`iGdZddlmZddlZddlmZddlZddlZ ddl m Z ddl m Z mZmZmZmZmZmZmZddlmZmZmZmZmZmZmZmZmZmZm Z m!Z!dd l"m#Z#m$Z$m%Z%dd l&m'Z'm(Z(m)Z)m*Z*m+Z+m,Z,d d l-m.Z.e+j/e0Z1d Z2dZ3dZ4dZ5dZ6dZ7dZ8dZ9dZ:dZ;dZGddZ?Gddej@jAZBGddej@jCZDGd d!ej@jAZEeDeEd"ZFGd#d$ej@jAZGGd%d&ej@jAZHGd'd(ej@jAZIGd)d*ej@jAZJGd+d,ej@jAZKGd-d.ej@jAZLGd/d0ej@jAZMGd1d2ej@jAZNGd3d4ej@jAZOGd5d6ej@jAZPGd7d8ej@jAZQGd9d:ej@jAZRGd;dej@jAZTGd?d@ej@jAZUGdAdBej@jAZVe GdCdDej@jAZWGdEdFeZXeGdGdHe'ZYdIZZdJZ[e)dKeZGdLdMeXZ\e)dNeZGdOdPeXe?Z]e)dQeZGdRdSeXeZ^GdTdUej@jAZ_e)dVeZGdWdXeXeZ`e)dYeZGdZd[eXeZae)d\eZGd]d^eXeZbe)d_eZGd`daeXeZce)dbeZGdcddeXeZdgdeZedS)fzTF 2.0 MobileBERT model.) annotationsN) dataclass)get_tf_activation)TFBaseModelOutputTFBaseModelOutputWithPoolingTFMaskedLMOutputTFMultipleChoiceModelOutputTFNextSentencePredictorOutputTFQuestionAnsweringModelOutputTFSequenceClassifierOutputTFTokenClassifierOutput) TFMaskedLanguageModelingLossTFModelInputTypeTFMultipleChoiceLossTFNextSentencePredictionLossTFPreTrainedModelTFQuestionAnsweringLossTFSequenceClassificationLossTFTokenClassificationLossget_initializerkeraskeras_serializable unpack_inputs)check_embeddings_within_bounds shape_liststable_softmax) ModelOutputadd_code_sample_docstringsadd_start_docstrings%add_start_docstrings_to_model_forwardloggingreplace_return_docstrings)MobileBertConfigzgoogle/mobilebert-uncasedr%z"vumichien/mobilebert-finetuned-nerzK['I-ORG', 'I-ORG', 'O', 'O', 'O', 'O', 'O', 'I-LOC', 'O', 'I-LOC', 'I-LOC']gQ?z%vumichien/mobilebert-uncased-squad-v2z'a nice puppet'gףp= @ zvumichien/emo-mobilebertz'others'z4.72ceZdZdZddZdS) TFMobileBertPreTrainingLossz Loss function suitable for BERT-like pretraining, that is, the task of pretraining a language model by combining NSP + MLM. .. note:: Any label of -100 will be ignored (along with the corresponding logits) in the loss computation. labels tf.Tensorlogitsreturnctjdtjjj}|t j|d|d}t j|ddk|j }||z}t j |t j |z }|t j|d|d }t j|ddk|j } || z} t j | t j | z } t j || zd S) NT) from_logits reductionr*r)y_truey_predidtypenext_sentence_labelr$)r$) rlossesSparseCategoricalCrossentropy ReductionNONEtfnnrelucastr4 reduce_sumreshape) selfr*r,loss_fnunmasked_lm_losses lm_loss_maskmasked_lm_lossesreduced_masked_lm_lossunmasked_ns_loss ns_loss_maskmasked_ns_lossreduced_masked_ns_losss /home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/transformers/models/mobilebert/modeling_tf_mobilebert.pyhf_compute_lossz+TFMobileBertPreTrainingLoss.hf_compute_loss^s>,<<Y^YeYoYt<uu%WBEJJvh7G,H,HQWXYQZ[[[wvh/47?Q?WXXX - <!#/?!@!@2=Q]C^C^!^#7"%**V!>|A\A\!\z03II4PPPN)r*r+r,r+r-r+)__name__ __module__ __qualname____doc__rKrLrJr)r)Ws8 QQQQQQrLr)c,eZdZfdZdZddZxZS)TFMobileBertIntermediatec tjdi|tj|jd|_t|jtrt|j|_ n |j|_ ||_ dS)NdensenamerQ) super__init__rlayersDenseintermediate_sizerU isinstance hidden_actstrrintermediate_act_fnconfigr@rakwargs __class__s rJrYz!TFMobileBertIntermediate.__init__ts""6"""\''(@w'OO f' - - 9'89J'K'KD $ $'-'8D $ rLcZ||}||}|SN)rUr`r@ hidden_statess rJcallzTFMobileBertIntermediate.calls, =11 00?? rLNc|jrdSd|_t|dd`tj|jj5|jdd|jjgddddS#1swxYwYdSdSNTrU) builtgetattrr: name_scoperUrWbuildratrue_hidden_sizer@ input_shapes rJrozTFMobileBertIntermediate.builds :  F 4$ ' ' 3tz// M M   $dk.J!KLLL M M M M M M M M M M M M M M M M M M 4 3(A55A9<A9rfrMrNrOrYriro __classcell__rds@rJrSrSssc      MMMMMMMMrLrSc*eZdZfdZdfd ZxZS) TFLayerNormcH||_tj|i|dSrf) feat_sizerXrY)r@rzargsrcrds rJrYzTFLayerNorm.