)__name__ __module__ __qualname____doc__bool__annotations__y/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/transformers/models/beit/image_processing_beit.pyrr&sr rF)totalceZdZdZeZejZe Z e Z dddZ dZdddZdZdZdZdZdZdeeffd Ze d(d ed edzdeed effd Z d(d ed edzd ededeezdzdeedfdzd efdZ de!dd e!dfdZ" d)d e!ddede#dddede#dede$dede$e!e$zdzd e$e!e$zdzd!edzd"ed e!dfd#Z% d*d$e!e&dzd%ed d&fd'Z'xZ(S)+BeitImageProcessorz/PIL backend for BEiT with reduce_label support.)heightwidthTFkwargsc $t|di|y)Nr)super__init__)selfr( __class__s r!r+zBeitImageProcessor.__init__Cs "6"r Nimagessegmentation_mapsreturnc &t|||fi|S)zp segmentation_maps (`ImageInput`, *optional*): The segmentation maps to preprocess. )r* preprocess)r,r.r/r(r-s r!r2zBeitImageProcessor.preprocessFsw!&*;FvFFr do_convert_rgbinput_data_formatreturn_tensorsdevicez torch.devicec |j||||}|j}d|d<i} |j|fi|| d<||j|ddtj} |j} | j ddd|jd d| i| } | D cgc]0} | j d jtj2} } | | d <t| | Scc} w) z"Handle extra inputs beyond images.)r.r3r4r6Fr pixel_values)r.expected_ndimsr3r4) do_normalize do_rescaler.rlabels)data tensor_typer) _prepare_image_like_inputscopy _preprocessr FIRSTupdatesqueezetotorchint64r) r,r.r/r3r4r5r6r( images_kwargsr>processed_segmentation_mapssegmentation_maps_kwargsprocessed_segmentation_maps r!_preprocess_image_like_inputsz0BeitImageProcessor._preprocess_image_like_inputsSs+00.L]fl1  ,1 ()/t//H-H^  (*.*I*I( $"2"8"8 +J+ '(.{{} $ $ + +URW,X Y*:$*:*:+2+6N+ '3N+.+221588E+ '+9DN>BB +s$5C-r=z torch.Tensorc ftt|D]}||}tj|dk(tjd|j |j |}|dz }tj|dk(tjd|j |j |}|||<|S)z/Reduce label values by 1, replacing 0 with 255.r)dtyper6)rangelenrGwheretensorrPr6)r,r=idxlabels r! reduce_labelzBeitImageProcessor.reduce_labelsV% C3KEKK ELLEKKX]XdXd,eglmEAIEKK ell3ekkZ_ZfZf.ginoEF3K   r do_resizesizeresamplez7PILImageResampling | tvF.InterpolationMode | int | Nonedo_center_crop crop_sizer<rescale_factorr; image_mean image_stddisable_groupingrc | r|j|}t|| \}}i}|jD]\}}|r|j|||}|||<!t ||}t|| \}}i}|jD]4\}}|r|j ||}|j |||| | | }|||<6t ||}|S)zCustom preprocessing for BEiT.)rb)rYr itemsresizer center_croprescale_and_normalize)r,r.rZr[r\r]r^r<r_r;r`rarbrr(grouped_imagesgrouped_images_indexresized_images_groupedshapestacked_imagesresized_imagesprocessed_images_groupedprocessed_imagess r!rBzBeitImageProcessor._preprocesss$ &&v.F0EV^n/o,,!#%3%9%9%; ; !E>!%^T8!L,: "5 ) ;((>@TU0E^fv/w,,#% %3%9%9%; = !E>!%!1!1.)!L!77 NL*V_N/= $U + =**BDXYr target_sizesreturn_segmentation_scoreszBlist[torch.Tensor] | list[SemanticSegmentationPostProcessorOutput]cts td|j}|t|t|k7r t dt |t jr|j}g}tt|D]g}tj||jd||dd}|djd}|jt||dd inJ|jd } t|j dD cgc]} t| | || d }} |s|D cgc]} | j"}} |Scc} wcc} w) a Converts the output of [`BeitForSemanticSegmentation`] into semantic segmentation maps. Args: outputs ([`BeitForSemanticSegmentation`]): Raw outputs of the model. target_sizes (`list[Tuple]` of length `batch_size`, *optional*): List of tuples corresponding to the requested final size (height, width) of each prediction. If unset, predictions will not be resized. 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). z:PyTorch is required for post_process_semantic_segmentationzTMake sure that you pass in as many target sizes as the batch dimension of the logitsr)dimbilinearF)r[mode align_corners) segmentationsegmentation_scores)r>rQ)r ImportErrorlogitsrT ValueError isinstancerGTensornumpyrSF interpolate unsqueezeargmaxappendrrkrw) r,outputsrprqrzsemantic_segmentationrWresized_logits semantic_mapseg_mapsiitems r!"post_process_semantic_segmentationz5BeitImageProcessor.post_process_semantic_segmentationsy2"#Z[ [  #6{c,// j, 5+113 $& !S[) !"3K))a)0|C7Hzin" .a077A7> %,,;.:SabcSde }}}+H v||A/ %8*21+fUViX% !%*CX$Y4T%6%6$Y !$Y$$%%Zs EE)N)F)NF))rrrrr valid_kwargsrBICUBICr\r r`r rar[default_to_squarer^rZr]r<r;rrr+rrrr2rr strrrrMlistrYrfloatrBtupler __classcell__)r-s@r!r$r$1s99+L!))H'J%IC (D-IINJL#(@!A#04 G G&, G12 G  G G&59*C*C&,*C *C , *C j(4/ *Cc>)*T1*C *CX4#7Drsr& 7;O2E5CC|5E%+E%E%P r