^jHddlZddlmZddlZddlmZddlmcmZddl m Z ddl m Z mZddlmZmZddlmZddlmZmZdd lmZdd lmZdd lmZmZmZdd lm Z dd l!m"Z"ddl#m$Z$GddeZ%eeGddeZ&GddejNZ(GddejNZ)Gdde%Z*GddejNZ+GddeZ,GddejNZ-Gd d!ejNZ.Gd"d#ejNZ/Gd$d%e%Z0ed&'Gd(d)e%Z1gd*Z2y)+N) dataclass)initialization)ACT2CLSACT2FN)filter_output_hidden_states load_backbone)GradientCheckpointingLayer)BaseModelOutputBaseModelOutputWithNoAttention)PreTrainedModel)Unpack)TransformersKwargsauto_docstringcan_return_tuple)merge_with_config_defaults)capture_outputs) SLANetConfigcfeZdZUeed<dZdZdZdZgZ e jfdZ xZ S)SLANetPreTrainedModelconfigbackbone pixel_values)imageTct||t|tjr|j dkDr"dt j|j z nd}tj|j| |tj|j| ||j"tj|j| ||j"tj|j| |t|trdt j|jj dzz }|j fD]}|j#D]n}t|tj$stj|j&| ||j(Mtj|j(| |pyy)zInitialize the weightsrg?N)super _init_weights isinstancennGRUCell hidden_sizemathsqrtinituniform_ weight_ih weight_hhbias_ihbias_hh SLANetSLAHeadrstructure_generatorchildrenLinearweightbias)selfmodulestd generatorlayer __class__s u/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/transformers/models/slanet/modeling_slanet.pyrz#SLANetPreTrainedModel._init_weights2sf f% fbjj )9?9K9Ka9O# &"4"455UVC MM&**SD# 6 MM&**SD# 6~~) fnnsdC8~~) fnnsdC8 fm , $++"9"9C"?@@C$88: A &//1AE!%3 ellSD#> ::1 MM%**sdC@ A A -)__name__ __module__ __qualname__r__annotations__base_model_prefixmain_input_nameinput_modalitiessupports_gradient_checkpointing_keep_in_fp32_modules_stricttorchno_gradr __classcell__r6s@r7rr*sB "$O!&*##% U]]_AAr8rcbeZdZUdZdZej dzed<dZej dzed<y)SLANetForTableRecognitionOutputak head_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Hidden-states of the SLANetSLAHead at each prediction step, varies up to max `self.config.max_text_length` states (depending on early exits). head_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Attentions of the SLANetSLAHead at each prediction step, varies up to max `self.config.max_text_length` attentions (depending on early exits). Nhead_hidden_stateshead_attentions) r9r:r;__doc__rHrB FloatTensorr<rIr8r7rGrGMs448))D0704OU&&-4r8rGc xeZdZfdZdej dej dej deefdZxZ S)SLANetAttentionGRUCellct|tj||d|_tj|||_tj|dd|_tj||z||_y)NF)r0r) r__init__r r.input_to_hiddenhidden_to_hiddenscorer!rnn)r1 input_sizer"num_embeddingsr6s r7rPzSLANetAttentionGRUCell.__init__\sa !yy[uM " +{ CYY{AE: ::j>9;Gr8 prev_hidden batch_hidden char_onehotskwargsc |j|}|j|jd}||z}tj|}|j |}t j|dtjj|j}|jdd}tj||jd} tj| |gd} |j| |} | |fS)Nrdimdtype)rQrR unsqueezerBtanhrSFsoftmaxfloat32tor^ transposematmulsqueezecatrT) r1rWrXrYrZbatch_hidden_projprev_hidden_projattention_scores attn_weightscontextconcat_context hidden_statess r7forwardzSLANetAttentionGRUCell.forwardes!00>00=GGJ,/?? ::&67::&67yy!1q NQQRbRhRhi #--a3 ,,|\:BB1EG\#:A>= l**r8) r9r:r;rPrBrKrrrqrDrEs@r7rNrN[sLH+&&+''+'' + +, +r8rNc&eZdZdfd ZdZxZS) SLANetMLPct|tj|||_tj|||_|tj |_yt||_yN) rrPr r.fc1fc2Identityract_fn)r1r" out_channels activationr6s r7rPzSLANetMLP.