sgGJdZddlmZddlZddlmZddlmZmZm Z ddl Z ddl Z ddlmZddlmZmZmZmZmZmZmZmZdd lmZmZmZmZmZmZm Z m!Z!m"Z"m#Z#m$Z$m%Z%dd l&m'Z'm(Z(m)Z)dd l*m+Z+m,Z,m-Z-m.Z.m/Z/m0Z0d d l1m2Z2e/jfe4Z5dZ6dZ7dZ8dZ9dZ:dZ;dZdZ?dZ@dZAdZBGddZCGdde#jjZFGdd e#jjZHGd!d"e#jjZIeHeId#ZJGd$d%e#jjZKGd&d'e#jjZLGd(d)e#jjZMGd*d+e#jjZNGd,d-e#jjZOGd.d/e#jjZPGd0d1e#jjZQGd2d3e#jjZRGd4d5e#jjZSGd6d7e#jjZTGd8d9e#jjZUGd:d;e#jjZVGd<d=e#jjZWGd>d?e#jjZXGd@dAe#jjZYGdBdCe#jjZZe$GdDdEe#jjZ[GdFdGeZ\eGdHdIe+Z]dJZ^dKZ_e-dLe^GdMdNe\Z`e-dOe^GdPdQe\eCZae-dRe^GdSdTe\eZbGdUdVe#jjZce-dWe^GdXdYe\eZde-dZe^Gd[d\e\e Zee-d]e^Gd^d_e\eZfe-d`e^Gdadbe\eZge-dce^Gdddee\e!Zhy)fzTF 2.0 MobileBERT model.) annotationsN) dataclass)OptionalTupleUnion)get_tf_activation)TFBaseModelOutputTFBaseModelOutputWithPoolingTFMaskedLMOutputTFMultipleChoiceModelOutputTFNextSentencePredictorOutputTFQuestionAnsweringModelOutputTFSequenceClassifierOutputTFTokenClassifierOutput) TFMaskedLanguageModelingLossTFModelInputTypeTFMultipleChoiceLossTFNextSentencePredictionLossTFPreTrainedModelTFQuestionAnsweringLossTFSequenceClassificationLossTFTokenClassificationLossget_initializerkeraskeras_serializable unpack_inputs)check_embeddings_within_bounds shape_liststable_softmax) ModelOutputadd_code_sample_docstringsadd_start_docstrings%add_start_docstrings_to_model_forwardloggingreplace_return_docstrings)MobileBertConfigzgoogle/mobilebert-uncasedr(z"vumichien/mobilebert-finetuned-nerzK['I-ORG', 'I-ORG', 'O', 'O', 'O', 'O', 'O', 'I-LOC', 'O', 'I-LOC', 'I-LOC']gQ?z%vumichien/mobilebert-uncased-squad-v2z'a nice puppet'gףp= @ zvumichien/emo-mobilebertz'others'z4.72ceZdZdZddZy)TFMobileBertPreTrainingLossz Loss function suitable for BERT-like pretraining, that is, the task of pretraining a language model by combining NSP + MLM. .. note:: Any label of -100 will be ignored (along with the corresponding logits) in the loss computation. ctjjdtjjj}|t j j|d|d}t j|ddk7|j}||z}t j|t j|z }|t j j|d|d }t j|ddk7|j} || z} t j| t j| z } t j|| zd S) NT) from_logits reductionlabelsr)y_truey_predidtypenext_sentence_labelr')r') rlossesSparseCategoricalCrossentropy ReductionNONEtfnnrelucastr4 reduce_sumreshape) selfr0logitsloss_fnunmasked_lm_losses lm_loss_maskmasked_lm_lossesreduced_masked_lm_lossunmasked_ns_loss ns_loss_maskmasked_ns_lossreduced_masked_ns_losss h/var/www/html/venv/lib/python3.12/site-packages/transformers/models/mobilebert/modeling_tf_mobilebert.pyhf_compute_lossz+TFMobileBertPreTrainingLoss.hf_compute_loss_s-,,<<Y^YeYeYoYoYtYt<u%BEEJJvh7G,HQWXYQZ[wwvh/47?Q?W?WX - <!#/?!@2==Q]C^!^#"%%**V|A\!\zz03II4PPN)r0 tf.TensorrArNreturnrN)__name__ __module__ __qualname____doc__rLrMrKr,r,Xs  QrMr,c,eZdZfdZdZddZxZS)TFMobileBertIntermediatec ,t|di|tjj |j d|_t|jtr"t|j|_ ||_ y|j|_ ||_ y)NdensenamerT) super__init__rlayersDenseintermediate_sizerX isinstance hidden_actstrr intermediate_act_fnconfigr@rdkwargs __class__s rKr\z!TFMobileBertIntermediate.__init__usw "6"\\''(@(@w'O f'' -'89J9J'KD $ (.'8'8D $ rMcJ|j|}|j|}|SN)rXrcr@ hidden_statess rKcallzTFMobileBertIntermediate.calls& =1 00? rMc(|jryd|_t|ddetj|jj 5|jj dd|jjgdddyy#1swYyxYwNTrX) builtgetattrr: name_scoperXrZbuildrdtrue_hidden_sizer@ input_shapes rKrrzTFMobileBertIntermediate.builds} ::  4$ ' 3tzz/ M   $dkk.J.J!KL M M 4 M M 3BBrirPrQrRr\rlrr __classcell__rgs@rKrVrVts  MrMrVc*eZdZfdZdfd ZxZS) TFLayerNormc2||_t||i|yri) feat_sizer[r\)r@r}argsrfrgs rKr\zTFLayerNorm.__init__s" $)&)rMc>t|dd|jgyri)r[rrr}r@rurgs rKrrzTFLayerNorm.builds  tT4>>23rMri)rPrQrRr\rrrxrys@rKr{r{s*44rMr{c2eZdZdfd ZfdZddZxZS)TFNoNormc 2t|di|||_y)NrT)r[r\r})r@r}epsilonrfrgs rKr\zTFNoNorm.