__init__s,"$)&)))))rLNcZtdd|jgdSrf)rXrorzr@rrrds rJrozTFLayerNorm.builds(  tT4>233333rLrf)rMrNrOrYrorurvs@rJrxrxsV*****4444444444rLrxc2eZdZdfd ZfdZddZxZS) TFNoNormNc Htjdi|||_dS)NrQ)rXrYrz)r@rzepsilonrcrds rJrYzTFNoNorm.__init__s+""6""""rLc|d|jgd|_|d|jgd|_t |dS)Nbiaszeros)shape initializerweightones) add_weightrzrrrXror}s rJrozTFNoNorm.builds]OOF4>2BPWOXX ooht~6FTZo[[   k"""""rLinputsr+c&||jz|jzSrf)rr)r@rs rJriz TFNoNorm.calls #di//rLrf)rr+)rMrNrOrYrorirurvs@rJrrsj########### 00000000rLr) layer_normno_normc2eZdZdZfdZddZddZxZS) TFMobileBertEmbeddingszGConstruct the embeddings from word, position and token_type embeddings.c tjdi||j|_|j|_||_|j|_|j|_|j|_tj |jd|_ t|j |j|jd|_tj |j|_|j|jrdndz|_dS) Nembedding_transformationrV LayerNormrrW)raterr$rQ)rXrY trigram_inputembedding_sizera hidden_sizemax_position_embeddingsinitializer_rangerrZr[rNORM2FNnormalization_typelayer_norm_epsrDropouthidden_dropout_probdropoutembedded_input_sizerbs rJrYzTFMobileBertEmbeddings.__init__s""6"""#1$3 !-'-'E$!'!9(- (:(:6;MTn(:(o(o%!!:;  (=K   |++1K+LL #'#6t?Q:X!!WX#Y   rLNcDtjd5|d|jj|jgt |j|_dddn #1swxYwYtjd5|d|jj |j gt |j|_ dddn #1swxYwYtjd5|d|j |j gt |j|_ dddn #1swxYwY|jrdSd|_t|d dXtj|jj5|jdd|jgdddn #1swxYwYt|d dStj|jj5|jdddddS#1swxYwYdSdS) Nword_embeddingsr)r)rWrrtoken_type_embeddings embeddingsposition_embeddingsTrr)r:rnrra vocab_sizerrrrtype_vocab_sizerrrrrlrmrrWrorrrqs rJrozTFMobileBertEmbeddings.buildsc ], - -  //{-t/BC+d>TUUU*DK               ]2 3 3  )-!{2D4DE+d>TUUU*9**D &               ]0 1 1  '+!3T5EF+d>TUUU(7((D $                :  F 43T : : Ft<ABB \ \-33T4AY4Z[[[ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ 4d + + 7t~233 + +$$T*** + + + + + + + + + + + + + + + + + + 8 7s[AA##A'*A'ACCC/=D88D<?D<#F11F58F5+HHHFc ,||J|5t||jjtj|j|}t |dd}|tj|d}|jrWtj tj |ddddfd|tj |ddddfdgd }|js|j |j kr| |}|0tjtjd|d d }tj|j|}tj|j|}||z|z} || } || | } | 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)rrarr:gatherrrfillrconcatpadrrr expand_dimsrangerrrr) r@ input_ids position_idstoken_type_ids inputs_embedsrrrposition_embedstoken_type_embedsfinal_embeddingss rJrizTFMobileBertEmbeddings.calls%-*?*?@  *9dk6L M M MIT[)LLLM //4  !W+Q???N   IF=ABB/1IJJ!F=CRC02JKK  M   I!48H!H!H 99-HHM  >"(+b/*R*R*RYZ[[)4+C\ZZZIT-GQ_```(?:=NN>>1A>BB<>!CCCC#&v'>A['[#\#\ !58PP\''  ?6C[3\3\cj(   <%%  ?6C[3\3\ch&  \''  ?6C[3\3\cj(   |++F,OPP  rLcxtj||d|j|jf}tj|gdS)Nrrrr$rperm)r:r?rr transpose)r@x batch_sizes rJtranspose_for_scoresz.TFMobileBertSelfAttention.transpose_for_scores,s; Jq:r4+CTE]^ _ _|ALLL1111rLFcHt|d}||} ||} ||} || |} || |} || |}t j| | d}t jt| d|j}|t j |z }| t j||j}||z}t|d}| ||}|||z}t j||}t j |gd }t j||d|jf}|r||fn|f}|S) NrT) transpose_brr3rrrr)rrrrrr:matmulr=r4mathsqrtrrrr?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 rJrizTFMobileBertSelfAttention.call1s //2  JJ|44((:.. JJ|44//0A:NN --ozJJ //0A:NN 9    WZ **2.6F6L M M M+bgll2.>.>>  %W^;K;QRRRN/.@ ))9CCC,,,JJ  - 9O /;??  ]FFF  JD,>?  7H]=/22mM]rLNc|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|dd]tj|j j5|j dd|jjgdddn #1swxYwYt|ddxtj|j j5|j dd|jj r |jjn |jj gddddS#1swxYwYdSdS)NTrrr) rlrmr:rnrrWrorarprruse_bottleneck_attentionrrqs rJrozTFMobileBertSelfAttention.build^sH :  F 4$ ' ' 3tz// M M   $dk.J!KLLL M M M M M M M M M M M M M M M 4 % % 1tx}-- K KdDK,HIJJJ K K K K K K K K K K K K K K K 4$ ' ' 3tz//     ;?5 44![4                    4 3s7(A44A8;A8.(C""C&)C&AE))E-0E-Frf)rMrNrOrYrrirorurvs@rJrrsw6222 ns++++ZrLrc.eZdZfdZddZddZxZS)TFMobileBertSelfOutputc tjdi||j|_tj|jt|jd|_ t|j |j|j d|_ |js)tj|j|_||_dS)NrUrrrrQ)rXrYuse_bottleneckrrZr[rprrrUrrrrrrrrarbs rJrYzTFMobileBertSelfOutput.