__init__}sR 99[+699[,7'1'9bkkm wz?R?T r8cl|j|}|j|}|j|}|Sru)rvrwryr1rps r7rqzSLANetMLP.forwards2/ /  M2 r8ru)r9r:r;rPrqrDrEs@r7rsrs|sU r8rsc eZdZdeiZ d dedzffd Zeee d de jde jdzde efdZxZS) r+ attentionsNrc t||t|j|j|j |_t|j|j |_|jyru) rrPrNpost_conv_out_channelsr"rzstructure_attention_cellrsr, post_init)r1rrZr6s r7rPzSLANetSLAHead.__init__s_  (>  ) )6+=+=v?R?R) %$-V-?-?ATAT#U  r8rptargetsrZc 6tj|jd|jjftj |j }tj|jdgtj|j }g}g}t|jjdzD]}tj||jjj} |j||j| \}}|j|} | j!d}|j#| |j#|tj$|dj'|jjdz j)dj+sntj,tj$|ddtj j/|j0} t3| |S) Nr)r^device)sizer^rrr]r\last_hidden_staterp)rBzerosshaperr"rdrlongrangemax_text_lengthrbone_hotrzfloatrr,argmaxappendstackeqanyallrcrer^r ) r1rprrZfeaturespredicted_charsstructure_preds_liststructure_ids_list_embedding_featurestructure_stepstructure_predss r7rqzSLANetSLAHead.forwards;;   #T[[%<%< =U]][h[o[o  ++M,?,?,B+C5::^k^r^rs!t{{22Q67 A ! /4;;;S;S T Z Z \ 77-BUBUBWYjkKHa!55h?N,333:O ' ' 7  % %o 6{{-15889Q9QTU9UVZZ[]^bbd ))EKK0D!$LRT\a\i\ijmm    Pdeer8ru)r9r:r;rN_can_record_outputsdictrPrrrrBrKTensorrrrqrDrEs@r7r+r+s, # t    (,f((f$f+, f! fr8r+ceZdZ d dedededededeeeefzdedeffd Zd ejd ejfd Z xZ S)SLANetConvLayer in_channelsrz kernel_sizestrider0dilationgroupsr{c t |tj|||||dz||||_tj ||_|t||_ ytj|_ y)Nr_)rrzrrpaddingr0rr) rrPr Conv2d convolution BatchNorm2d normalizationrrxr{) r1rrzrrr0rrr{r6s r7rPzSLANetConvLayer.__init__sn 99#%#1$   ^^L90:0F&,BKKMr8rpreturncl|j|}|j|}|j|}|Sru)rrr{r}s r7rqzSLANetConvLayer.forwards6((7 **=9  6 r8)rrFrr hardswish) r9r:r;intbooltuplestrrPrBrrqrDrEs@r7rrs *+%ZZZ Z  Z  Zc3h'ZZZ2U\\ellr8rc(eZdZdZfdZdZxZS)!SLANetDepthwiseSeparableConvLayerz Depthwise Separable Convolution Layer: Depthwise Conv -> Pointwise Conv Core component of lightweight models (e.g., MobileNet, PP-LCNet) that significantly reduces the number of parameters and computational cost. ct|t||||||j|_t j |_t|d|d|j|_y)N)rrzrrrr{r)rrrzrr{) rrPr hidden_actdepthwise_convolutionr rxsqueeze_excitation_modulepointwise_convolution)r1rrzrrrr6s r7rPz*SLANetDepthwiseSeparableConvLayer.__init__sg %4#$#(( & "*,&%4#%(( & "r8cl|j|}|j|}|j|}|Sru)rrr)r1 hidden_states r7rqz)SLANetDepthwiseSeparableConvLayer.forwards811,? 55lC 11,? r8)r9r:r;rJrPrqrDrEs@r7rrs  4r8rcVeZdZfdZdej dej fdZxZS)SLANetBottleneckctt|t||d||_t |||d||_y)Nrrrzrr{)rrzrrr)rrPrconv1rconv2)r1rrzrr{rr6s r7rPzSLANetBottleneck.__init__ sC $#,AZd  7$%#   r8rprcJ|j|}|j|}|Sru)rrr}s r7rqzSLANetBottleneck.forwards$ =1  =1 r8)r9r:r;rPrBrKrqrDrEs@r7rrs' (U%6%65;L;Lr8rcdeZdZdZ dfd Zdej dej fdZxZS)SLANetCSPLayerz Cross Stage Partial (CSP) network layer. Similar in structure to DFineCSPRepLayer, but with a different forward computation. c Bt |t||z}t||d||_t||d||_td|z|d||_tjt|D cgc]} t|||||c} |_ ycc} w)Nr)r{r_) rrPrrrrconv3r ModuleListrr bottlenecks) r1rrrzr expansion num_blocksr{hidden_channelsrr6s r7rPzSLANetCSPLayer.