__init__s "6""rMc|jd|jgd|_|jd|jgd|_t||y)Nbiaszeros)shape initializerweightones) add_weightr}rrr[rrrs rKrrzTFNoNorm.buildsKOOF4>>2BPWOX ooht~~6FTZo[   k"rMc:||jz|jzSri)rr)r@inputss rKrlz TFNoNorm.calls #dii//rMri)rrN)rPrQrRr\rrrlrxrys@rKrrs## 0rMr) layer_normno_normc2eZdZdZfdZddZddZxZS)TFMobileBertEmbeddingszGConstruct the embeddings from word, position and token_type embeddings.c Vt|di||j|_|j|_||_|j |_|j |_|j|_tjj|j d|_ t|j|j |jd|_tjj!|j"|_|j|jr dz|_ydz|_y) Nembedding_transformationrY LayerNormrrZ)raterr'rT)r[r\ trigram_inputembedding_sizerd hidden_sizemax_position_embeddingsinitializer_rangerr]r^rNORM2FNnormalization_typelayer_norm_epsrDropouthidden_dropout_probdropoutembedded_input_sizeres rKr\zTFMobileBertEmbeddings.__init__s "6"#11$33 !--'-'E'E$!'!9!9(- (:(:6;M;MTn(:(o%!!:!:;   (=(=K ||++1K1K+L #'#6#6t?Q?Q!#Y WX#Y rMctjd5|jd|jj|j gt |j|_dddtjd5|jd|jj|jgt |j|_ dddtjd5|jd|j|jgt |j|_ ddd|jryd|_t|d dZtj|j j"5|j j%dd|j&gdddt|d dNtj|j(j"5|j(j%ddddyy#1swYxYw#1swYOxYw#1swYxYw#1swYxYw#1swYyxYw) Nword_embeddingsr)r)rZrrtoken_type_embeddings embeddingsposition_embeddingsTrr)r:rqrrd vocab_sizerrrrtype_vocab_sizerrrrrorprrZrrrrrts rKrrzTFMobileBertEmbeddings.builds ]], - //{{--t/B/BC+d>T>TU*DK ]]2 3 )-!{{22D4D4DE+d>T>TU*9*D & ]]0 1 '+!33T5E5EF+d>T>TU(7(D $  ::  43T : Ft<<AAB \--33T4AYAY4Z[ \ 4d + 7t~~223 +$$T* + + 87       \ \ + +s>AHAH,AH))H5)IHH&)H25H>I c R||J|At||jjtj|j |}t |dd}|tj|d}|jrTtjtj|ddddfd|tj|ddddfdgd }|js|j|jk7r|j|}|/tjtjd|d d }tj|j |}tj|j"|}||z|z} |j%| } |j'| | } | S)z Applies embedding based on inputs tensor. Returns: final_embeddings (`tf.Tensor`): output embedding tensor. N)paramsindicesr)dimsvaluer')rr)rr'r)r)r'rraxis)startlimit)r)rtraining)rrdrr:gatherrrfillrconcatpadrrr expand_dimsrangerrrr) r@ input_ids position_idstoken_type_ids inputs_embedsrruposition_embedstoken_type_embedsfinal_embeddingss rKrlzTFMobileBertEmbeddings.calls%-*?@@  *9dkk6L6L MIIT[[)LM /4  !WW+Q?N   IIFF=AB/1IJ!FF=CRC02JK  M   !4!48H8H!H 99-HM  >>"((+b/*RYZ))4+C+C\ZIIT-G-GQ_`(?:=NN>>1A>B<$>>!CCC#&v'>'>A[A['[#\ !558P8PP\\''   ?6C[C[3\cj(  <<%%   ?6C[C[3\ch& \\''   ?6C[C[3\cj(  ||++F,O,OP  rMctj||d|j|jf}tj|gdS)Nrrrr'rperm)r:r?rr transpose)r@x batch_sizes rKtranspose_for_scoresz.TFMobileBertSelfAttention.transpose_for_scores-s8 JJq:r4+C+CTE]E]^ _||AL11rMct|d}|j|} |j|} |j|} |j | |} |j | |} |j | |}t j | | d}t jt| d|j}|t jj|z }|&t j||j}||z}t|d}|j||}|||z}t j ||}t j|gd }t j||d|jf}|r||f}|S|f}|S) NrT) transpose_brr3rrrr)rrrrrr:matmulr=r4mathsqrtr rrr?r)r@ query_tensor key_tensor value_tensorattention_mask head_maskrrrmixed_query_layermixed_key_layermixed_value_layer query_layer key_layer value_layerattention_scoresdkattention_probs context_layeroutputss rKrlzTFMobileBertSelfAttention.call2s /2  JJ|4((:. JJ|4//0A:N --ozJ //0A:N 99  WWZ *2.6F6L6L M+bggll2.>>  %WW^;K;Q;QRN/.@ ))9C,,,J  - 9O /;?  ]F  JD,>,>? 7H=/2O\M]rMcv|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|dddtj|jj 5|jj dd|jjgdddt|ddtj|jj 5|jj dd|jjr|jjn|jjgdddyy#1swYxYw#1swYxYw#1swYyxYw)NTrrr) rorpr:rqrrZrrrdrsrruse_bottleneck_attentionrrts rKrrzTFMobileBertSelfAttention.build_sW ::  4$ ' 3tzz/ M   $dkk.J.J!KL M 4 % 1txx}}- KdDKK,H,HIJ K 4$ ' 3tzz/    ;;?? 44![[44    4  M M K K  s% 3F<3F#-AF/F #F,/F8Fri)rPrQrRr\rrlrrrxrys@rKrrs62 ns+ZrMrc.eZdZfdZddZddZxZS)TFMobileBertSelfOutputc t|di||j|_tjj |j t|jd|_ t|j|j |jd|_ |js.tjj|j|_||_y)NrXrrrrT)r[r\use_bottleneckrr]r^rsrrrXrrrrrrrrdres rKr\zTFMobileBertSelfOutput.