__init__vs""6"""$3\''  #H`8a8aho(   !!:;  #V-B   " L <//0JKKDL rLFc||}|js|||}|||z}|SNr)rUrrr)r@rhresidual_tensorrs rJrizTFMobileBertSelfOutput.callsP =11 " K LLLJJM}'FGG rLNc|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|ddStj|j j5|j dddddS#1swxYwYdSdSNTrUr rlrmr:rnrUrWrorarprrqs rJrozTFMobileBertSelfOutput.buildo :  F 4$ ' ' 3tz// M M   $dk.J!KLLL M M M M M M M M M M M M M M M 4d + + 7t~233 + +$$T*** + + + + + + + + + + + + + + + + + + 8 7$(A44A8;A8.CCCrrfrtrvs@rJrrus`      + + + + + + + +rLrc6eZdZfdZdZ ddZddZxZS) TFMobileBertAttentionc tjdi|t|d|_t |d|_dS)Nr@rVoutputrQ)rXrYrr@rmobilebert_outputrbs rJrYzTFMobileBertAttention.__init__sO""6"""-f6BBB !7X!N!N!NrLctrfNotImplementedError)r@headss rJ prune_headsz!TFMobileBertAttention.prune_headss!!rLFc ||||||||} || d||} | f| ddz} | S)Nrrr$)r@r) r@rrr layer_inputrrrr self_outputsattention_outputrs rJrizTFMobileBertAttention.callsjyy *lNIO`ks!   11,q/;Ya1bb#% QRR(88rLNc|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|ddStj|jj5|jdddddS#1swxYwYdSdS)NTr@r)rlrmr:rnr@rWrorrqs rJrozTFMobileBertAttention.buildsT :  F 4 & & 2ty~.. & & %%% & & & & & & & & & & & & & & & 4,d 3 3 ?t5:;; 3 3&,,T222 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 @ ?$A''A+.A+!C  C C rrf)rMrNrOrYrrirorurvs@rJrrsyOOOOO """& 3 3 3 3 3 3 3 3rLrc.eZdZfdZddZddZxZS)TFOutputBottleneckc Htjdi|tj|jd|_t|j|j|j d|_ tj |j |_ ||_dSNrUrVrrrQ)rXrYrrZr[rrUrrrrrrrrarbs rJrYzTFOutputBottleneck.__init__s""6"""\''(:'II  !:;  (=K   |++F,FGG  rLFc||}|||}|||z}|Sr)rUrr)r@rhrr layer_outputss rJrizTFOutputBottleneck.callsD =11  ]X FF }'FGG rLNc|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|ddStj|j j5|j dddddS#1swxYwYdSdSrrrqs rJrozTFOutputBottleneck.buildrrrrfrtrvs@rJrrs` + + + + + + + +rLrc.eZdZfdZddZddZxZS)TFMobileBertOutputc tjdi||j|_tj|jt|jd|_ t|j |j|j d|_ |js*tj|j|_nt#|d|_||_dS)NrUrrr bottleneckrVrQ)rXrYrrrZr[rprrrUrrrrrrrrrrarbs rJrYzTFMobileBertOutput.__init__s""6"""$3\''  #H`8a8aho(   !!:;  #V-B   " L <//0JKKDLL0lKKKDO rLFc||}|js0|||}|||z}n.|||z}|||}|Sr)rUrrrr)r@rhresidual_tensor_1residual_tensor_2rs rJrizTFMobileBertOutput.calls~ =11 " N LLLJJM NN=;L+LMMMM NN=;L+LMMM OOM;LMMMrLNc|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|ddPtj|j j5|j ddddn #1swxYwYt|ddStj|j j5|j dddddS#1swxYwYdSdS)NTrUrr) rlrmr:rnrUrWrorar\rrrqs rJrozTFMobileBertOutput.builds :  F 4$ ' ' 3tz// N N   $dk.K!LMMM N N N N N N N N N N N N N N N 4d + + 7t~233 + +$$T*** + + + + + + + + + + + + + + + 4t , , 8t344 , ,%%d+++ , , , , , , , , , , , , , , , , , , 9 8s6(A44A8;A8.CCCD77D;>D;rrfrtrvs@rJrrs`      , , , , , , , ,rLrc,eZdZfdZdZddZxZS)TFBottleneckLayerc tjdi|tj|jd|_t|j|j|j d|_ ||_ dSr) rXrYrrZr[intra_bottleneck_sizerUrrrrrarbs rJrYzTFBottleneckLayer.__init__sw""6"""\''(D7'SS  !:;  (&2Gk    rLcZ||}||}|SrfrUr)r@rrhs rJrizTFBottleneckLayer.call s* 6** }55 rLNc|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|ddStj|j j5|j dddddS#1swxYwYdSdSr rlrmr:rnrUrWrorarrrqs rJrozTFBottleneckLayer.buildo :  F 4$ ' ' 3tz// H H   $dk.E!FGGG H H H H H H H H H H H H H H H 4d + + 7t~233 + +$$T*** + + + + + + + + + + + + + + + + + + 8 7rrfrtrvs@rJrr[ + + + + + + + +rLrc,eZdZfdZdZddZxZS) TFBottleneckc tjdi||j|_|j|_t |d|_|jrt |d|_dSdS)NinputrV attentionrQ)rXrYkey_query_shared_bottleneckrrbottleneck_inputr*rbs rJrYzTFBottleneck.__init__s{""6"""+1+M((.