T^(h0vdZddlmZddlZddlmZmZmZmZddl Z ddl Z ddl mZddlmZmZmZmZmZmZmZmZddlmZmZmZmZmZmZmZm 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-d d l.m/Z/e-j`e1Z2d Z3Gdde"jhjjZ6Gdde"jhjjZ7Gdde"jhjjZ8Gdde"jhjjZ9Gdde"jhjjZ:Gdde"jhjjZ;Gdde"jhjjZ<Gdde"jhjjZ=Gdde"jhjjZ>Gd d!e"jhjjZ?Gd"d#e"jhjjZ@e#Gd$d%e"jhjjZAGd&d'eZBd(ZCd)ZDe+d*eCGd+d,eBZEe+d-eCGd.d/eBeZFe+d0eCGd1d2eBeZGe+d3eCGd4d5eBeZHe+d6eCGd7d8eBeZIe+d9eCGd:d;eBe ZJe+deBeZKgd?ZLy)@zTF 2.0 RemBERT model.) annotationsN)DictOptionalTupleUnion)get_tf_activation)+TFBaseModelOutputWithPastAndCrossAttentions.TFBaseModelOutputWithPoolingAndCrossAttentions#TFCausalLMOutputWithCrossAttentionsTFMaskedLMOutputTFMultipleChoiceModelOutputTFQuestionAnsweringModelOutputTFSequenceClassifierOutputTFTokenClassifierOutput) TFCausalLanguageModelingLossTFMaskedLanguageModelingLossTFModelInputTypeTFMultipleChoiceLossTFPreTrainedModelTFQuestionAnsweringLossTFSequenceClassificationLossTFTokenClassificationLossget_initializerkeraskeras_serializable unpack_inputs)check_embeddings_within_bounds shape_liststable_softmax)add_code_sample_docstringsadd_start_docstrings%add_start_docstrings_to_model_forwardlogging) RemBertConfigr&cXeZdZdZdfd ZddZ d ddZxZS) TFRemBertEmbeddingszGConstruct the embeddings from word, position and token_type embeddings.c Vt|di|||_|j|_|j|_|j |_t jj|jd|_ t jj|j|_ y)N LayerNormepsilonnamerate)super__init__configinput_embedding_sizemax_position_embeddingsinitializer_rangerlayersLayerNormalizationlayer_norm_epsr*Dropouthidden_dropout_probdropoutselfr3kwargs __class__s m/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/transformers/models/rembert/modeling_tf_rembert.pyr2zTFRemBertEmbeddings.__init__Ds "6" $*$?$?!'-'E'E$!'!9!988AVAV]h8i||++1K1K+L ctjd5|jd|jj|j gt |j|_dddtjd5|jd|jj|j gt |j|_ dddtjd5|jd|j|j gt |j|_ ddd|jryd|_ t|ddetj|jj 5|jj#dd|jj gdddyy#1swY[xYw#1swYxYw#1swYxYw#1swYyxYw) Nword_embeddingsweightr-shape initializertoken_type_embeddings embeddingsposition_embeddingsTr*)tf name_scope add_weightr3 vocab_sizer4rr6rEtype_vocab_sizerIr5rKbuiltgetattrr*r-buildr> input_shapes rArSzTFRemBertEmbeddings.buildNs ]], - //{{--t/H/HI+D,B,BC*DK ]]2 3 )-!{{22D4M4MN+D,B,BC*9*D & ]]0 1 '+!33T5N5NO+D,B,BC(7(D $  ::  4d + 7t~~223 U$$dD$++2R2R%ST U U 81       U Us2A F<A G *AG?3G!<G GG!G*c,||J|At||jjtj|j |}t |dd}|tj|d}|2tjtj||d|zd}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)paramsindicesrdimsvaluer%)startlimitaxisinputsrbtraining)rr3rOrLgatherrErfill expand_dimsrangerKrIr*r<) r> input_ids position_idstoken_type_ids inputs_embedspast_key_values_lengthrdrUposition_embedstoken_type_embedsfinal_embeddingss rAcallzTFRemBertEmbeddings.callks%-*?