__init__ws "6"$33\\''  # #H`H`8aho(  !!:!:;  # #V-B-B "" <<//0J0JKDL rMc|j|}|js|j||}|j||z}|SNr)rXrrr)r@rkresidual_tensorrs rKrlzTFMobileBertSelfOutput.callsD =1 "" LLLJM}'FG rMc|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYwNTrXr rorpr:rqrXrZrrrdrsrrts rKrrzTFMobileBertSelfOutput.build ::  4$ ' 3tzz/ M   $dkk.J.J!KL M 4d + 7t~~223 +$$T* + + 8 M M + + 3C"<C."C+.C7rrirwrys@rKrrvs  +rMrc6eZdZfdZdZ ddZddZxZS)TFMobileBertAttentionc lt|di|t|d|_t |d|_y)Nr@rYoutputrT)r[r\rr@rmobilebert_outputres rKr\zTFMobileBertAttention.__init__s0 "6"-f6B !7X!NrMctriNotImplementedError)r@headss rK prune_headsz!TFMobileBertAttention.prune_headss!!rMc v|j|||||||} |j| d||} | f| ddz} | S)Nrrr')r@r) r@rrr layer_inputrrrr self_outputsattention_outputrs rKrlzTFMobileBertAttention.calls`yy *lNIO`ks!  11,q/;Ya1b#% QR(88rMc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYw)NTr@r)rorpr:rqr@rZrrrrts rKrrzTFMobileBertAttention.builds ::  4 & 2tyy~~. & % & 4,d 3 ?t55::; 3&&,,T2 3 3 @ & & 3 3 C %C CC rri)rPrQrRr\r rlrrrxrys@rKrrsO "& 3rMrc.eZdZfdZddZddZxZS)TFOutputBottleneckc Vt|di|tjj |j d|_t|j|j |jd|_ tjj|j|_ ||_yNrXrYrrrT)r[r\rr]r^rrXrrrrrrrrdres rKr\zTFOutputBottleneck.__init__s "6"\\''(:(:'I  !:!:;   (=(=K ||++F,F,FG  rMcv|j|}|j||}|j||z}|Sr)rXrr)r@rkrr layer_outputss rKrlzTFOutputBottleneck.calls; =1  ]X F }'FG rMc|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYwrrrts rKrrzTFOutputBottleneck.buildrrrrirwrys@rKrrs +rMrc.eZdZfdZddZddZxZS)TFMobileBertOutputc t|di||j|_tjj |j t|jd|_ t|j|j |jd|_ |js6tjj|j|_||_yt#|d|_||_y)NrXrrr bottleneckrYrT)r[r\rrr]r^rsrrrXrrrrrrrrrrdres rKr\zTFMobileBertOutput.__init__s "6"$33\\''  # #H`H`8aho(  !!:!:;  # #V-B-B "" <<//0J0JKDL 1lKDO rMc|j|}|js)|j||}|j||z}|S|j||z}|j ||}|Sr)rXrrrr)r@rkresidual_tensor_1residual_tensor_2rs rKrlzTFMobileBertOutput.callst =1 "" LLLJM NN=;L+LMM!NN=;L+LMM OOM;LMMrMc|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYxYw#1swYqxYw#1swYyxYw)NTrXrr) rorpr:rqrXrZrrrdr_rrrts rKrrzTFMobileBertOutput.builds ::  4$ ' 3tzz/ N   $dkk.K.K!LM N 4d + 7t~~223 +$$T* + 4t , 8t334 ,%%d+ , , 9  N N + + , ,s$ 3D<<EE<EEErrirwrys@rKrrs  ,rMrc,eZdZfdZdZddZxZS)TFBottleneckLayerc t|di|tjj |j d|_t|j|j |jd|_ ||_ yr) r[r\rr]r^intra_bottleneck_sizerXrrrrrdres rKr\zTFBottleneckLayer.__init__sg "6"\\''(D(D7'S  !:!:;  ( (&2G2Gk  rMcJ|j|}|j|}|SrirXr)r@rrks rKrlzTFBottleneckLayer.call s$ 6* }5 rMc|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYwr rorpr:rqrXrZrrrdrrrts rKrrzTFBottleneckLayer.build ::  4$ ' 3tzz/ H   $dkk.E.E!FG H 4d + 7t~~223 +$$T* + + 8 H H + +rrirwrys@rKr r  +rMr c,eZdZfdZdZddZxZS) TFBottleneckc t|di||j|_|j|_t |d|_|jrt |d|_yy)NinputrY attentionrT)r[r\key_query_shared_bottleneckrr bottleneck_inputr-res rKr\zTFBottleneck.__init__sZ "6"+1+M+M((.(G(G% 1&w G  + +.vKHDN ,rMc|j|}|jr|fdzS|jr|j|}||||fS||||fS)N)r/rr.r-)r@rkbottlenecked_hidden_statesshared_attention_inputs rKrlzTFBottleneck.call'se"&*%:%:=%I"  ( (.014 4  - -%)^^M%B "*,BMSmn n!=-A[\ \rMc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYw)NTr/r-)rorpr:rqr/rZrrr-rts rKrrzTFBottleneck.buildAs ::  4+T 2 >t4499: 2%%++D1 2 4d + 7t~~223 +$$T* + + 8 2 2 + +rrirwrys@rKr*r*sI]4 +rMr*c,eZdZfdZdZddZxZS) TFFFNOutputc t|di|tjj |j d|_t|j|j |jd|_ ||_ yr) r[r\rr]r^rsrXrrrrrdres rKr\zTFFFNOutput.__init__Nsg "6"\\''(?(?g'N  !:!:;  # #V-B-B  rMcP|j|}|j||z}|Srir$)r@rkrs rKrlzTFFFNOutput.callVs) =1 }'FG rMc|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYwr) rorpr:rqrXrZrrrdr_rrts rKrrzTFFFNOutput.build[s ::  4$ ' 3tzz/ N   $dkk.K.K!LM N 4d + 7t~~223 +$$T* + + 8 N N + +rrirwrys@rKr6r6Mr(rMr6c,eZdZfdZdZddZxZS) TFFFNLayerc lt|di|t|d|_t |d|_y)N intermediaterYrrT)r[r\rVr=r6rres rKr\zTFFFNLayer.