tj:UddlmZddlZddlmZddlmcmZddlm Z ddl m Z ddl m Z mZmZmZddlmZddlmZdd lmZdd lmZmZdd lmZmZdd lmZdd lm Z m!Z!ddl"m#Z#ddl$m%Z%ddl&m'Z'm(Z(m)Z)ddl*m+Z+m,Z,ddl-m.Z.ddl/m0Z0ddl1m2Z2ddl3m4Z4ddl5m6Z6e(de Gdde eZ7e(de GddeZ8e(e.dGd d!eZ9Gd"d#e4Z:Gd$d%ej;Z<Gd&d'ej;Z=Gd(d)ej;Z>Gd*d+eZ?Gd,d-ej;Z@e(Gd.d/e6ZAGd0d1eAZBe(d23Gd4d5eeAZCGd6d7ej;ZDe(d83Gd9d:eAZEgd;ZFdS)<)SequenceN)strict)ACT2FN)BackboneConfigMixin BackboneMixin%consolidate_backbone_kwargs_to_configfilter_output_hidden_states)PreTrainedConfig) BatchFeature)TorchvisionBackend)group_images_by_shapereorder_images)PILImageResamplingSizeDict)GradientCheckpointingLayer)BackboneOutputBaseModelOutputWithNoAttention)PreTrainedModel)Unpack)TransformersKwargsauto_docstringcan_return_tuple) TensorTypemerge_with_config_defaults)requires)capture_outputs) AutoConfig)PPLCNetConvLayer)PPOCRV5ServerDetPreTrainedModelzPaddlePaddle/UVDoc_safetensors) checkpointcLeZdZUdZdZdZeedzed<dZ ee dzed<dZ e ee e e dfzed<d Ze e e e e e efee ezzed <d Ze e e e dfee zed <d Ze ed<fdZxZS)UVDocBackboneConfiga resnet_head (`Sequence[list[int] | tuple[int, ...]]`, *optional*, defaults to `((3, 32), (32, 32))`): Configuration for the ResNet head layers in format [in_channels, out_channels]. resnet_configs (`Sequence[Sequence[tuple[int, int, int, bool] | list[int | bool]]]`, *optional*, defaults to `(((32, 32, 1, False), (32, 32, 3, False), (32, 32, 3, False)), ((32, 64, 1, True), (64, 64, 3, False), (64, 64, 3, False), (64, 64, 3, False)), ((64, 128, 1, True), (128, 128, 3, False), (128, 128, 3, False), (128, 128, 3, False), (128, 128, 3, False), (128, 128, 3, False)))`): Configuration for the ResNet stages in format [in_channels, out_channels, dilation_value, downsample]. stage_configs (Sequence[Sequence[tuple[int, ...] | list[int]]], *optional*, defaults to `(((128, 1),), ((128, 2),), ((128, 5),), ((128, 8),(128, 3),(128, 2),), ((128, 12), (128, 7), (128, 4),), ((128, 18), (128, 12), (128, 6),),)`): Configuration for the bridge module stages in format [in_channels, dilation_value]. Each inner sequence corresponds to a single bridge block, and the outer sequence groups blocks by bridge stage. uvdoc_backboneN _out_features _out_indices))r )r(r(. resnet_head)))r(r(Fr(r(rFr+))r(@r*Tr,r,rFr-r-))r,r*Tr.r.rFr/r/r/r/resnet_configs)))r.r*)r.r))r.))r.)r.rr1)r. )r.)r.))r.r4)r. stage_configsr2 kernel_sizec Td|jD|_dgdtdt|jdzDz|_||dd|ddtjdi|dS) Nc,g|]}t|S)len).0stagess e/var/www/html/CA-Chatbot/venv/lib/python3.11/site-packages/transformers/models/uvdoc/modular_uvdoc.py z5UVDocBackboneConfig.__post_init__..xsDDDvs6{{DDDstemcg|]}d|S)stager>)r@idxs rBrCz5UVDocBackboneConfig.__post_init__..ys&f&f&f}s}}&f&f&frDr* out_indices out_features)rIrJr>) r:depthsranger? stage_names"set_output_features_output_indicespopsuper __post_init__selfkwargs __class__s rBrQz!UVDocBackboneConfig.