`idZddlmZddlZddlZddlZddlZddl m Z ddl m Z m Z mZmZmZmZddlmZmZmZmZmZmZmZmZmZmZddlmZmZm Z dd l!m"Z"m#Z#m$Z$m%Z%d d l&m'Z'e$j(e)Z*d Z+Gd dej,j-Z.Gddej,j-Z/Gddej,j-Z0Gddej,j-Z1Gddej,j-Z2Gddej,j-Z3Gddej,j-Z4Gddej,j-Z5Gddej,j-Z6Gdd ej,j-Z7Gd!d"ej,j-Z8Gd#d$ej,j-Z9eGd%d&ej,j-Z:Gd'd(eZ;d)Ze"d.e<Gd/d0e;eZ?e"d1e<Gd2d3e;eZ@e"d4e<Gd5d6e;eZAe"d7e<Gd8d9e;eZBgd:ZCdS);zTF 2.0 LayoutLM model.) annotationsN)get_tf_activation)+TFBaseModelOutputWithPastAndCrossAttentions.TFBaseModelOutputWithPoolingAndCrossAttentionsTFMaskedLMOutputTFQuestionAnsweringModelOutputTFSequenceClassifierOutputTFTokenClassifierOutput) TFMaskedLanguageModelingLossTFModelInputTypeTFPreTrainedModelTFQuestionAnsweringLossTFSequenceClassificationLossTFTokenClassificationLossget_initializerkeraskeras_serializable unpack_inputs)check_embeddings_within_bounds shape_liststable_softmax)add_start_docstrings%add_start_docstrings_to_model_forwardloggingreplace_return_docstrings)LayoutLMConfigrcBeZdZdZdfd ZddZ dddZxZS)TFLayoutLMEmbeddingszGConstruct the embeddings from word, position and token_type embeddings.configrc Rtjdi|||_|j|_|j|_|j|_|j|_tj |j d|_ tj |j |_dS)N LayerNormepsilonnamerate)super__init__r! hidden_sizemax_position_embeddingsmax_2d_position_embeddingsinitializer_rangerlayersLayerNormalizationlayer_norm_epsr#Dropouthidden_dropout_probdropoutselfr!kwargs __class__s /home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/transformers/models/layoutlm/modeling_tf_layoutlm.pyr+zTFLayoutLMEmbeddings.__init__;s""6""" !-'-'E$*0*K'!'!988AV]h8ii|++1K+LL Nctjd5|d|jj|jgt |j|_dddn #1swxYwYtjd5|d|jj |jgt |j|_ dddn #1swxYwYtjd5|d|j |jgt |j|_ dddn #1swxYwYtjd5|d|j |jgt |j|_dddn #1swxYwYtjd5|d|j |jgt |j|_dddn #1swxYwYtjd 5|d|j |jgt |j|_dddn #1swxYwYtjd 5|d|j |jgt |j|_dddn #1swxYwY|jrdSd |_t'|d d`tj|jj5|jdd|jjgddddS#1swxYwYdSdS) Nword_embeddingsweight)r&shape initializertoken_type_embeddings embeddingsposition_embeddingsx_position_embeddingsy_position_embeddingsh_position_embeddingsw_position_embeddingsTr#)tf name_scope add_weightr! vocab_sizer,rr/r>type_vocab_sizerAr-rCr.rDrErFrGbuiltgetattrr#r&buildr7 input_shapes r:rOzTFLayoutLMEmbeddings.buildFsR ], - -  //{-t/?@+D,BCC*DK               ]2 3 3  )-!{2D4DE+D,BCC*9**D &               ]0 1 1  '+!3T5EF+D,BCC(7((D $               ]2 3 3  )-!68HI+D,BCC*9**D &               ]2 3 3  )-!68HI+D,BCC*9**D &               ]2 3 3  )-!68HI+D,BCC*9**D &               ]2 3 3  )-!68HI+D,BCC*9**D &                :  F 4d + + 7t~233 L L$$dD$+2I%JKKK L L L L L L L L L L L L L L L L L L 8 7sAA""A&)A&ACCC-rfill expand_dimsrangerDrE IndexErrorrFrGrCrAr#r5)r7rRrTrUrVrWrXrQleft_position_embeddingsupper_position_embeddingsright_position_embeddingslower_position_embeddingserFrGposition_embedstoken_type_embedsfinal_embeddingss r:callzTFLayoutLMEmbeddings.calls6%-*?*?@  *9dk6L M M MIT[)LLLM //4  !W+Q???N  >"(+b/*R*R*RYZ[[  >"(+b/*R*R*RYZ[[ <7;!,A666D b')y1KTRSRSRSUVUVUVXYRY]'['[ $(* $2LdSTSTSTVWVWVWYZSZm(\(\ %(* $2LdSTSTSTVWVWVWYZSZm(\(\ %(* $2LdSTSTSTVWVWVWYZSZm(\(\ % % b b bYZZ`a a b " $*Dd111aaaQR7mVZ[\[\[\^_^_^_ab[bVcFc d d " $*Dd111aaaQR7mVZ[\[\[\^_^_^_ab[bVcFc d d)4+C\ZZZIT-GQ_```   ' '(  ( (  ( (  ($ $$ $  >>1A>BB<  b."<&q)I(+K3NN  b11$((BW(2X2XZdeeI33DJJF[J4\4\^hiiK3NN  '11$((-(2P2PR\]]I33DJJmJ4T4TV`aaK >!#4i"@qIIII)^A%6 $D1MMMKK11$((-(2P2PR\]]I33DJJmJ4T4TV`aaK//0A:NN ? 6(5N9[)NNN WT,4D4J K K K9%5r::  %!v&6GG )0@rJJJ,,o,QQ   k/9EEO9_kBB<(8|||LLL:-=jRTVZVhEijjj9Jc#_55QaPc ? 2 11Gr;Nc|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|dd]tj|j j5|j dd|jjgdddn #1swxYwYt|dd`tj|j j5|j dd|jjgddddS#1swxYwYdSdS)NTrrrb) rMrNrHrIrr&rOr!r,rrbrPs r:rOzTFLayoutLMSelfAttention.build4s: :  F 4$ ' ' 3tz// H H   $dk.E!FGGG H H H H H H H H H H H H H H H 4 % % 1tx}-- F FdDK,CDEEE F F F F F F F F F F F F F F F 4$ ' ' 3tz// H H   $dk.E!FGGG H H H H H H H H H H H H H H H H H H 4 3s6(A44A8;A8.