__init__hs1 "6"4V.Q!,V(!CrMcL|j|}|j||}|Sri)r=r)r@rkintermediate_outputrs rKrlzTFFFNLayer.callms,"// >../BMR rMc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYw)NTr=r)rorpr:rqr=rZrrrrts rKrrzTFFFNLayer.buildrs ::  4 . :t00556 .!!''- . 4,d 3 ?t55::; 3&&,,T2 3 3 @ . . 3 3rrirwrys@rKr;r;gsD  3rMr;c.eZdZfdZddZddZxZS)TFMobileBertLayerc t|di||j|_|j|_t |d|_t |d|_t|d|_ |jrt|d|_ |jdkDr:t|jdz Dcgc]}t|d|c}|_yycc}w) Nr-rYr=rrr'zffn.rT)r[r\rnum_feedforward_networksrr-rVr=rrr*rrr;ffnr@rdrfirgs rKr\zTFMobileBertLayer.__init__s "6"$33(.(G(G%.vKH4V.Q!3F!J   *6 EDO  * *Q .EJ6KjKjmnKnEop 6$qc ;pDH /ps0Cc |jr|j|\}}}} n |gdz\}}}} |j|||| ||||} | d} | f} |jdk7r+t |j D]\} }|| } | | fz } |j | }|j|| ||}|f| ddztjd|||| | |fz| z}|S)Nr1rrr') rrr-rD enumeraterEr=rr:constant)r@rkrrrrrrrr attention_outputsr srG ffn_moduler? layer_outputrs rKrlzTFMobileBertLayer.callsD   BF//R_B` ?L*lKCP/TUBU ?L*lK NN       +  -Q/    ( (A -!*488!4 ) :#-.>#? &(( )#//0@A--.ACSUbmu-v O# $ A #   rMc(|jryd|_t|ddMtj|jj 5|jj ddddt|ddMtj|jj 5|jj ddddt|ddMtj|jj 5|jj ddddt|ddMtj|jj 5|jj ddddt|ddK|jD];}tj|j 5|j dddd=yy#1swYqxYw#1swY$xYw#1swYxYw#1swYxYw#1swY{xYw)NTr-r=rrrE) rorpr:rqr-rZrrr=rrrEr@rulayers rKrrzTFMobileBertLayer.builds ::  4d + 7t~~223 +$$T* + 4 . :t00556 .!!''- . 4,d 3 ?t55::; 3&&,,T2 3 4t , 8t334 ,%%d+ , 4 % 1 &]]5::.&KK%&& & 2 + + . . 3 3 , ,&&s< G%G#?G0G<8HG #G-0G9<HH rrirwrys@rKrBrB~s q+Z&rMrBc0eZdZfdZ ddZddZxZS)TFMobileBertEncoderc t|di||j|_|j|_t |j Dcgc]}t |d|c}|_ycc}w)Nzlayer_._rYrT)r[r\routput_hidden_statesrnum_hidden_layersrBrQrFs rKr\zTFMobileBertEncoder.__init__s^ "6"!'!9!9$*$?$?!NSTZTlTlNmn'xs^Dn ns A*c|rdnd}|rdnd} t|jD].\} } |r||fz}| |||| ||} | d}|s&| | dfz} 0|r||fz}|std||| fDSt||| S)NrTrrr'c3&K|] }|| ywrirT).0vs rK z+TFMobileBertEncoder.call..shqZ[Zghs)last_hidden_staterk attentions)rIrQtupler ) r@rkrrrrU return_dictrall_hidden_statesall_attentionsrG layer_modulers rKrlzTFMobileBertEncoder.calls#7BD0d(4 FOA|#$58H$H!(~y|=NYaM*!,M !/=3C2E!E F 1]4D D h]4E~$Vhh h +;LYg  rMc|jryd|_t|ddK|jD];}tj|j 5|j dddd=yy#1swYIxYw)NTrQ)rorprQr:rqrZrrrPs rKrrzTFMobileBertEncoder.buildsp ::  4$ ' 3 &]]5::.&KK%&& & 4&&s A..A7 rrirwrys@rKrSrSso! F&rMrSc,eZdZfdZdZddZxZS)TFMobileBertPoolerc t|di||j|_|jrEtj j |jt|jdd|_ ||_ y)NtanhrX)r activationrZrT) r[r\classifier_activation do_activaterr]r^rrrrXrdres rKr\zTFMobileBertPooler.__init__sk "6"!77   ++""#263K3K#L! ,DJ  rMcV|dddf}|js|S|j|}|SNr)rjrX)r@rkfirst_token_tensor pooled_outputs rKrlzTFMobileBertPooler.calls7+1a40% % JJ'9:M rMc(|jryd|_t|ddetj|jj 5|jj dd|jjgdddyy#1swYyxYwrn) rorpr:rqrXrZrrrdrrts rKrrzTFMobileBertPooler.builds} ::  4$ ' 3tzz/ H   $dkk.E.E!FG H H 4 H Hrvrirwrys@rKreres !HrMrec,eZdZfdZdZddZxZS)#TFMobileBertPredictionHeadTransformc t|di|tjj |j t |jd|_t|jtrt|j|_ n|j|_ td|j |jd|_||_y)NrXrrrrrT)r[r\rr]r^rrrrXr`rarbr transform_act_fnrrrrdres rKr\z,TFMobileBertPredictionHeadTransform.__init__$s "6"\\''   ?6C[C[3\cj(   f'' -$5f6G6G$HD !$*$5$5D ! .v/A/A6K`K`grs rMcl|j|}|j|}|j|}|Sri)rXrsrrjs rKrlz(TFMobileBertPredictionHeadTransform.call0s4 =1 --m< }5 rMc|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYwrr&rts rKrrz)TFMobileBertPredictionHeadTransform.build6r'rrirwrys@rKrqrq#s  +rMrqcDeZdZfdZddZdZdZdZdZdZ xZ S) TFMobileBertLMPredictionHeadc Vt|di|t|d|_||_y)N transformrYrT)r[r\rqryrdres rKr\z%TFMobileBertLMPredictionHead.__init__Cs( "6"rMrwc,eZdZfdZdZddZxZS)TFMobileBertMLMHeadc Ht|di|t|d|_y)N predictionsrYrT)r[r\rwrres rKr\zTFMobileBertMLMHead.