__post_init__wsDD1CDDD "8&f&faTM_I`I`cdId@e@e&f&f&ff // =$77fjjQ_aeFfFf 0    '''''''rD)__name__ __module__ __qualname____doc__ model_typer&liststr__annotations__r'intr)rtupler0boolr:r;rQ __classcell__rUs@rBr$r$0s0  "J&*M49t#***%)L$s)d"))):K$s)eCHo56 YNHXeCc4,?&@4d CS&STU,FM8HU38_tCy%@AB*K(((((((((rDr$ceZdZUdZdZdeiZdZee zdze d<dZ e e d<dZ e e d<d Zee d <d Zeeeed fze d <dZeeeeed fze d<fdZxZS) UVDocConfiga padding_mode (`str`, *optional*, defaults to `"reflect"`): Padding mode for convolutional layers. Supported modes are `"reflect"`, `"constant"`, and `"replicate"`. kernel_size (`int`, *optional*, defaults to 5): Kernel size for convolutional layers in the backbone network. bridge_connector (`list[int] | tuple[int, ...]`, *optional*, defaults to `(128, 128)`): Configuration for the bridge connector in format [in_channels, out_channels]. out_point_positions2D (`Sequence[list[int] | tuple[int, ...]]`, *optional*, defaults to `((128, 32), (32, 2))`): Configuration for the output point positions 2D layer in format [in_channels, out_channels]. uvdocbackbone_configNprelu hidden_actreflect padding_moder2r;)r.r..bridge_connector))r.r()r(rout_point_positions2Dc rtd|jdd|\|_}tjdi|dS)Nr%)rfdefault_config_typer>)r rfrPrQrRs rBrQzUVDocConfig.__post_init__s\'L(  0 0( ( ( ( $f '''''''rD)rVrWrXrYrZr sub_configsrfdictr r]rhr\rjr;r^rkr[r_rlrrQrarbs@rBrdrds  J$j1K6:OT,,t3:::J!L#!!!K4>d3i%S/1>>>CW8DIc3h$?@WWW(((((((((rDrd)torch)backendsceZdZdZdZdddZejZde dde de d e d e d e d e e e zd zde e e zd zde d zde ezd zdefdZ ddejde ejde de ee ejffdZd S)UVDocImageProcessorTii)heightwidthimagesz torch.Tensor do_resizesize do_rescalerescale_factor do_normalize image_meanN image_stddisable_groupingreturn_tensorsreturnc t|| \} } i}| D]8\}}|||||||}|ddgdddddf}|||<9t|| }|}t|| \} } i}| D]0\}}|r$t j||j|jfdd}|||<1t|| }t||d| dgS) N)r)rr*rbilinearTrymode align_corners) pixel_valuesoriginal_imagesr)data tensor_typeskip_tensor_conversion) ritemsrescale_and_normalizercopyF interpolaterurvr )rSrwrxryrzr{r|r}r~rrrTgrouped_imagesgrouped_images_indexprocessed_images_groupedshapestacked_imagesrescale_and_normalize_imagesrinterpolated_images_groupedrs rB _preprocesszUVDocImageProcessor._preprocesss0EV^n/o/o/o,,#% %3%9%9%;%; = = !E>!77 NL*V_N,AAAyyy!!!QQQ,>?N.< $U + +'56NPd'e'e$6;;==/D (;K0 0 0 ,,')#%3%9%9%;%; @ @ !E> !""$+tz)Bcg"""2@ ' . .