(G% 1&w G G G  + I.vKHHHDNNN I IrLc||}|jr|fdzS|jr||}||||fS||||fS)N)r,rr+r*)r@rhbottlenecked_hidden_statesshared_attention_inputs rJrizTFBottleneck.call&sq"&*%:%:=%I%I"  ( ].014 4  - ]%)^^M%B%B "*,BMSmn n!=-A[\ \rLNc|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|ddStj|jj5|jdddddS#1swxYwYdSdS)NTr,r*)rlrmr:rnr,rWror*rqs rJrozTFBottleneck.build@sV :  F 4+T 2 2 >t49:: 2 2%++D111 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 4d + + 7t~233 + +$$T*** + + + + + + + + + + + + + + + + + + 8 7r rfrtrvs@rJr'r'scIIIII]]]4 + + + + + + + +rLr'c,eZdZfdZdZddZxZS) TFFFNOutputc tjdi|tj|jd|_t|j|j|j d|_ ||_ dSr) rXrYrrZr[rprUrrrrrarbs rJrYzTFFFNOutput.__init__Msw""6"""\''(?g'NN  !:;  #V-B    rLc`||}|||z}|Srfr!)r@rhrs rJrizTFFFNOutput.callUs/ =11 }'FGG rLNc|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|ddStj|j j5|j dddddS#1swxYwYdSdSr) rlrmr:rnrUrWrorar\rrqs rJrozTFFFNOutput.buildZso :  F 4$ ' ' 3tz// N N   $dk.K!LMMM N N N N N N N N N N N N N N N 4d + + 7t~233 + +$$T*** + + + + + + + + + + + + + + + + + + 8 7rrfrtrvs@rJr3r3Lr%rLr3c,eZdZfdZdZddZxZS) TFFFNLayerc tjdi|t|d|_t |d|_dS)N intermediaterVrrQ)rXrYrSr:r3rrbs rJrYzTFFFNLayer.__init__gsP""6"""4V.QQQ!,V(!C!C!CrLc\||}|||}|Srf)r:r)r@rhintermediate_outputrs rJrizTFFFNLayer.callls2"// >>../BMRR rLNc|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|ddStj|jj5|jdddddS#1swxYwYdSdS)NTr:r)rlrmr:rnr:rWrorrqs rJrozTFFFNLayer.buildqsY :  F 4 . . :t0566 . .!''--- . . . . . . . . . . . . . . . 4,d 3 3 ?t5:;; 3 3&,,T222 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 @ ?r rfrtrvs@rJr8r8fs`DDDDD  3 3 3 3 3 3 3 3rLr8c.eZdZfdZddZddZxZS)TFMobileBertLayerc tjdi|j|_j|_t d|_t d|_td|_ |jrtd|_ jdkr*fdtjdz D|_ dSdS) Nr*rVr:rrr$c8g|]}td|S)zffn.rV)r8.0iras rJ z.TFMobileBertLayer.__init__..s,ppp 6 q ;;;ppprLrQ)rXrYrnum_feedforward_networksrr*rSr:rrr'rrffnrbs ` rJrYzTFMobileBertLayer.__init__~s""6"""$3(.(G%.vKHHH4V.QQQ!3F!J!J!J   F*6 EEEDO  *Q . .ppppU6KjmnKnEoEopppDHHH / .rLFc |jr||\}}}} n |gdz\}}}} ||||| ||||} | d} | f} |jdkr+t |jD]\} }|| } | | fz } || }||| ||}|f| ddztj d|||| | |fz| z}|S)Nr.rrr$) rrr*rF enumeraterGr:rr:constant)r@rhrrrrrrrrattention_outputsr srD ffn_moduler< layer_outputrs rJrizTFMobileBertLayer.calls_   VBF//R_B`B` ?L*lKKCP/TUBU ?L*lK NN       +   -Q/    (A - -!*48!4!4 ) ) :#-:.>#?#? &(("//0@AA--.ACSUbmu-vv O# $ A #   rLNc|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|ddPtj|jj5|jddddn #1swxYwYt|ddPtj|jj5|jddddn #1swxYwYt|ddPtj|j j5|j ddddn #1swxYwYt|ddP|j D]J}tj|j5|ddddn #1swxYwYIdSdS)NTr*r:rrrG) rlrmr:rnr*rWror:rrrGr@rrlayers rJrozTFMobileBertLayer.builds :  F 4d + + 7t~233 + +$$T*** + + + + + + + + + + + + + + + 4 . . :t0566 . .!''--- . . . . . . . . . . . . . . . 4,d 3 3 ?t5:;; 3 3&,,T222 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 4t , , 8t344 , ,%%d+++ , , , , , , , , , , , , , , , 4 % % 1 & &]5:..&&KK%%%&&&&&&&&&&&&&&& 2 1 & &sZA''A+.A+!CC C D))D-0D-#F  FFG**G. 1G. rrfrtrvs@rJr?r?}sf q q q q q++++Z&&&&&&&&rLr?c0eZdZfdZ ddZddZxZS)TFMobileBertEncoderc tjdi|j|_j|_fdt jD|_dS)Nc8g|]}td|S)zlayer_._rV)r?rBs rJrEz0TFMobileBertEncoder.__init__..s-nnn'^^^DDDnnnrLrQ)rXrYroutput_hidden_statesrnum_hidden_layersrQrbs ` rJrYzTFMobileBertEncoder.__init__sb""6"""!'!9$*$?!nnnneTZTlNmNmnnn rLFc&|rdnd}|rdnd} t|jD]9\} } |r||fz}| |||| ||} | d}|r | | dfz} :|r||fz}|std||| fDSt||| S)NrQrrr$c3K|]}||V dSrfrQ)rCvs rJ z+TFMobileBertEncoder.call..s(hhqZ[ZgZgZgZgZghhrL)last_hidden_staterh attentions)rIrQtupler) r@rhrrrrV return_dictrall_hidden_statesall_attentionsrD layer_modulers rJrizTFMobileBertEncoder.calls #7@BBD0:d(44 F FOA|# I$58H$H!