__init__ts" "6"7]SrMc(|j|}|Srir)r@sequence_outputprediction_scoress rKrlzTFMobileBertMLMHead.callxs ,,_=  rMc|jryd|_t|ddNtj|jj 5|jj ddddyy#1swYyxYw)NTr)rorpr:rqrrZrrrts rKrrzTFMobileBertMLMHead.build|sm ::  4 - 9t//445 -  &&t, - - : - - A11A:rirwrys@rKrrssT!-rMrcdeZdZeZdfd ZdZdZdZe ddZ d dZ xZ S) TFMobileBertMainLayerc :t|di|||_|j|_|j|_|j |_|j |_t|d|_ t|d|_ |rt|d|_ yd|_ y)NrrYencoderpoolerrT)r[r\rdrVrrUuse_return_dictr_rrrSrrer)r@rdadd_pooling_layerrfrgs rKr\zTFMobileBertMainLayer.__init__s "6" !'!9!9!'!9!9$*$?$?!!110lK*6 B CT(h? Z^ rMc|jSri)rr~s rKget_input_embeddingsz*TFMobileBertMainLayer.get_input_embeddingss rMc`||j_t|d|j_yrl)rrrrrs rKset_input_embeddingsz*TFMobileBertMainLayer.set_input_embeddingss$!&%/%6q%9"rMct)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)r@heads_to_prunes rK _prune_headsz"TFMobileBertMainLayer._prune_headss "!rMc `| | td| t|} n|t|dd} n td|tj| d}|tj| d}|j ||||| } tj || ddd| df} tj | | j} tjd| j}tjd | j}tjtj|| |} |tdg|jz}|j| | |||| | }|d}|j|j|nd}| s ||f|ddzSt|||j |j" S) NzDYou cannot specify both input_ids and inputs_embeds at the same timerz5You have to specify either input_ids or inputs_embedsr'rrr3g?g)r\ pooler_outputrkr])rrr:rrr?r=r4rJmultiplysubtractrrVrrr rkr])r@rrrrrrrrUr_rruembedding_outputextended_attention_maskone_cstten_thousand_cstencoder_outputsrrns rKrlzTFMobileBertMainLayer.calls  ]%>cd d  "$Y/K  &$]3CR8KTU U  !WW[!4N  !WW[!4N??9lNTalt?u#%**^k!naQRT_`aTb=c"d#%''*AIYI_I_"`++c)9)?)?@;;x7G7M7MN"$++bkk'CZ.[]m"n  % %!7!77I,,  #   ' *!,8< 8O O4UY  #$ $ ,-')77&11   rMc|jryd|_t|ddMtj|jj 5|jj ddddt|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYxYw#1swYqxYw#1swYyxYw)NTrrr) rorpr:rqrrZrrrrrts rKrrzTFMobileBertMainLayer.builds ::  4t , 8t334 ,%%d+ , 4D ) 5t||001 ) ""4( ) 44 ( 4t{{//0 ( !!$' ( ( 5  , , ) ) ( ($ D%%D1?D=%D.1D:=E)T NNNNNNNNNFri) rPrQrRr( config_classr\rrrrrlrrrxrys@rKrrsY#L _:"!Q Q f (rMrceZdZdZeZdZy)TFMobileBertPreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. mobilebertN)rPrQrRrSr(rbase_model_prefixrTrMrKrrs $L$rMrcXeZdZUdZdZded<dZded<dZded<dZded <dZ ded <y) TFMobileBertForPreTrainingOutputaE Output type of [`TFMobileBertForPreTraining`]. Args: prediction_logits (`tf.Tensor` of shape `(batch_size, sequence_length, config.vocab_size)`): Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). seq_relationship_logits (`tf.Tensor` of shape `(batch_size, 2)`): Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation before SoftMax). hidden_states (`tuple(tf.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `tf.Tensor` (one for the output of the embeddings + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the initial embedding outputs. attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Nztf.Tensor | NonelossrNprediction_logitsseq_relationship_logitszTuple[tf.Tensor] | Nonerkr]) rPrQrRrSr__annotations__rrrkr]rTrMrKrrs@,"D !#'y')-Y--1M*1*.J'.rMra 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! Parameters: config ([`MobileBertConfig`]): 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 [`~PreTrainedModel.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) 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. This argument can be used only in eager mode, in graph mode the value in the config will be used instead. 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. This argument can be used only in eager mode, in graph mode the value in the config will be used instead. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This argument can be used in eager mode, in graph mode the value will always be set to True. 