%&ACWXX ".?SS&$5#6    rDo@ predictionrscalect|}tjt||j}g}t |D]#\}}|jdkr|d}||j}|j dd\}} tj |||dz|| fdd } | dddd} tj || d } | d ddd} | |z} | d g tjdd } |d| i%|S)a Post-process document rectification predictions to convert them into rectified images. Args: prediction: Predicted 2D Bezier mesh coordinates, shape (B, 2, H, W) original_images: List of original input tensors, each of shape (C, H_i, W_i). Images may have different sizes. scale: Scaling factor for output images (default: 255.0) Returns: List of dictionaries containing rectified images. Each dictionary has: - "images": Rectified image tensor of shape (H, W, 3) with dtype torch.uint8 and BGR channel order (suitable for OpenCV visualization) )devicerrrNr*rTr)r)dimsF)dtype non_blockingrrw)r[rqtensorfloatr enumeratendim unsqueezetorrrpermute grid_samplesqueezeflipuint8append)rSrrr image_listresultsioriginal_imageoriginal_heightoriginal_widthupsampled_meshrearranged_mesh rectifiedimages rB#post_process_document_rectificationz7UVDocImageProcessor.post_process_document_rectifications&/**  U5\\*2CDDD!*:!6!6 . . A~"a''!/!9!9!!rDrBrtrtsCJIC ( (D!*H/ ^$/ /  /  /  / / DK'$./ 4;&-/ +/ j(4//  / / / / j 55L5el+5 5 d3 $% & 555555rDrtcXeZdZdZ dded ed ed ed ed edededeffd ZxZS)UVDocConvLayerz>-00 66  66 %0  M22 rD)r*rr*Fr) rVrWrXrYr^r`r\rrqrrrarbs@rBrr/s44  0V0V0V0V 0V  0V  0V0V0V0V0V0V0V0V0VdU\elrDrcFeZdZdZfdZdejdejfdZxZS)UVDocResNetStagez3A ResNet stage containing multiple residual blocks.c t|j|}tjg|_|D]A\}}}}|jt|||rdnd|dz|||jBdS)Nrr*)rrrrrrr;) rPrr0r ModuleListlayersrrr;) rSconfig stage_indexrArrrrrUs rBrzUVDocResNetStage.__init__ps &{3mB'' ?E   ;Kx K  " +!- *111$qL%) & 2      rDrrc0|jD] }||}|Sr)r)rSrlayers rBrzUVDocResNetStage.forwards*[ 1 1E!E-00MMrD rVrWrXrYrrqrrrarbs@rBrrmsd==$U\elrDrcFeZdZdZfdZdejdejfdZxZS) UVDocResNetz$Initial resnet_head and resnet_down.c Fttjg|_t t |jD]]}|jt|j|d|j|d|j d|j dz^tjg|_ t t |j D],}t||}|j |-dS)Nrr*r)rrr;rr) rPrrrr)rLr?rrr; resnet_downr0r)rSrrrrGrUs rBrzUVDocResNet.__init__s =,,s6-..//  A   # # & 21 5a 8!'!3A!6q!9 & 2".!3     =,, V%:!;!;<< + +K$V[99E   # #E * * * * + +rDrrcZ|jD] }||}|jD] }||}|Sr)r)r)rSrheadrGs rBrzUVDocResNet.forwardsN$ 0 0D D//MM% 1 1E!E-00MMrDrrbs@rBrrsd..+++++&U\elrDrcVeZdZdZfdZdejdeedejfdZ xZ S)UVDocBridgeBlockzDBridge module with dilated convolutions for long-range dependencies.c ttjg|_|j|}|D]0\}}|jt||||1dS)N)rr)rPrrrblocksr:rr)rSr bridge_indexbridgerrrUs rBrzUVDocBridgeBlock.__init__s mB'' %l3%+ n n !K K  ~k;PXcklll m m m m n nrDrrTrc 0|jD] }||}|Sr)r)rSrrTblocks rBrzUVDocBridgeBlock.forwards, [ 1 1E!E-00MMrD) rVrWrXrYrrqrrrrrarbs@rBrrs}NNnnnnn|+,  rDrcFeZdZdZfdZdejdejfdZxZS)UVDocPointPositions2DzDModule for predicting 2D point positions for document rectification.c tt|jdd|jdd|jd|jdz|j|j|_tj |jdd|jdd|jd|jdz|j|_ dS)Nrr*r)rrr;rrrjr)rrr;rrrj) rPrrrlr;rjrhrrrconv_uprSrrUs rBrzUVDocPointPositions2D.__init__s '4Q7:5a8;*&!+,(   y4Q7:5a8;*&!