(L~y|=NYaM*!,M  F!/=3C2E!E  E 1]4D D  ihh]4E~$Vhhhhh h +;LYg    rLNc|jrdSd|_t|ddP|jD]J}tj|j5|ddddn #1swxYwYIdSdS)NTrQ)rlrmrQr:rnrWrorPs rJrozTFMobileBertEncoder.builds :  F 4$ ' ' 3 & &]5:..&&KK%%%&&&&&&&&&&&&&&& 4 3 & &sA&&A* -A* rrfrtrvs@rJrSrSskooooo! ! ! ! F&&&&&&&&rLrSc,eZdZfdZdZddZxZS)TFMobileBertPoolerc tjdi||j|_|jr?tj|jt|j dd|_ ||_ dS)NtanhrU)r activationrWrQ) rXrYclassifier_activation do_activaterrZr[rrrrUrarbs rJrYzTFMobileBertPooler.__init__s{""6"""!7   ++"#263K#L#L! ,DJ  rLcZ|dddf}|js|S||}|SNr)rjrU)r@rhfirst_token_tensor pooled_outputs rJrizTFMobileBertPooler.calls@+111a40 !% % JJ'9::M rLNc|jrdSd|_t|dd`tj|jj5|jdd|jjgddddS#1swxYwYdSdSrk) rlrmr:rnrUrWrorarrqs rJrozTFMobileBertPooler.builds :  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 3rsrfrtrvs@rJreresc     !!!HHHHHHHHrLrec,eZdZfdZdZddZxZS)#TFMobileBertPredictionHeadTransformc tjdi|tj|jt |jd|_t|j trt|j |_ n |j |_ td|j|jd|_||_dS)NrUrrrrrQ)rXrYrrZr[rrrrUr]r^r_rtransform_act_fnrrrrarbs rJrYz,TFMobileBertPredictionHeadTransform.__init__#s""6"""\''  ?6C[3\3\cj(    f' - - 6$5f6G$H$HD ! !$*$5D ! .v/A6K`grsss rLc||}||}||}|Srf)rUrsrrgs rJriz(TFMobileBertPredictionHeadTransform.call/s= =11 --m<< }55 rLNc|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|ddStj|j j5|j dddddS#1swxYwYdSdSrr#rqs rJroz)TFMobileBertPredictionHeadTransform.build5r$rrfrtrvs@rJrqrq"s[      + + + + + + + +rLrqcDeZdZfdZd dZdZdZdZdZdZ xZ S) TFMobileBertLMPredictionHeadc ttjdi|t|d|_||_dS)N transformrVrQ)rXrYrqryrarbs rJrYz%TFMobileBertLMPredictionHead.__init__Bs?""6""">>rLrwc,eZdZfdZdZddZxZS)TFMobileBertMLMHeadc ftjdi|t|d|_dS)N predictionsrVrQ)rXrYrwrrbs rJrYzTFMobileBertMLMHead.__init__ss;""6"""7]SSSrLc0||}|Srfr)r@sequence_outputprediction_scoress rJrizTFMobileBertMLMHead.callws ,,_==  rLNc|jrdSd|_t|ddStj|jj5|jdddddS#1swxYwYdSdS)NTr)rlrmr:rnrrWrorqs rJrozTFMobileBertMLMHead.build{s :  F 4 - - 9t/455 - - &&t,,, - - - - - - - - - - - - - - - - - - : 9A((A,/A,rfrtrvs@rJrrrs`TTTTT!!!--------rLrcjeZdZeZd fd ZdZdZdZe d dZ d d Z xZ S) TFMobileBertMainLayerTc 4tjdi|||_|j|_|j|_|j|_|j|_t|d|_ t|d|_ |rt|dnd|_ dS)NrrVencoderpoolerrQ)rXrYrarWrrVuse_return_dictr_rrrSrrer)r@raadd_pooling_layerrcrds rJrYzTFMobileBertMainLayer.__init__s""6""" !'!9!'!9$*$?!!10lKKK*6 BBB CT^(h????Z^ rLc|jSrf)rr~s rJget_input_embeddingsz*TFMobileBertMainLayer.get_input_embeddingss rLc\||j_t|d|j_dSrl)rrrrrs rJset_input_embeddingsz*TFMobileBertMainLayer.set_input_embeddingss)!&%/%6%6q%9"""rLct)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 rJ _prune_headsz"TFMobileBertMainLayer._prune_headss "!rLNFc r||td|t|} n)|t|dd} ntd|tj| d}|tj| d}|||||| } tj|| ddd| df} tj| | j} tjd| j}tjd | j}tj tj || |} |tdg|j z}| | | |||| | }|d}|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_outputrhr])rrr:rrr?r=r4rJmultiplysubtractrrWrrrrhr])r@rrrrrrrrVr_rrrembedding_outputextended_attention_maskone_cstten_thousand_cstencoder_outputsrrns rJrizTFMobileBertMainLayer.calls   ]%>cdd d  "$Y//KK  &$]33CRC8KKTUU U  !W[!44N  !W[!44N??9lNTalt?uu#%*^k!naQRT_`aTb=c"d"d#%'*AIYI_"`"`"`+c)9)?@@@;x7G7MNNN"$+bk'CZ.[.[]m"n"n  % %!77I,,  #   '  *!,8< 8O O444UY  $ #$ $ ,-')7&1     rLcr|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|ddPtj|jj5|jddddn #1swxYwYt|ddStj|jj5|jdddddS#1swxYwYdSdS)NTrrr) rlrmr:rnrrWrorrrqs rJrozTFMobileBertMainLayer.builds :  F 4t , , 8t344 , ,%%d+++ , , , , , , , , , , , , , , , 4D ) ) 5t|011 ) ) ""4((( ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) 44 ( ( 4t{/00 ( ( !!