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). zdThe bare MobileBert Model transformer outputting raw hidden-states without any specific head on top.ceZdZfdZeeejdee e e  d ddZ ddZ xZS) TFMobileBertModelcPt||g|i|t|d|_y)NrrY)r[r\rrr@rdrrfrgs rKr\zTFMobileBertModel.__init__s( 3&3F3/\JrMbatch_size, sequence_length checkpoint output_typerc <|j||||||||| |  } | S)N) rrrrrrrrUr_r)r) r@rrrrrrrrUr_rrs rKrlzTFMobileBertModel.calls<(//))%'/!5#"  rMc|jryd|_t|ddNtj|jj 5|jj ddddyy#1swYyxYw)NTr)rorpr:rqrrZrrrts rKrrzTFMobileBertModel.buildsi ::  4t , 8t334 ,%%d+ , , 9 , ,rr)rTFModelInputType | Nonernp.ndarray | tf.Tensor | NonerrrrrrrrrOptional[bool]rUrr_rrrrOz*Union[Tuple, TFBaseModelOutputWithPooling]ri)rPrQrRr\rr$MOBILEBERT_INPUTS_DOCSTRINGformatr"_CHECKPOINT_FOR_DOCr _CONFIG_FOR_DOCrlrrrxrys@rKrrs K*+F+M+MNk+lm&0$.28<8<6:377;,0/3&*#(*66  4  1 5*-$! 4 n8,rMrz MobileBert Model with two heads on top as done during the pretraining: a `masked language modeling` head and a `next sentence prediction (classification)` head. ceZdZfdZdZdZeeejde e e  d d dZ d dZdZxZS) TFMobileBertForPreTrainingct||g|i|t|d|_t |d|_t |d|_y)NrrYpredictions___clsseq_relationship___cls)r[r\rrrrTFMobileBertOnlyNSPHeadseq_relationshiprs rKr\z#TFMobileBertForPreTraining.__init__sH 3&3F3/\J.v>> import tensorflow as tf >>> from transformers import AutoTokenizer, TFMobileBertForPreTraining >>> tokenizer = AutoTokenizer.from_pretrained("google/mobilebert-uncased") >>> model = TFMobileBertForPreTraining.from_pretrained("google/mobilebert-uncased") >>> input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1 >>> outputs = model(input_ids) >>> prediction_scores, seq_relationship_scores = outputs[:2] ``` rrrrrrrUr_rNrr0r5r0rA)rrrrkr])rrrrLrrkr])r@rrrrrrrrUr_r0r5rrrrnrseq_relationship_score total_lossd_labelsrs rKrlzTFMobileBertForPreTraining.callsB// ))%'/!5#"  *1!& ,,_=!%!6!6}!E  "5"A &)H.AH* +--XGXZpFq-rJ')?@712;NF/9/EZMF* Q6 Q//$:!//))   rMc|jryd|_t|ddMtj|jj 5|jj ddddt|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYxYw#1swYqxYw#1swYyxYw)NTrrr) rorpr:rqrrZrrrrrts rKrrz TFMobileBertForPreTraining.build s  ::  4t , 8t334 ,%%d+ , 4 - 9t//445 -  &&t, - 4+T 2 >t4499: 2%%++D1 2 2 ?  , , - - 2 2rc|dk(r|dfS|fSNzcls.predictions.decoder.weightz,mobilebert.embeddings.word_embeddings.weightrTr@ tf_weights rKtf_to_pt_weight_renamez1TFMobileBertForPreTraining.tf_to_pt_weight_rename. 8 8LL L< rM NNNNNNNNNNNF)rrrrrrrrrrrrrrrUrr_rr0rr5rrrrOz.Union[Tuple, TFMobileBertForPreTrainingOutput]ri)rPrQrRr\rrrr$rrr&rrrlrrrrxrys@rKrrs_ ,a*+F+M+MNk+lm+KZij.28<8<6:377;,0/3&*04=A#(? *? 6? 6 ? 4 ? 1 ? 5? *? -? $? .? ;? !?  8? kn? B 2 rMrz8MobileBert Model with a `language modeling` head on top.c eZdZgdZfdZdZdZeee jde e e edd d d d Zdd Zd ZxZS)TFMobileBertForMaskedLM)rrcls.seq_relationshipcvt||g|i|t|dd|_t |d|_y)NFrrrZrrY)r[r\rrrrrs rKr\z TFMobileBertForMaskedLM.__init__>s; 3&3F3/%Vbc.vt4499: S%%++T49P9P,QR S S ? S Srvrirwrys@rKrrs &SrMrzPMobileBert Model with a `next sentence prediction (classification)` head on top.ceZdZddgZfdZeeejde e e  d d dZ d dZ xZS) %TFMobileBertForNextSentencePredictionrcls.predictionsctt||g|i|t|d|_t |d|_y)NrrYr)r[r\rrrclsrs rKr\z.TFMobileBertForNextSentencePrediction.