+,     rDrrcZ||}||}|Sr)rr)rSrs rBrzUVDocPointPositions2D.forwards*}55  ]33 rDrrbs@rBrrsdNN     ,U\elrDrcJeZdZdZdeiZejdZdS)UVDocPreTrainedModelTrctj||t|tjr|dSdS)zInitialize the weights.N)r _init_weights isinstancerPReLUreset_parameters)rSmodules rBrz"UVDocPreTrainedModel._init_weightssM %dF333 fbh ' ' &  # # % % % % % & &rDN) rVrWrXsupports_gradient_checkpointingr_can_record_outputsrqno_gradrr>rDrBrrsK&*#)U]__&&_&&&rDrcreZdZfdZeedejdee dejfdZ xZ S) UVDocBridgec<t|tjg|_t t |jD]*}|jt||+| dSr) rPrrrrrLr?r:rr post_init)rSrrrUs rBrzUVDocBridge.__init__s    mB'' !#f&:";";<< G GL K  / EE F F F F rDrrTrc L|jD] }||}t|S)N)last_hidden_state)rr)rSrrTrfeatures rBrzUVDocBridge.forwards7[ + +EeM**GG-HHHHrD) rVrWrXrrrrqrrrrrarbs@rBrrs I|I+,I  III_ IIIIIrDrz6 UVDoc backbone model for feature extraction. ) custom_introc eZdZdZdZdeffd Zeee de j de e defdZxZS) UVDocBackboneFbackbonercNt||jddg}|jD]#}||dd$||_t ||_t||_ | dS)Nrrr*) rPrr)r:r num_featuresrresnetrrr)rSrrrGrUs rBrzUVDocBackbone.__init__s    *2.r23 ) - -E   a , , , ,(!&)) !&))  rDrrTrc ||}|j|fi|}d}t|jD]\}}||jvr||j|fz } t ||jS)Nr>) feature_mapsr)rrrrMrJrr)rSrrTroutputsrrHrGs rBrzUVDocBackbone.forwards L11 $+m66v66 #D$455 > >JC)))!6s!; == %!/    rD)rVrWrXhas_attentionsbase_model_prefixr$rrr rrq FloatTensorrrrrrarbs@rBr r s N" 2        ' +,      ^!      rDr c\eZdZfdZdejdeedejjfdZxZ S) UVDocHeadctt|jj|_t |jd|jz|jddddd|_t||_ dS)Nrr*)rrr;rrr) rPrr?rfr:num_bridge_layersrrkrrlrs rBrzUVDocHead.__init__)s !$V%;%I!J!J ./2T5KK03 ! ! ! &;6%B%B"""rDrrTrc Z||}||}|Sr)rkrl)rSrrTs rBrzUVDocHead.forward8s0 --m<< 22=AA rD) rVrWrXrrqrrrrrarbs@rBrr(sz C C C C C|+,   rDrz The model takes raw document images (pixel values) as input, processes them through the UVDoc backbone to predict spatial transformation parameters, and outputs the rectified (corrected) document image tensor. c eZdZdeffd Zeedejde e de eje zfdZ xZS) UVDocModelrct|t|j|_t ||_|dSr)rPrr rfr rrrrs rBrzUVDocModel.__init__IsR    %f&<== f%%  rDrrTrc |j|fi|}tj|jd}|j|fi|}t ||jS)Nr*)dim)rr)r rqcatrrrr)rSrrTbackbone_outputs fused_outputsrs rBrzUVDocModel.forwardPsp)4=@@@@ "2"?QGGG %DIm>>v>>-/*8    rD)rVrWrXrdrrrrqrrrr_rrrarbs@rBrrBs{  '  +,   u !$B B    ^      rDr)rr r$rtrdrr)Gcollections.abcrrqtorch.nnrtorch.nn.functional functionalrhuggingface_hub.dataclassesr activationsrbackbone_utilsrrr r configuration_utilsr feature_extraction_utilsr image_processing_backendsr image_transformsrr image_utilsrrmodeling_layersrmodeling_outputsrrmodeling_utilsrprocessing_utilsrutilsrrr utils.genericrrutils.import_utilsrutils.output_capturingrautorpp_lcnet.modeling_pp_lcnetr 0pp_ocrv5_server_det.modeling_pp_ocrv5_server_detr!r$rdrtrModulerrrrrrrr rr__all__r>rDrBr=sN %$$$$$ ......!!!!!! 433333444444;;;;;;EEEEEEEE77777777999999NNNNNNNN------&&&&&&IIIIIIIIIICCCCCCCC******555555999999^^^^^^;<<<K(K(K(K(K(-/?K(K(=<K(\;<<<((((("((=<(> :lllll,lll^     %   :;;;;;;;;|ry6")<1(BI> & & & & &: & & &IIIII&III( $ $ $ $ $ M#7$ $  $ N 4      %    2   rD