$''' ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( 5 46A''A+.A+!CC C D**D.1D.)T NNNNNNNNNFrf) rMrNrOr% config_classrYrrrrrirorurvs@rJrrs#L _ _ _ _ _ _:::"""!Q Q Q ]Q f ( ( ( ( ( ( ( (rLrceZdZdZeZdZdS)TFMobileBertPreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. mobilebertN)rMrNrOrPr%rbase_model_prefixrQrLrJrrs' $L$rLrcZeZdZUdZdZded<dZded<dZded<dZded<dZ ded <dS) 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 | Nonelossprediction_logitsseq_relationship_logitsztuple[tf.Tensor] | Nonerhr]) rMrNrOrPr__annotations__rrrhr]rQrLrJrrsu,"D!!!!*.....044444-1M1111*.J......rLra 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.ceZdZfdZeeedee e e  dddZ ddZ xZS)TFMobileBertModelcntj|g|Ri|t|d|_dS)NrrV)rXrYrrr@rarrcrds rJrYzTFMobileBertModel.__init__sB3&333F333/\JJJrLbatch_size, sequence_length checkpoint output_typerNFrTFModelInputType | Nonernp.ndarray | tf.Tensor | Nonerrrrr bool | NonerVr_rr-$tuple | TFBaseModelOutputWithPoolingc D|||||||||| |  } | S)N) rrrrrrrrVr_r)r) r@rrrrrrrrVr_rrs rJrizTFMobileBertModel.callsA(//))%'/!5#"   rLc|jrdSd|_t|ddStj|jj5|jdddddS#1swxYwYdSdS)NTr)rlrmr:rnrrWrorqs rJrozTFMobileBertModel.builds :  F 4t , , 8t344 , ,%%d+++ , , , , , , , , , , , , , , , , , , 9 8rr)rrrrrrrrrrrrrrrVrr_rrrr-rrf)rMrNrOrYrr!MOBILEBERT_INPUTS_DOCSTRINGformatr_CHECKPOINT_FOR_DOCr_CONFIG_FOR_DOCrirorurvs@rJrrs KKKKK**+F+M+MNk+l+lmm&0$.28<8<6:377;)-,0#' % nm]8,,,,,,,,rLrz MobileBert Model with two heads on top as done during the pretraining: a `masked language modeling` head and a `next sentence prediction (classification)` head. ceZdZfdZdZdZeee de e e  dddZ ddZdZxZS)TFMobileBertForPreTrainingctj|g|Ri|t|d|_t |d|_t |d|_dS)NrrVpredictions___clsseq_relationship___cls)rXrYrrrrTFMobileBertOnlyNSPHeadseq_relationshiprs rJrYz#TFMobileBertForPreTraining.__init__sp3&333F333/\JJJ.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] ``` rrrrrrrVr_rNrr*r5r*r,)rrrrhr])rrrrKrrhr])r@rrrrrrrrVr_r*r5rrrrnrseq_relationship_score total_lossd_labelsrs rJrizTFMobileBertForPreTraining.calls B// ))%'/!5#"   *1!& ,,_==!%!6!6}!E!E  "5"A &)H.AH* +--XGXZpFq-rrJ R')?@7122;NF/9/EZMF**6 Q//$:!/)     rLcr|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|ddPtj|jj5|jddddn #1swxYwYt|ddStj|jj5|jdddddS#1swxYwYdSdS)NTrrr) rlrmr:rnrrWrorrrqs rJroz TFMobileBertForPreTraining.builds :  F 4t , , 8t344 , ,%%d+++ , , , , , , , , , , , , , , , 4 - - 9t/455 - - &&t,,, - - - - - - - - - - - - - - - 4+T 2 2 >t49:: 2 2%++D111 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 ? >rc|dkr|dfS|fSNzcls.predictions.decoder.weightz,mobilebert.embeddings.word_embeddings.weightrQr@ tf_weights rJtf_to_pt_weight_renamez1TFMobileBertForPreTraining.tf_to_pt_weight_rename-" 8 8 8LL L< rL NNNNNNNNNNNF)rrrrrrrrrrrrrrrVrr_rr*rr5rrrr-rrf)rMrNrOrYrrrr!rrr#rrrirorrurvs@rJrrs_____ ,,,aaa**+F+M+MNk+l+lmm+KZijjj.28<8<6:377;)-,0#'04=A %? ? ? ? kjnm]? B 2 2 2 2       rLrz8MobileBert Model with a `language modeling` head on top.c eZdZgdZfdZdZdZeee de e e edd dddZd dZdZxZS)!TFMobileBertForMaskedLM)rrcls.seq_relationshipctj|g|Ri|t|dd|_t |d|_dS)NFrrrWrrV)rXrYrrrrrs rJrYz TFMobileBertForMaskedLM.__init__=s]3&333F333/%Vbccc.vt49:: S S%++T49P,QRRR S S S S S S S S S S S S S S S S S S ? >rsrfrtrvs@rJrrsc &&&SSSSSSSSrLrzPMobileBert Model with a `next sentence prediction (classification)` head on top.ceZdZddgZfdZeeede e e  dddZ ddZ xZS)%TFMobileBertForNextSentencePredictionrcls.predictionsctj|g|Ri|t|d|_t |d|_dS)NrrVr)rXrYrrrclsrs rJrYz.TFMobileBertForNextSentencePrediction.