__init__s7 3&3F3/\J*68PQrMrrc |j||||||||| |  } | d} |j| }| dn|j| |}| s|f| ddz}||f|zS|St||| j| j S)a Return: Examples: ```python >>> import tensorflow as tf >>> from transformers import AutoTokenizer, TFMobileBertForNextSentencePrediction >>> tokenizer = AutoTokenizer.from_pretrained("google/mobilebert-uncased") >>> model = TFMobileBertForNextSentencePrediction.from_pretrained("google/mobilebert-uncased") >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced." >>> next_sentence = "The sky is blue due to the shorter wavelength of blue light." >>> encoding = tokenizer(prompt, next_sentence, return_tensors="tf") >>> logits = model(encoding["input_ids"], token_type_ids=encoding["token_type_ids"])[0] ```rr'Nrrr)rrrLrrkr])r@rrrrrrrrUr_r5rrrnseq_relationship_scoresnext_sentence_lossrs rKrlz*TFMobileBertForNextSentencePrediction.callsF// ))%'/!5#"    "&((="9#* %%-@I`%a  -/'!"+=F7I7U')F2 a[a a,#*!//))   rMc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYw)NTrr)rorpr:rqrrZrrrrts rKrrz+TFMobileBertForNextSentencePrediction.builds ::  4t , 8t334 ,%%d+ , 4 % 1txx}}- %t$ % % 2 , , % %rr)rrrrrrrrrrrrrrrUrr_rr5rrrrOz+Union[Tuple, TFNextSentencePredictorOutput]ri)rPrQrRrr\rr$rrr&rrrlrrrxrys@rKrrs +?@R)S&R *+F+M+MNk+lm+HWfg.28<8<6:377;,0/3&*=A#(> *> 6> 6 > 4 > 1 > 5> *> -> $> ;> !>  5> hn> @ %rMrz MobileBert Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled output) e.g. for GLUE tasks. c eZdZgdZdgZfdZeeejde e e e ee d d dZd dZxZS) %TFMobileBertForSequenceClassificationrrrrrct||g|i||j|_t|d|_|j |j n |j }tjj||_ tjj|jt|jd|_||_y)NrrY classifierrr[r\ num_labelsrrclassifier_dropoutrrr]rrr^rrr rdr@rdrrfr rgs rKr\z.TFMobileBertForSequenceClassification.__init__s 3&3F3 ++/\J)/)B)B)NF % %TZTnTn ||++,>? ,,,,   /&BZBZ2[bn-  rMrrc &|j||||||||| |  } | d} |j| | } |j| }| dn|j| |}| s|f| ddz}||f|zS|St ||| j | j S)a labels (`tf.Tensor` 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). rr'rNrr)rrr rLrrkr])r@rrrrrrrrUr_r0rrrnrArrs rKrlz*TFMobileBertForSequenceClassification.call%s:// ))%'/!5#"     ]X F /~t4+?+?+OY,F)-)9TGf$ Ev E)!//))   rMc|jryd|_t|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYwNTrr  rorpr:rqrrZrrr rdrrts rKrrz+TFMobileBertForSequenceClassification.build` ::  4t , 8t334 ,%%d+ , 4t , 8t334 M%%tT4;;3J3J&KL M M 9 , , M M C"%3C."C+.C7r)rrrrrrrrrrrrrrrUrr_rr0rrrrOz(Union[Tuple, TFSequenceClassifierOutput]ri)rPrQrRr_keys_to_ignore_on_load_missingr\rr$rrr"'_CHECKPOINT_FOR_SEQUENCE_CLASSIFICATIONrr_SEQ_CLASS_EXPECTED_OUTPUT_SEQ_CLASS_EXPECTED_LOSSrlrrrxrys@rKrrs*& (2l# *+F+M+MNk+lm:.$2. .28<8<6:377;,0/3&*04#(0 *0 60 6 0 4 0 1 0 50 *0 -0 $0 .0 !0  20 n0 d MrMrz MobileBert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of the hidden-states output to compute `span start logits` and `span end logits`). c eZdZgdZfdZeeejde e e e e eee d ddZd dZxZS) TFMobileBertForQuestionAnsweringrrrrrc t||g|i||j|_t|dd|_t j j|jt|jd|_ ||_ y)NFrr qa_outputsr) r[r\r rrrr]r^rrrrdrs rKr\z)TFMobileBertForQuestionAnswering.__init__}su 3&3F3 ++/%Vbc,,,,   /&BZBZ2[bn-  rMr)rrrqa_target_start_indexqa_target_end_indexrrc |j||||||||| |  } | d}|j|}tj|dd\}}tj|d}tj|d}d}| | | | d}|j |||f}| s||f| ddz}||f|zS|St |||| j| jS) a start_positions (`tf.Tensor` 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` 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. rrrrrN)start_position end_position)r start_logits end_logitsrkr]) rrr:splitsqueezerLrrkr])r@rrrrrrrrUr_start_positions end_positionsrrrrAr#r$rr0rs rKrlz%TFMobileBertForQuestionAnswering.callsH// ))%'/!5#"  "!*1#%88FAB#? jzz,R8 ZZ 4   &=+D(7WF''z0JKD"J/'!"+=F)-)9TGf$ Ev E-%!!