__init__)s lY67$[/1Q[\ $[/1Q[\ $Q%8,V`a ==z* !/;PZ\bc   s9Brprc|j|}|j|}|jD] }||} tj||fd}|j |}|S)Nrr)rrrrBrir)r1rpresidual bottlenecks r7rqzSLANetCSPLayer.forward?sh::m, =1 ** 6J&}5M 6 =(";C  =1 r8)rg?rr r9r:r;rJrPrBrKrqrDrEs@r7rr$s: , U%6%6 5;L;L r8rcZeZdZdZfdZdej dej fdZxZS) SLANetCSPPANz; CSP-PAN: Path Aggregation Network with CSP layers c Lt ||j}|j}|j}|j }t jtt|Dcgc]}t|||d|c}|_ t jdd|_ t jtt|dz ddDcgc]}t||dz||||c}|_t jtt|dz Dcgc]}t!|||d| c}|_t jtt|dz Dcgc]}t||dz||||c}|_ycc}wcc}wcc}wcc}w) Nrrr_nearest) scale_factormoderr)rrr{)rrr)rrPrrcsp_kernel_sizecsp_num_blocksr rrlenrchannel_projectorUpsampleupsamplertop_down_blocksr downsamplesbottom_up_blocks) r1rin_channel_listrzr{rrirr6s r7rPzSLANetCSPPAN.__init__Qs 44 && ,, ..!# s?34   / 2[\is "  C !}}s?3a7B?  1$ +-)   ==s?3a78 2  +!    !# s?3a78  1$ +-)   ! K    s%FFF*F!rprcg}tt|jD])}|j|j|||+|dg}t |j t |ddD]_\}}|d}tj||jddd}|tj||gd} |j| att |} | d} t |j|j| ddD]-\} } }| | }| tj||gd} /| jdj!dd}|S) Nrr)rrrrrr_)rrrrziprreversedrb interpolaterrBrilistrrflattenrf)r1rpprojected_featuresidxtop_down_featurestop_down_blocklow_level_featurehigh_level_featureupsampled_feature fused_featurepyramid_featuresoutput_featuredownsample_layerbottom_up_blockdownsampled_features r7rqzSLANetCSPPAN.forwardsT3345 WC  % %&Ad&<&"B ,UYY8KM_7`fg-hiN  j '..q1;;AqA r8rrEs@r7rrLs-= ~U%6%65;L;Lr8rc eZdZdeffd Zeedejde e de eje zfdZ xZS)SLANetBackbonerct||t||_t ||jj dd|_|jy)Nr_)rrPr vision_backboner num_features post_csp_panrr1rr6s r7rPzSLANetBackbone.__init__sK  ,V4(1E1E1R1RSTSU1VW r8rprZrc |j|fi|}|j|j}t||jS)Nr)rr feature_mapsr rp)r1rprZoutputss r7rqzSLANetBackbone.forwardsJ '$&&}??))'*>*>? -+!//  r8)r9r:r;rrPrrrBrKrrrr rqrDrEs@r7rrs`| ".. :@AS:T u  !$B B  r8rz SLANet Table Recognition model for table recognition tasks. Wraps the core SLANetPreTrainedModel and returns outputs compatible with the Transformers table recognition API. ) custom_introc eZdZdgZdeffd Zeedejde e de eje zfdZxZS)SLANetForTableRecognitionnum_batches_trackedrct||t||_t ||_|j y)N)r)rrPrrr+headrrs r7rPz"SLANetForTableRecognition.__init__s2  &f5 !0  r8rrZrc |j|fi|}|j|jfi|}t|j|j|j|j S)N)rrprHrI)rr rrGrpr)r1rrZr head_outputss r7rqz!SLANetForTableRecognition.forwardse  $-- 77 tyy!:!:EfE .*<<!//+99(33   r8)r9r:r;_keys_to_ignore_on_load_missingrrPrrrBrKrrrrGrqrDrEs@r7rrsk(=&=#|   !--  9?@R9S  u  !$C C    r8r)rrr+r)3r# dataclassesrrBtorch.nnr torch.nn.functional functionalrbrr% activationsrrbackbone_utilsrr modeling_layersr modeling_outputsr r modeling_utilsr processing_utilsrutilsrrr utils.genericrutils.output_capturingrconfiguration_slanetrrrGModulerNrsr+rrrrrrr__all__rLr8r7rs/, ! &*H9O-&II75. AO AF  5&D 5  5+RYY+B  1f)1fhbiiB&(B&Rryy8%RYY%P]299]@ * *  5   2 dr8