__init__sW3&333F333/\JJJ*68PQQQrLrrNFrrrrrrrrrrrVr_r5rr-%tuple | TFNextSentencePredictorOutputc |||||||||| |  } | d} || }| dn|| |}| s|f| ddz}||f|zn|St||| j| jS)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)rrrKr rhr])r@rrrrrrrrVr_r5rrrnseq_relationship_scoresnext_sentence_lossrs rJriz*TFMobileBertForNextSentencePrediction.callsF// ))%'/!5#"     "&((="9"9#* D%%-@I`%aa   b-/'!""+=F7I7U')F22[a a,#*!/)     rLc|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|ddStj|jj5|jdddddS#1swxYwYdSdS)NTrr)rlrmr:rnrrWrorrqs rJroz+TFMobileBertForNextSentencePrediction.buildsP :  F 4t , , 8t344 , ,%%d+++ , , , , , , , , , , , , , , , 4 % % 1tx}-- % %t$$$ % % % % % % % % % % % % % % % % % % 2 1r r)rrrrrrrrrrrrrrrVrr_rr5rrrr-rrf)rMrNrOrrYrr!rrr#r rrirorurvs@rJrrs +?@R)S&RRRRR **+F+M+MNk+l+lmm+HWfggg.28<8<6:377;)-,0#'=A %> > > > hgnm]> @ % % % % % % % %rLrz 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 eZdZgdZdgZfdZeee de e e e ee dddZddZxZS)%TFMobileBertForSequenceClassificationrrrrrctj|g|Ri||j|_t|d|_|j|jn|j}tj ||_ tj |jt|j d|_||_dS)NrrV classifierrrXrY num_labelsrrclassifier_dropoutrrrZrrr[rrrrar@rarrcrrds rJrYz.TFMobileBertForSequenceClassification.__init__s3&333F333 +/\JJJ)/)B)NF % %TZTn |++,>?? ,,,  /&BZ2[2[bn-   rLrrNFrrrrrrrrrrrVr_r*rr-"tuple | TFSequenceClassifierOutputc J|||||||||| |  } | d} || | } || }| dn|| |}| s|f| ddz}||f|zn|St ||| j| jS)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$rNrr)rrrrKr rhr])r@rrrrrrrrVr_r*rrrnr,rrs rJriz*TFMobileBertForSequenceClassification.call$s:// ))%'/!5#"      ]X FF //~tt4+?+?+O+O FY,F)-)9TGf$$v E)!/)     rLc|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|dd`tj|jj5|jdd|jj gddddS#1swxYwYdSdSNTrr rlrmr:rnrrWrorrarrqs rJroz+TFMobileBertForSequenceClassification.build_s :  F 4t , , 8t344 , ,%%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+!(CCCr)rrrrrrrrrrrrrrrVrr_rr*rrrr-rrf)rMrNrOr_keys_to_ignore_on_load_missingrYrr!rrr'_CHECKPOINT_FOR_SEQUENCE_CLASSIFICATIONr r_SEQ_CLASS_EXPECTED_OUTPUT_SEQ_CLASS_EXPECTED_LOSSrirorurvs@rJr r s***& (2l#     **+F+M+MNk+l+lmm:.$2. .28<8<6:377;)-,0#'04 %0 0 0 0 nm]0 d M M M M M M M MrLr z 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 eZdZgdZfdZeeede e e e e eee dddZddZxZS) TFMobileBertForQuestionAnsweringrrrrrctj|g|Ri||j|_t|dd|_t j|jt|j d|_ ||_ dS)NFrr qa_outputsr) rXrYrrrrrZr[rrr"rars rJrYz)TFMobileBertForQuestionAnswering.__init__|s3&333F333 +/%Vbccc,,,  /&BZ2[2[bn-   rLr)rrrqa_target_start_indexqa_target_end_indexrrNFrrrrrrrrrrrVr_start_positions end_positionsrr-&tuple | TFQuestionAnsweringModelOutputc |||||||||| |  } | d}||}tj|dd\}}tj|d}tj|d}d}| | | | d}||||f}| s||f| ddz}||f|zn|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_logitsrhr]) rr"r:splitsqueezerKr rhr])r@rrrrrrrrVr_r%r&rrrr,r+r,rr*rs rJriz%TFMobileBertForQuestionAnswering.calls4H// ))%'/!5#"   "!*11#%8FAB#?#?#? jz,R888 Z 444   &=+D(7WWF''z0JKKD F"J/'!""+=F)-)9TGf$$v E-%!!/)     rLc|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|dd`tj|jj5|jdd|jj gddddS#1swxYwYdSdS)NTrr") rlrmr:rnrrWror"rarrqs rJroz&TFMobileBertForQuestionAnswering.buildrrr)rrrrrrrrrrrrrrrVrr_rr%rr&rrrr-r'rf)rMrNrOrrYrr!rrr_CHECKPOINT_FOR_QAr r_QA_TARGET_START_INDEX_QA_TARGET_END_INDEX_QA_EXPECTED_OUTPUT_QA_EXPECTED_LOSSrirorurvs@rJrrks ***&**+F+M+MNk+l+lmm%2$40+'.28<8<6:377;)-,0#'9=7; %; ; ; ; nm]; z M M M M M M M MrLrz MobileBert Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a softmax) e.g. for RocStories/SWAG tasks. ceZdZgdZdgZfdZeee de e e e  dddZddZxZS)TFMobileBertForMultipleChoicer rc@tj|g|Ri|t|d|_tj|j|_tj dt|j d|_ ||_ dS)NrrVr$rr)rXrYrrrrZrrrr[rrrrars rJrYz&TFMobileBertForMultipleChoice.