//))   rMc|jryd|_t|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYw)NTrr) rorpr:rqrrZrrrrdrrts rKrrz&TFMobileBertForQuestionAnswering.buildrrr)rrrrrrrrrrrrrrrUrr_rr'rr(rrrrOz,Union[Tuple, TFQuestionAnsweringModelOutput]ri)rPrQrRrr\rr$rrr"_CHECKPOINT_FOR_QArr_QA_TARGET_START_INDEX_QA_TARGET_END_INDEX_QA_EXPECTED_OUTPUT_QA_EXPECTED_LOSSrlrrrxrys@rKrrls*&*+F+M+MNk+lm%2$40+'.28<8<6:377;,0/3&*9=7;#(; *; 6; 6 ; 4 ; 1 ; 5; *; -; $; 7; 5; !;  6; n; z MrMrz MobileBert Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a softmax) e.g. for RocStories/SWAG tasks. ceZdZgdZdgZfdZeeejde e e e  d d dZd dZxZS) TFMobileBertForMultipleChoicerrc.t||g|i|t|d|_tj j |j|_tj jdt|jd|_ ||_ y)NrrYr'r r)r[r\rrrr]rrrr^rrr rdrs rKr\z&TFMobileBertForMultipleChoice.__init__s{ 3&3F3/\J||++F,F,FG ,,,, /&2J2J"KR^-  rMz(batch_size, num_choices, sequence_lengthrc |t|d} t|d} nt|d} t|d} |tj|d| fnd}|tj|d| fnd}|tj|d| fnd}|tj|d| fnd}|%tj|d| t|dfnd}|j||||||||| |  }|d}|j || }|j |}tj|d| f}| dn|j | |}| s|f|ddz}||f|zS|St|||j|jS) a5 labels (`tf.Tensor` of shape `(batch_size,)`, *optional*): Labels for computing the multiple choice classification loss. Indices should be in `[0, ..., num_choices]` where `num_choices` is the size of the second dimension of the input tensors. (See `input_ids` above) Nr'rrr)r_rrr) rr:r?rrr rLr rkr])r@rrrrrrrrUr_r0r num_choices seq_lengthflat_input_idsflat_attention_maskflat_token_type_idsflat_position_idsflat_inputs_embedsrrnrAreshaped_logitsrrs rKrlz"TFMobileBertForMultipleChoice.calls8  $Y/2K#I.q1J$]3A6K#M215JDMDYIJ/?@_cN\Nhbjj"j9IJnrN\Nhbjj"j9IJnrJVJbBJJ|b*5EFhl( JJ}r:z-7PQR7S&T U  //        #"     ]X F /**Vb+->?~t4+?+?+X%''!"+5F)-)9TGf$ Ev E*"!//))   rMc|jryd|_t|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYwrrrts rKrrz#TFMobileBertForMultipleChoice.build@rrr)rrrrrrrrrrrrrrrUrr_rr0rrrrOz)Union[Tuple, TFMultipleChoiceModelOutput]ri)rPrQrRrrr\rr$rrr"rr rrlrrrxrys@rKr0r0s*& (2l#*#**+UV &/$.28<8<6:377;,0/3&*04#(? *? 6? 6 ? 4 ? 1 ? 5? *? -? $? .? !?  3? ? B MrMr0z MobileBert 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. c eZdZgdZdgZfdZeeejde e e e ee d d dZd dZxZS) "TFMobileBertForTokenClassificationrrct||g|i||j|_t|dd|_|j |j n |j }tjj||_ tjj|jt|jd|_||_y)NFrrr rr rs rKr\z+TFMobileBertForTokenClassification.__init__^s 3&3F3 ++/%Vbc)/)B)B)NF % %TZTnTn ||++,>? ,,,,   /&BZBZ2[bn-  rMrrc &|j||||||||| |  } | d} |j| | } |j| }| dn|j| |}| s|f| ddz}||f|zS|St ||| j | j S)z labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`. rrrNrr)rrr rLrrkr])r@rrrrrrrrUr_r0rrrrArrs rKrlz'TFMobileBertForTokenClassification.callls6// ))%'/!5#"  "!*,,,J1~t4+?+?+OY,F)-)9TGf$ Ev E&!//))   rMc|jryd|_t|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYwrrrts rKrrz(TFMobileBertForTokenClassification.buildrrr)rrrrrrrrrrrrrrrUrr_rr0rrrrOz%Union[Tuple, TFTokenClassifierOutput]ri)rPrQrRrrr\rr$rrr"$_CHECKPOINT_FOR_TOKEN_CLASSIFICATIONrr_TOKEN_CLASS_EXPECTED_OUTPUT_TOKEN_CLASS_EXPECTED_LOSSrlrrrxrys@rKr=r=Ls*&(2l# *+F+M+MNk+lm7+$40 .28<8<6:377;,0/3&*04#(. *. 6. 6 . 4 . 1 . 5. *. -. $. .. !.  /. n. ` MrMr=)irS __future__rr dataclassesrtypingrrrnumpynp tensorflowr:activations_tfr modeling_tf_outputsr r r r rrrrmodeling_tf_utilsrrrrrrrrrrrrtf_utilsrrr utilsr!r"r#r$r%r&configuration_mobilebertr( get_loggerrPloggerrrrArBrCr*r-r.r+r,rrrr,r]LayerrVLayerNormalizationr{rrrrrrrrr r*r6r;rBrSrerqrwrrrrMOBILEBERT_START_DOCSTRINGrrrrrrrrr0r=rTrMrKrUs "!))/       SR7   H %1$(L$l!='+E''!QQ8Mu||11M64%,,114 0u||!! 0% :e U\\//e Pb 2 2bJ+U\\//+B%3ELL..%3P++++8&,++&,R+ **+4,+5<<%%,+^+%,,$$+43##3.N& **N&b1&%,,,,1&hH++H@+%,,*<*<+>.5<<#5#5.b-%,,,,-$~(ELL..~(~(B%"3% /{/ /<(T5pj.,3., .,b d !<>Yd d NTVpq] 9;W] r] @Sell00S&ZV%,GIeV% V%r \M,GIe\M\M~ eM'BD[eMeMP gM$?AUgMgMT [M)DF_[M[MrM