(C""C&)C&(EEErz)rr[rrrZr[F)rr[rr[rr[rr[rr[rrrrYrXrYrZrr{)r|r}r~r+rryrOrrs@r:rrs87777 OOOOOb H H H H H H H Hr;rc2eZdZdfd Zddd Zdd ZxZS)TFLayoutLMSelfOutputr!rc ntjdi|tj|jt |jd|_tj |j d|_ tj |j |_||_dSNdenserr#r$r'r)r*r+rr0rr,rr/rr1r2r#r3r4r5r!r6s r:r+zTFLayoutLMSelfOutput.__init__E""6"""\''$Ia9b9bip(   88AV]h8ii|++1K+LL  r;Frr[ input_tensorrXrYrZc||}|||}|||z}|SNrirkrr5r#r7rrrXs r:ryzTFLayoutLMSelfOutput.callOH - 88  MH MM ml.JKK r;Nc|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|dd`tj|j j5|j dd|jjgddddS#1swxYwYdSdSNTrr# rMrNrHrIrr&rOr!r,r#rPs r:rOzTFLayoutLMSelfOutput.buildV :  F 4$ ' ' 3tz// H H   $dk.E!FGGG H H H H H H H H H H H H H H H 4d + + 7t~233 L L$$dD$+2I%JKKK L L L L L L L L L L L L L L L L L L 8 7$(A44A8;A8.(C##C'*C'rzrrr[rr[rXrYrZr[r{r|r}r~r+ryrOrrs@r:rrDr L L L L L L L Lr;rc:eZdZdfd ZdZ dddZddZxZS)TFLayoutLMAttentionr!rc tjdi|t|d|_t |d|_dS)Nr7r&outputr))r*r+rself_attentionr dense_outputr6s r:r+zTFLayoutLMAttention.__init__dsP""6"""5f6JJJ0hGGGr;ctr{NotImplementedError)r7headss r: prune_headszTFLayoutLMAttention.prune_headsjs!!r;Frr[rrrrrrrrYrXrZc |||||||||} || d||} | f| ddz} | S)NrrrrrrrrXrrrrXr)rr) r7rrrrrrrrX self_outputsrrs r:ryzTFLayoutLMAttention.callms{**&)"7#9)/+    ,,&q/ x-  $% QRR(88r;Nc|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|ddStj|jj5|jdddddS#1swxYwYdSdS)NTrr)rMrNrHrIrr&rOrrPs r:rOzTFLayoutLMAttention.buildsX :  F 4)4 0 0 <t2788 0 0#))$/// 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 4 . . :t0566 . .!''--- . . . . . . . . . . . . . . . . . . ; :$A''A+.A+!C  C C rzr)rr[rr[rr[rr[rr[rrrrYrXrYrZrr{)r|r}r~r+rryrOrrs@r:rrcsHHHHHH """: . . . . . . . .r;rc0eZdZd fd Zd dZd d ZxZS) TFLayoutLMIntermediater!rc Dtjdi|tj|jt |jd|_t|j trt|j |_ n |j |_ ||_dS)Nrrr))r*r+rr0rintermediate_sizerr/r isinstance hidden_actstrrintermediate_act_fnr!r6s r:r+zTFLayoutLMIntermediate.__init__s""6"""\''*vOg?h?hov(    f' - - 9'89J'K'KD $ $'-'8D $ r;rr[rZc\||}||}|SNri)rrr7rs r:ryzTFLayoutLMIntermediate.calls. - 88 00?? r;Nc|jrdSd|_t|dd`tj|jj5|jdd|jjgddddS#1swxYwYdSdSNTr rMrNrHrIrr&rOr!r,rPs r:rOzTFLayoutLMIntermediate.build :  F 4$ ' ' 3tz// H H   $dk.E!FGGG H H H H H H H H H H H H H H H H H H 4 3(A55A9<A9rzrr[rZr[r{rrs@r:rrsm       HHHHHHHHr;rc2eZdZdfd Zddd Zdd ZxZS)TFLayoutLMOutputr!rc ntjdi|tj|jt |jd|_tj |j d|_ tj |j |_||_dSrrr6s r:r+zTFLayoutLMOutput.__init__rr;Frr[rrXrYrZc||}|||}|||z}|Srrrs r:ryzTFLayoutLMOutput.callrr;Nc|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|dd`tj|j j5|j dd|jj gddddS#1swxYwYdSdSr) rMrNrHrIrr&rOr!rr#r,rPs r:rOzTFLayoutLMOutput.builds :  F 4$ ' ' 3tz// N N   $dk.K!LMMM N N N N N N N N N N N N N N N 4d + + 7t~233 L L$$dD$+2I%JKKK L L L L L L L L L L L L L L L L L L 8 7rrzrrr{rrs@r:rrrr;rc4eZdZdfd Z dddZddZxZS)TFLayoutLMLayerr!rc Ztjdi|t|d|_|j|_|j|_|jr/|jst |dt|d|_t|d|_ t|d|_ dS)N attentionrz> should be used as a decoder model if cross attention is addedcrossattention intermediaterr)) r*r+rrradd_cross_attentionrrrrr bert_outputr6s r:r+zTFLayoutLMLayer.__init__s""6""",V+FFF +#)#=  # U? j D!h!h!hiii"5fCS"T"T"TD 26OOO+FBBBr;Frr[rrrrSrrtuple[tf.Tensor] | NonerrYrXrZrc ,| |ddnd} ||||dd| ||} | d} |jr| dd} | d} n | dd} d}|jrr|pt|dstd|d| |d dnd}|| |||||||}|d} | |ddz} |d}| |z} || }||| | }|f| z} |jr| | fz} | S) Nrh)rrrrrrrrXrrr_rz'If `encoder_hidden_states` are passed, z` has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`rr)rrhasattrrrrr)r7rrrrrrrrXself_attn_past_key_valueself_attention_outputsrrpresent_key_valuecross_attn_present_key_valuecross_attn_past_key_valuecross_attention_outputsintermediate_output layer_outputs r:ryzTFLayoutLMLayer.calls:H9S>"1"#5#5Y] !