__init__s3&333F333/\JJJ|++F,FGG ,,, /&2J"K"KR^-   rLz(batch_size, num_choices, sequence_lengthrNFrrrrrrrrrrrVr_r*rr-#tuple | TFMultipleChoiceModelOutputc \|+t|d} t|d} n*t|d} t|d} |tj|d| fnd}|tj|d| fnd}|tj|d| fnd}|tj|d| fnd}|+tj|d| t|dfnd}|||||||||| |  }|d}||| }||}tj|d| f}| dn|| |}| s|f|ddz}||f|zn|St|||j|j S) 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?rrrrKr rhr])r@rrrrrrrrVr_r*r num_choices seq_lengthflat_input_idsflat_attention_maskflat_token_type_idsflat_position_idsflat_inputs_embedsrrnr,reshaped_logitsrrs rJriz"TFMobileBertForMultipleChoice.calls8  $Y//2K#I..q1JJ$]33A6K#M2215JDMDYIJ/?@@@_cN\Nhbj"j9IJJJnrN\Nhbj"j9IJJJnrJVJbBJ|b*5EFFFhl( J}r:z-7P7PQR7S&T U U U  //        #"      ]X FF //*Vb+->??~tt4+?+?+X+X F%''!""+5F)-)9TGf$$v E*"!/)     rLc|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|dd`tj|jj5|jdd|jj gddddS#1swxYwYdSdSrrrqs rJroz#TFMobileBertForMultipleChoice.build?rrr)rrrrrrrrrrrrrrrVrr_rr*rrrr-r8rf)rMrNrOrrrYrr!rrrrr rrirorurvs@rJr6r6s***& (2l#**#**+UVV &/$.28<8<6:377;)-,0#'04 %? ? ? ? ]? B M M M M M M M MrLr6z MobileBert Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. c eZdZgdZdgZfdZeee de e e e ee dddZddZxZS)"TFMobileBertForTokenClassificationr rctj|g|Ri||j|_t|dd|_|j|jn|j}tj ||_ tj |jt|j d|_||_dS)NFrrrrrrs rJrYz+TFMobileBertForTokenClassification.__init__]s3&333F333 +/%Vbccc)/)B)NF % %TZTn |++,>?? ,,,  /&BZ2[2[bn-   rLrrNFrrrrrrrrrrrVr_r*rr-tuple | TFTokenClassifierOutputc J|||||||||| |  } | d} || | } || }| dn|| |}| s|f| ddz}||f|zn|St ||| j| jS)z labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`. rrrNrr)rrrrKrrhr])r@rrrrrrrrVr_r*rrrr,rrs rJriz'TFMobileBertForTokenClassification.callks6// ))%'/!5#"   "!*,,,JJ11~tt4+?+?+O+O FY,F)-)9TGf$$v E&!/)     rLc|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|dd`tj|jj5|jdd|jj gddddS#1swxYwYdSdSrrrqs rJroz(TFMobileBertForTokenClassification.buildrrr)rrrrrrrrrrrrrrrVrr_rr*rrrr-rFrf)rMrNrOrrrYrr!rrr$_CHECKPOINT_FOR_TOKEN_CLASSIFICATIONrr_TOKEN_CLASS_EXPECTED_OUTPUT_TOKEN_CLASS_EXPECTED_LOSSrirorurvs@rJrDrDKs***&(2l#     **+F+M+MNk+l+lmm7+$40 .28<8<6:377;)-,0#'04 %. . . . nm]. ` M M M M M M M MrLrD) rr6rrrr rDrrr)frP __future__rr dataclassesrnumpynp tensorflowr:activations_tfrmodeling_tf_outputsrrr r r r r rmodeling_tf_utilsrrrrrrrrrrrrtf_utilsrrrutilsrrr r!r"r#configuration_mobilebertr% get_loggerrMloggerrrrIrJrKr0r3r4r1r2rrrr)rZLayerrSLayerNormalizationrxrrrrrrrrrr'r3r8r?rSrerqrwrrrrMOBILEBERT_START_DOCSTRINGrrrrrrr rr6rD__all__rQrLrJr]s  """"""!!!!!!//////                                                SRRRRRRRRR766666  H % %1$(L$l!='+E''!QQQQQQQQ8MMMMMu|1MMM644444%,1444 0 0 0 0 0u|! 0 0 0% : :e e e e e U\/e e e Pbbbbb 2bbbJ+++++U\/+++B%3%3%3%3%3EL.%3%3%3P+++++++++8&,&,&,&,&,+&,&,&,R+++++ *+++4,+,+,+,+,+5<%,+,+,+^+++++%,$+++433333#333.N&N&N&N&N& *N&N&N&b1&1&1&1&1&%,,1&1&1&hHHHHH+HHH@+++++%,*<+++>.....5<#5...b-----%,,---$~(~(~(~(~(EL.~(~(~(B%%%%%"3%%% /////{// /<(T5pj.,.,.,.,.,3.,., .,b d d d d d !<>Yd d d NTVpqq] ] ] ] ] 9;W] ] rq] @SSSSSel0SSS&ZV%V%V%V%V%,GIeV%V% V%r \M\M\M\M\M,GIe\M\M\M~ eMeMeMeMeM'BD[eMeMeMP gMgMgMgMgM$?AUgMgMgMT [M[M[M[M[M)DF_[M[M[M|   rL