%&)"&#'3/"0 " " 2!4 ? 1,QrT2G 6r :  ,QRR0G'+$ ? Q4@4!122  DdDDD @N?Yrss(;(;_c %&*&9&9--#&;'=8"3!': ' ' # 7q9  7" ==G,C2+F ( 14P P "//>N/OO''-z.TFLayoutLMEncoder.__init__..@s,lllqof>a>>BBBlllr;r))r*r+r!ronum_hidden_layerslayerr6s ` r:r+zTFLayoutLMEncoder.__init__=sS""6""" llllERXRjLkLklll r;Frr[rrrrSrpast_key_valuestuple[tuple[tf.Tensor]] | None use_cache bool | NonerrYoutput_hidden_states return_dictrXrZ>TFBaseModelOutputWithPastAndCrossAttentions | tuple[tf.Tensor]c | rdnd} |rdnd} |r|jjrdnd}|rdnd}t|jD]p\}}| r| |fz} |||nd}|||||||||| }|d}|r ||dfz }|r&| |dfz} |jjr| ||dfz}q| r| |fz} | st d|| | |fDSt ||| | |S) Nr)rrr_rrhc3K|]}||V dSr{r))rvs r: z)TFLayoutLMEncoder.call..ts1ghgtgtgtgtgtr;)last_hidden_staterr attentionscross_attentions)r!r enumeratertupler)r7rrrrrrrrrr rXall_hidden_statesall_attentionsall_cross_attentionsnext_decoder_cacher layer_moduler layer_outputss r:ryzTFLayoutLMEncoder.callBs#7@BBD0:d%6d4;;Zdrr`d#,6RR$(44 V VOA|# I$58H$H!3B3N_Q//TXN(L+-#A,&;'=-"3!   M*!,M ;"}R'8&::"  V!/=3C2E!E;2V7L7X+?=QRCSBU+U(  E 1]4D D  )+r#rs r:ryz&TFLayoutLMPredictionHeadTransform.callsA - 88 --m<< m<< r;Nc|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|dd`tj|j j5|j dd|jjgddddS#1swxYwYdSdSrrrPs r:rOz'TFLayoutLMPredictionHeadTransform.buildrrrzrr{rrs@r:r<r<sm" L L L L L L L Lr;r<cPeZdZdfd ZddZdd Zdd ZddZddZddZ xZ S)TFLayoutLMLMPredictionHeadr!rinput_embeddingskeras.layers.Layerc tjdi|||_|j|_t |d|_||_dS)N transformrr))r*r+r!r,r<rFrCr7r!rCr8r9s r:r+z#TFLayoutLMLMPredictionHead.__init__sV""6""" !-:6 TTT!1r;Nc@||jjfddd|_|jrdSd|_t |ddSt j|jj 5|j dddddS#1swxYwYdSdS)NzerosTbias)r?r@ trainabler&rF) rJr!rKrJrMrNrHrIrFr&rOrPs r:rOz TFLayoutLMLMPredictionHead.buildsOO4;+A*CQXdhouOvv :  F 4d + + 7t~233 + +$$T*** + + + + + + + + + + + + + + + + + + 8 7s)BBBrZc|jSr{)rCr7s r:get_output_embeddingsz0TFLayoutLMLMPredictionHead.get_output_embeddingss $$r;rb tf.Variablec\||j_t|d|j_dSNr)rCr>rrKr7rbs r:set_output_embeddingsz0TFLayoutLMLMPredictionHead.set_output_embeddingss+',$+5e+<+r!rKnnbias_addrJ)r7r seq_lengths r:ryzTFLayoutLMLMPredictionHead.calls]CC  ..q1  -DDT?UVVV  MT5J5Q_cddd  -JPTP[Pf?ghhh ]KK r;r!rrCrDr{rZrDrbrO)rZrTr) r|r}r~r+rOrNrSrVrXryrrs@r:rBrBs 1 1 1 1 1 1++++%%%%@@@@####>>>>r;rBc0eZdZd fd Zd d Zdd ZxZS)TFLayoutLMMLMHeadr!rrCrDc htjdi|t||d|_dS)N predictionsrr))r*r+rBrerGs r:r+zTFLayoutLMMLMHead.__init__s?""6"""5f>NUbcccr;sequence_outputr[rZc2||}|S)Nr)re)r7rfprediction_scoress r:ryzTFLayoutLMMLMHead.calls ,,?,KK  r;Nc|jrdSd|_t|ddStj|jj5|jdddddS#1swxYwYdSdS)NTre)rMrNrHrIrer&rOrPs r:rOzTFLayoutLMMLMHead.build s :  F 4 - - 9t/455 - - &&t,,, - - - - - - - - - - - - - - - - - - : 9A((A,/A,r_)rfr[rZr[r{rrs@r:rcrcskdddddd !!!! --------r;rccxeZdZeZd#d$fd Zd%d Zd&d Zd Ze d'd(d!Z d)d"Z xZ S)*TFLayoutLMMainLayerTr!radd_pooling_layerrYc tjdi|||_t|d|_t |d|_|rt|dnd|_dS)NrBrencoderpoolerr)) r*r+r!r rBrror3rp)r7r!rmr8r9s r:r+zTFLayoutLMMainLayer.__init__ss""6""" .vLIII(i@@@ AR\&vH====X\ r;rZrDc|jSr{)rBrMs r:get_input_embeddingsz(TFLayoutLMMainLayer.get_input_embeddings!s r;rbrOc\||j_t|d|j_dSrQ)rBr>rrKrRs r:set_input_embeddingsz(TFLayoutLMMainLayer.set_input_embeddings$s)!&%/%6%6q%9"""r;ct)z Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base class PreTrainedModel r)r7heads_to_prunes r: _prune_headsz TFLayoutLMMainLayer._prune_heads(s "!r;NFrRTFModelInputType | NonerTnp.ndarray | tf.Tensor | NonerrVrUrrWrrrrrr rXATFBaseModelOutputWithPoolingAndCrossAttentions | tuple[tf.Tensor]c ||td|t|}n)|t|dd}ntd|tj|d}|tj|d}|tj|dgzd}||||||| }tj||ddd|df}tj||j }tjd |j }tjd |j }tj tj |||}|tdg|j j z}|||||ddd | | | | }|d}|j||nd}| s||f|ddzSt!|||j|j|jS)NzDYou cannot specify both input_ids and inputs_embeds at the same timer_z5You have to specify either input_ids or inputs_embedsrr`rrg)rRrTrUrVrWrXrg?gF) rrrrrrrrrr rXr)r& pooler_outputrr'r()rrrHrmrBrrrconstantrsubtractrr!rrorprrr'r()r7rRrTrrVrUrrWrrrrr rXrQembedding_outputextended_attention_maskone_cstten_thousand_cstencoder_outputsrfr9s r:ryzTFLayoutLMMainLayer.call/sU"  ]%>cdd d  "$Y//KK  &$]33CRC8KKTUU U  !W+Q???N  !W+Q???N <7 qc 1;;;D??%)' +  #%*^k!naQRT_`aTb=c"d"d#%'*AIYI_"`"`"`+c)9)?@@@;x7G7MNNN"$+bk'CZ.[.[]m"n"n  % %!>>I,,*2"7#' /!5#'   *!,FJkF] / BBBcg  $ #$ $ >-')7&1,=     r;cr|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|ddPtj|jj5|jddddn #1swxYwYt|ddStj|jj5|jdddddS#1swxYwYdSdS)NTrBrorp) rMrNrHrIrBr&rOrorprPs r:rOzTFLayoutLMMainLayer.builds :  F 4t , , 8t344 , ,%%d+++ , , , , , , , , , , , , , , , 4D ) ) 5t|011 ) ) ""4((( ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) 44 ( ( 4t{/00 ( ( !!$''' ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( 5 4s6A''A+.A+!CC C D**D.1D.)T)r!rrmrYr`ra NNNNNNNNNNNNF)rRrxrTryrryrVryrUryrryrWryrryrryrrrrr rrXrYrZrzr{) r|r}r~r config_classr+rrrtrwrryrOrrs@r:rlrls!L]]]]]]]::::""".2.28<8<6:377;?C@D)-,0#'c c c c ]c J ( ( ( ( ( ( ( (r;rlc:eZdZdZeZdZefdZxZ S)TFLayoutLMPreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. layoutlmcvtj}tjdtjd|d<|S)N)NNrgrT)r?rr&)r*input_signaturerH TensorSpecint32)r7 signaturer9s r:rz)TFLayoutLMPreTrainedModel.input_signatures3GG+ MrxV\]]] &r;) r|r}r~rrrbase_model_prefixpropertyrrrs@r:rrsY "L" Xr;raz This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.) This model is also a [keras.Model](https://www.tensorflow.org/api_docs/python/tf/keras/Model) subclass. Use it as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage and behavior. TensorFlow models and layers in `transformers` accept two formats as input: - having all inputs as keyword arguments (like PyTorch models), or - having all inputs as a list, tuple or dict in the first positional argument. The reason the second format is supported is that Keras methods prefer this format when passing inputs to models and layers. Because of this support, when using methods like `model.fit()` things should "just work" for you - just pass your inputs and labels in any format that `model.fit()` supports! If, however, you want to use the second format outside of Keras methods like `fit()` and `predict()`, such as when creating your own layers or models with the Keras `Functional` API, there are three possibilities you can use to gather all the input Tensors in the first positional argument: - a single Tensor with `input_ids` only and nothing else: `model(input_ids)` - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring: `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])` - a dictionary with one or several input Tensors associated to the input names given in the docstring: `model({"input_ids": input_ids, "token_type_ids": token_type_ids})` Note that when creating models and layers with [subclassing](https://keras.io/guides/making_new_layers_and_models_via_subclassing/) then you don't need to worry about any of this, as you can just pass inputs like you would to any other Python function! Args: config ([`LayoutLMConfig`]): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the [`~TFPreTrainedModel.from_pretrained`] method to load the model weights. a Args: input_ids (`Numpy array` or `tf.Tensor` of shape `({0})`): Indices of input sequence tokens in the vocabulary. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.__call__`] and [`PreTrainedTokenizer.encode`] for details. [What are input IDs?](../glossary#input-ids) bbox (`Numpy array` or `tf.Tensor` of shape `({0}, 4)`, *optional*): Bounding Boxes of each input sequence tokens. Selected in the range `[0, config.max_2d_position_embeddings- 1]`. attention_mask (`Numpy array` or `tf.Tensor` of shape `({0})`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) token_type_ids (`Numpy array` or `tf.Tensor` of shape `({0})`, *optional*): Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0, 1]`: - 0 corresponds to a *sentence A* token, - 1 corresponds to a *sentence B* token. [What are token type IDs?](../glossary#token-type-ids) position_ids (`Numpy array` or `tf.Tensor` of shape `({0})`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.max_position_embeddings - 1]`. [What are position IDs?](../glossary#position-ids) head_mask (`Numpy array` or `tf.Tensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. inputs_embeds (`tf.Tensor` of shape `({0}, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. training (`bool`, *optional*, defaults to `False`): Whether or not to use the model in training mode (some modules like dropout modules have different behaviors between training and evaluation). zbThe bare LayoutLM Model transformer outputting raw hidden-states without any specific head on top.ceZdZdfd Zeeedee e  dddZ ddZ xZ S) TFLayoutLMModelr!rcntj|g|Ri|t|d|_dS)Nrr)r*r+rlrr7r!rjr8r9s r:r+zTFLayoutLMModel.__init__sB3&333F333+FDDD r;batch_size, sequence_length output_typerNFrRrxrTryrrVrUrrWrrrrrr rXrZrzc F||||||||| | | |  }|S)a  Returns: Examples: ```python >>> from transformers import AutoTokenizer, TFLayoutLMModel >>> import tensorflow as tf >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/layoutlm-base-uncased") >>> model = TFLayoutLMModel.from_pretrained("microsoft/layoutlm-base-uncased") >>> words = ["Hello", "world"] >>> normalized_word_boxes = [637, 773, 693, 782], [698, 773, 733, 782] >>> token_boxes = [] >>> for word, box in zip(words, normalized_word_boxes): ... word_tokens = tokenizer.tokenize(word) ... token_boxes.extend([box] * len(word_tokens)) >>> # add bounding boxes of cls + sep tokens >>> token_boxes = [[0, 0, 0, 0]] + token_boxes + [[1000, 1000, 1000, 1000]] >>> encoding = tokenizer(" ".join(words), return_tensors="tf") >>> input_ids = encoding["input_ids"] >>> attention_mask = encoding["attention_mask"] >>> token_type_ids = encoding["token_type_ids"] >>> bbox = tf.convert_to_tensor([token_boxes]) >>> outputs = model( ... input_ids=input_ids, bbox=bbox, attention_mask=attention_mask, token_type_ids=token_type_ids ... ) >>> last_hidden_states = outputs.last_hidden_state ``` rRrTrrVrUrrWrrr rX)r)r7rRrTrrVrUrrWrrrrr rXrs r:ryzTFLayoutLMModel.call sEn--))%'/!5#   r;c|jrdSd|_t|ddStj|jj5|jdddddS#1swxYwYdSdS)NTr)rMrNrHrIrr&rOrPs r:rOzTFLayoutLMModel.buildgs :  F 4T * * 6t}122 * * ##D))) * * * * * * * * * * * * * * * * * * 7 6rjrzr)rRrxrTryrryrVryrUryrryrWryrryrryrrrrr rrXrrZrzr{)r|r}r~r+rrLAYOUTLM_INPUTS_DOCSTRINGformatrr_CONFIG_FOR_DOCryrOrrs@r:rrs EEEEEE **+D+K+KLi+j+jkkBQ` .2.28<8<6:377;?C@D)-,0#' %@@@@lk] @D********r;rz6LayoutLM Model with a `language modeling` head on top.ceZdZgdZd fd Zd!dZd"d Zeee d e e e  d#d$dZd%dZxZS)&TFLayoutLMForMaskedLM)rpcls.seq_relationshipzcls.predictions.decoder.weight nsp___clsr!rctj|g|Ri||jrtdt |dd|_t||jjd|_ dS)NzqIf you want to use `TFLayoutLMForMaskedLM` make sure `config.is_decoder=False` for bi-directional self-attention.Trrmr& mlm___cls)rCr&) r*r+rloggerwarningrlrrcrBmlmrs r:r+zTFLayoutLMForMaskedLM.__init__zs3&333F333    NN1    ,FdQ[\\\ $Vdm>V]hiiir;rZrDc|jjSr{)rrerMs r: get_lm_headz!TFLayoutLMForMaskedLM.get_lm_heads x##r;rctjdt|jdz|jjzdz|jjjzS)NzMThe method get_prefix_bias_name is deprecated. Please use `get_bias` instead./)warningswarn FutureWarningr&rrerMs r:get_prefix_bias_namez*TFLayoutLMForMaskedLM.get_prefix_bias_names= egtuuuy3.4tx7K7PPPr;rrNFrRrxrTryrrVrUrrWrrrr labelsrX#TFMaskedLMOutput | tuple[tf.Tensor]c $|||||||||| | |  } | d}||| }| dn|| |}| s|f| ddz}||f|zn|St||| j| jS)aV labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]` Returns: Examples: ```python >>> from transformers import AutoTokenizer, TFLayoutLMForMaskedLM >>> import tensorflow as tf >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/layoutlm-base-uncased") >>> model = TFLayoutLMForMaskedLM.from_pretrained("microsoft/layoutlm-base-uncased") >>> words = ["Hello", "[MASK]"] >>> normalized_word_boxes = [637, 773, 693, 782], [698, 773, 733, 782] >>> token_boxes = [] >>> for word, box in zip(words, normalized_word_boxes): ... word_tokens = tokenizer.tokenize(word) ... token_boxes.extend([box] * len(word_tokens)) >>> # add bounding boxes of cls + sep tokens >>> token_boxes = [[0, 0, 0, 0]] + token_boxes + [[1000, 1000, 1000, 1000]] >>> encoding = tokenizer(" ".join(words), return_tensors="tf") >>> input_ids = encoding["input_ids"] >>> attention_mask = encoding["attention_mask"] >>> token_type_ids = encoding["token_type_ids"] >>> bbox = tf.convert_to_tensor([token_boxes]) >>> labels = tokenizer("Hello world", return_tensors="tf")["input_ids"] >>> outputs = model( ... input_ids=input_ids, ... bbox=bbox, ... attention_mask=attention_mask, ... token_type_ids=token_type_ids, ... labels=labels, ... ) >>> loss = outputs.loss ```rr)rfrXNrrrhlossrrr')rrhf_compute_lossrrr')r7rRrTrrVrUrrWrrr rrXrrfrhrrs r:ryzTFLayoutLMForMaskedLM.calls~--))%'/!5#   "!* HH_xHXX~tt4+?+?vVg+?+h+h F')GABBK7F)-)9TGf$$v E$!/)     r;c|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|ddStj|jj5|jdddddS#1swxYwYdSdS)NTrr)rMrNrHrIrr&rOrrPs r:rOzTFLayoutLMForMaskedLM.buildsP :  F 4T * * 6t}122 * * ##D))) * * * * * * * * * * * * * * * 4 % % 1tx}-- % %t$$$ % % % % % % % % % % % % % % % % % % 2 1rrzr`)rZr NNNNNNNNNNNF)rRrxrTryrryrVryrUryrryrWryrrrrr rrryrXrrZrr{)r|r}r~"_keys_to_ignore_on_load_unexpectedr+rrrrrrrrrryrOrrs@r:rrps****& j j j j j j$$$$QQQQ**+D+K+KLi+j+jkk+;/ZZZ.2.28<8<6:377;)-,0#'04 %V V V V [Zlk]V p % % % % % % % %r;rz LayoutLM Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled output) e.g. for GLUE tasks. ceZdZgdZdgZdfd Zeee de e e  dddZ d dZxZS)!#TFLayoutLMForSequenceClassification)rrcls.predictionsrr5r!rcdtj|g|Ri||j|_t|d|_t j|j|_ t j |jt|j d|_ ||_dS)Nrrr' classifierrr*r+ num_labelsrlrrr0r3r4r5rrr/rr!rs r:r+z,TFLayoutLMForSequenceClassification.__init__s3&333F333 ++FDDD |++1K+LL ,,,#.v/GHH-    r;rrNFrRrxrTryrrVrUrrWrrrr rrXrZ-TFSequenceClassifierOutput | tuple[tf.Tensor]c P|||||||||| | |  } | d}||| }||}| dn|| |}| s|f| ddz}||f|zn|St ||| j| jS) an labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). Returns: Examples: ```python >>> from transformers import AutoTokenizer, TFLayoutLMForSequenceClassification >>> import tensorflow as tf >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/layoutlm-base-uncased") >>> model = TFLayoutLMForSequenceClassification.from_pretrained("microsoft/layoutlm-base-uncased") >>> words = ["Hello", "world"] >>> normalized_word_boxes = [637, 773, 693, 782], [698, 773, 733, 782] >>> token_boxes = [] >>> for word, box in zip(words, normalized_word_boxes): ... word_tokens = tokenizer.tokenize(word) ... token_boxes.extend([box] * len(word_tokens)) >>> # add bounding boxes of cls + sep tokens >>> token_boxes = [[0, 0, 0, 0]] + token_boxes + [[1000, 1000, 1000, 1000]] >>> encoding = tokenizer(" ".join(words), return_tensors="tf") >>> input_ids = encoding["input_ids"] >>> attention_mask = encoding["attention_mask"] >>> token_type_ids = encoding["token_type_ids"] >>> bbox = tf.convert_to_tensor([token_boxes]) >>> sequence_label = tf.convert_to_tensor([1]) >>> outputs = model( ... input_ids=input_ids, ... bbox=bbox, ... attention_mask=attention_mask, ... token_type_ids=token_type_ids, ... labels=sequence_label, ... ) >>> loss = outputs.loss >>> logits = outputs.logits ```rrrkriNrrhr)rr5rrr rr')r7rRrTrrVrUrrWrrr rrXrr9rrrs r:ryz(TFLayoutLMForSequenceClassification.calls~--))%'/!5#      MH MM  66~tt4+?+?vV\+?+]+] FY,F)-)9TGf$$v E)!/)     r;c|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|dd`tj|jj5|jdd|jj gddddS#1swxYwYdSdSNTrr rMrNrHrIrr&rOrr!r,rPs r:rOz)TFLayoutLMForSequenceClassification.buildjs :  F 4T * * 6t}122 * * ##D))) * * * * * * * * * * * * * * * 4t , , 8t344 M M%%tT4;3J&KLLL M M M M M M M M M M M M M M M M M M 9 8$A''A+.A+!(CCCrzr)rRrxrTryrryrVryrUryrryrWryrrrrr rrryrXrrZrr{)r|r}r~r_keys_to_ignore_on_load_missingr+rrrrrr rryrOrrs@r:rrs*s)r)r&'1l#      **+D+K+KLi+j+jkk+ETcddd.2.28<8<6:377;)-,0#'04 %W W W W edlk]W r M M M M M M M Mr;rz LayoutLM Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. ceZdZgdZdgZdfd Zeee de e e  dddZ d dZxZS)! TFLayoutLMForTokenClassificationrprrrrr5r!rcftj|g|Ri||j|_t|dd|_t j|j|_ t j |jt|j d|_ ||_dS)NTrrr'rrrrs r:r+z)TFLayoutLMForTokenClassification.__init__s3&333F333 ++FdQ[\\\ |++1K+LL ,,,#.v/GHH-    r;rrNFrRrxrTryrrVrUrrWrrrr rrXrZ*TFTokenClassifierOutput | tuple[tf.Tensor]c P|||||||||| | |  } | d}||| }||}| dn|| |}| s|f| ddz}||f|zn|St ||| j| jS) a labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`. Returns: Examples: ```python >>> import tensorflow as tf >>> from transformers import AutoTokenizer, TFLayoutLMForTokenClassification >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/layoutlm-base-uncased") >>> model = TFLayoutLMForTokenClassification.from_pretrained("microsoft/layoutlm-base-uncased") >>> words = ["Hello", "world"] >>> normalized_word_boxes = [637, 773, 693, 782], [698, 773, 733, 782] >>> token_boxes = [] >>> for word, box in zip(words, normalized_word_boxes): ... word_tokens = tokenizer.tokenize(word) ... token_boxes.extend([box] * len(word_tokens)) >>> # add bounding boxes of cls + sep tokens >>> token_boxes = [[0, 0, 0, 0]] + token_boxes + [[1000, 1000, 1000, 1000]] >>> encoding = tokenizer(" ".join(words), return_tensors="tf") >>> input_ids = encoding["input_ids"] >>> attention_mask = encoding["attention_mask"] >>> token_type_ids = encoding["token_type_ids"] >>> bbox = tf.convert_to_tensor([token_boxes]) >>> token_labels = tf.convert_to_tensor([1, 1, 0, 0]) >>> outputs = model( ... input_ids=input_ids, ... bbox=bbox, ... attention_mask=attention_mask, ... token_type_ids=token_type_ids, ... labels=token_labels, ... ) >>> loss = outputs.loss >>> logits = outputs.logits ```rrrkriNrrhr)rr5rrr rr')r7rRrTrrVrUrrWrrr rrXrrfrrrs r:ryz%TFLayoutLMForTokenClassification.callsz--))%'/!5#   "!*,,o,QQ88~tt4+?+?vV\+?+]+] FY,F)-)9TGf$$v E&!/)     r;c|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|dd`tj|jj5|jdd|jj gddddS#1swxYwYdSdSrrrPs r:rOz&TFLayoutLMForTokenClassification.buildrrrzr)rRrxrTryrryrVryrUryrryrWryrrrrr rrryrXrrZrr{)r|r}r~rrr+rrrrrr rryrOrrs@r:rrvs ***&(2l#      **+D+K+KLi+j+jkk+BQ`aaa.2.28<8<6:377;)-,0#'04 %U U U U balk]U n M M M M M M M Mr;ra LayoutLM Model with a span classification head on top for extractive question-answering tasks such as [DocVQA](https://rrc.cvc.uab.es/?ch=17) (a linear layer on top of the final hidden-states output to compute `span start logits` and `span end logits`). ceZdZgdZdfd Zeeede e e  dddZ d dZ xZS)!TFLayoutLMForQuestionAnsweringrr!rctj|g|Ri||j|_t|dd|_t j|jt|j d|_ ||_ dS)NTrr qa_outputsr) r*r+rrlrrr0rrr/rr!rs r:r+z'TFLayoutLMForQuestionAnswering.__init__s3&333F333 ++FdQ[\\\ ,,,#.v/GHH-    r;rrNFrRrxrTryrrVrUrrWrrrr start_positions end_positionsrXrZ1TFQuestionAnsweringModelOutput | tuple[tf.Tensor]c |||||||||| | |  }|d}||}tj|dd\}}tj|d}tj|d}d}| $| "d | i}| |d <||||f }| s||f|ddz}||f|zn|St ||||j|j S) a start_positions (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*): Labels for position (index) of the start of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. end_positions (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*): Labels for position (index) of the end of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. Returns: Examples: ```python >>> import tensorflow as tf >>> from transformers import AutoTokenizer, TFLayoutLMForQuestionAnswering >>> from datasets import load_dataset >>> tokenizer = AutoTokenizer.from_pretrained("impira/layoutlm-document-qa", add_prefix_space=True) >>> model = TFLayoutLMForQuestionAnswering.from_pretrained("impira/layoutlm-document-qa", revision="1e3ebac") >>> dataset = load_dataset("nielsr/funsd", split="train") >>> example = dataset[0] >>> question = "what's his name?" >>> words = example["words"] >>> boxes = example["bboxes"] >>> encoding = tokenizer( ... question.split(), words, is_split_into_words=True, return_token_type_ids=True, return_tensors="tf" ... ) >>> bbox = [] >>> for i, s, w in zip(encoding.input_ids[0], encoding.sequence_ids(0), encoding.word_ids(0)): ... if s == 1: ... bbox.append(boxes[w]) ... elif i == tokenizer.sep_token_id: ... bbox.append([1000] * 4) ... else: ... bbox.append([0] * 4) >>> encoding["bbox"] = tf.convert_to_tensor([bbox]) >>> word_ids = encoding.word_ids(0) >>> outputs = model(**encoding) >>> loss = outputs.loss >>> start_scores = outputs.start_logits >>> end_scores = outputs.end_logits >>> start, end = word_ids[tf.math.argmax(start_scores, -1)[0]], word_ids[tf.math.argmax(end_scores, -1)[0]] >>> print(" ".join(words[start : end + 1])) M. Hamann P. Harper, P. Martinez ```rrrirhr_)rbnum_or_size_splitsrf)inputrfNstart_position end_positionr)r start_logits end_logitsrr') rrrHsplitsqueezerr rr')r7rRrTrrVrUrrWrrr rrrXrrfrrrrrrs r:ryz#TFLayoutLMForQuestionAnswering.callsDL--))%'/!5#   "!*88#%8&QUW#X#X#X jz 2>>> Zjr:::   &=+D&8F%2F> "''v|Z>X'YYD F"J/'!""+=F)-)9TGf$$v E-%!!/)     r;c|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|dd`tj|jj5|jdd|jj gddddS#1swxYwYdSdS)NTrr) rMrNrHrIrr&rOrr!r,rPs r:rOz$TFLayoutLMForQuestionAnswering.buildrrrzr)rRrxrTryrryrVryrUryrryrWryrrrrr rrryrryrXrrZrr{)r|r}r~rr+rrrrrr rryrOrrs@r:rrs***&      **+D+K+KLi+j+jkk+IXghhh.2.28<8<6:377;)-,0#'9=7; %h h h h ihlk]h T M M M M M M M Mr;r)rrrrrlrr)Dr __future__rrrnumpynp tensorflowrHactivations_tfrmodeling_tf_outputsrrrr r r modeling_tf_utilsr r rrrrrrrrtf_utilsrrrutilsrrrrconfiguration_layoutlmr get_loggerr|rrr0Layerr rrrrrrrr3r<rBrcrlrLAYOUTLM_START_DOCSTRINGrrrrrr__all__r)r;r:rs"""""" //////                        SRRRRRRRRRtttttttttttt222222  H % %"C C C C C 5<-C C C NAHAHAHAHAHel0AHAHAHJLLLLL5<-LLL>0.0.0.0.0.%,,0.0.0.hHHHHHU\/HHH<LLLLLu|)LLL>d0d0d0d0d0el(d0d0d0PK&K&K&K&K& *K&K&K&^HHHHHu|)HHH<"L"L"L"L"L (:"L"L"LL-----!3---b----- *---(L(L(L(L(L(%,,L(L(L(^      1    (T5phS*S*S*S*S*/S*S* S*lRTlmm@%@%@%@%@%57S@%@%nm@%F xMxMxMxMxM*CEaxMxMxMv |M|M|M|M|M'@B[|M|M|M~  LMLMLMLMLM%>@WLMLMLM^   r;