@@  *9dkk6L6L MIIT[[)LM /4  !WW+Q?N  >>5[^Nd=delmL))4+C+C\ZIIT-G-GQ_`(?:=NN>>1A>B<*.,0.2-1 % &% *% , % + % %  % rBr(c^eZdZdfd ZddZ d ddZd dZxZS) TFRemBertSelfAttentionc t|d i||j|jzdk7r&t d|jd|jd|j|_t |j|jz |_|j|j z|_tj|j |_ tjj|jt|jd|_tjj|jt|jd|_tjj|jt|jd|_tjj'|j( |_|j,|_||_y) NrzThe hidden size (z6) is not a multiple of the number of attention heads ()queryunitskernel_initializerr-keyr\r.r0)r1r2 hidden_sizenum_attention_heads ValueErrorintattention_head_size all_head_sizemathsqrtsqrt_att_head_sizerr7Denserr6rrr\r:attention_probs_dropout_probr< is_decoderr3r=s rAr2zTFRemBertSelfAttention.__init__s "6"    : : :a ?#F$6$6#78''-'A'A&B!E  $*#=#= #&v'9'9Fr batch_sizes rAtranspose_for_scoresz+TFRemBertSelfAttention.transpose_for_scoress;6*b$BZBZ\`\t\t1uv||F66rBc t|d} |j|} |du} | r||d} |d} |}n | rG|j|j|| } |j|j || } |}n|}|j|j|| } |j|j || } t j |d| gd} t j |d| gd} nD|j|j|| } |j|j || } |j| | }|jr| | f}t j|| d}t j|j|j}t j||}|t j||}t|d }|j|| }|t j ||}t j|| }t j"|gd  }t j$|| d |j&f}|r||fn|f}|jr||fz}|S)Nrrar%rr_T) transpose_bdtyperY)logitsr`rcrrr)rrrrr\rLconcatrmatmulcastrrdivideaddr r<multiplyrrr)r> hidden_statesattention_mask head_maskencoder_hidden_statesencoder_attention_maskpast_key_valueoutput_attentionsrdrmixed_query_layeris_cross_attention key_layer value_layer query_layerattention_scoresdkattention_probsattention_outputoutputss rArqzTFRemBertSelfAttention.callsy .q1  JJmJ< 3$> ."<&q)I(+K3N 11$((BW(2XZdeI33DJJF[J4\^hiK3N  '11$((-(2PR\]I33DJJmJ4TV`aK >!#4i"@qII))^A%6 $D1MK11$((-(2PR\]I33DJJmJ4TV`aK//0A:N ??(5N99[)N WWT,,4D4J4J K99%5r:  %!vv&6G )0@rJ,,o,Q   kk/9EO99_kB<<(8|L::-=jRTVZVhVhEij9J#_5QaPc ?? 11GrBc|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|dddtj|jj 5|jj dd|jjgdddt|ddetj|jj 5|jj dd|jjgdddyy#1swYxYw#1swYxYw#1swYyxYw)NTrrr\) rQrRrLrMrr-rSr3rrr\rTs rArSzTFRemBertSelfAttention.build s9 ::  4$ ' 3tzz/ H   $dkk.E.E!FG H 4 % 1txx}}- FdDKK,C,CDE F 4$ ' 3tzz/ H   $dkk.E.E!FG H H 4  H H F F H Hs$ 3E*<3E6-3F*E36E?F rr)rrwrrrvrwF)rrwrrwrrwrrwrrwrTuple[tf.Tensor]rrurdrurvrrs)rxryrzr2rrqrSr|r}s@rArrs87 O O"O O ) O !* O)O OO Ob HrBrc2eZdZdfd ZdddZddZxZS)TFRemBertSelfOutputc xt|di|tjj |j t |jd|_tjj|jd|_ tjj|j|_||_yNdenserr*r+r.r0r1r2rr7rrrr6rr8r9r*r:r;r<r3r=s rAr2zTFRemBertSelfOutput.__init__ "6"\\''$$IaIa9bip(  88AVAV]h8i||++1K1K+L  rBcz|j|}|j||}|j||z}|SNrarcrr<r*r>r input_tensorrds rArqzTFRemBertSelfOutput.call$? - 8  MH M ml.JK rBc"|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYwNTrr*) rQrRrLrMrr-rSr3rr*rTs rArSzTFRemBertSelfOutput.build+s ::  4$ ' 3tzz/ H   $dkk.E.E!FG H 4d + 7t~~223 L$$dD$++2I2I%JK L L 8 H H L L 3C9<3D9DDrrrrrwrrwrdrurvrwrsrxryrzr2rqrSr|r}s@rArr LrBrc\eZdZdfd ZdZ d ddZddZxZS) TFRemBertAttentionc lt|di|t|d|_t |d|_y)Nr>r-outputr0)r1r2rself_attentionr dense_outputr=s rAr2zTFRemBertAttention.__init__9s1 "6"4V&I/XFrBctrsNotImplementedError)r>headss rA prune_headszTFRemBertAttention.prune_heads?s!!rBc x|j||||||||} |j| d||} | f| ddz} | S)Nrrrrrrrrdrrrrdr%)rr) r>rrrrrrrrd self_outputsrrs rArqzTFRemBertAttention.callBso**&)"7#9)/+   ,,&q/ x- $% QR(88rBc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYw)NTrr)rQrRrLrMrr-rSrrTs rArSzTFRemBertAttention.build_s ::  4)4 0 <t22778 0##))$/ 0 4 . :t00556 .!!''- . . ; 0 0 . . C %C CC rrr)rrwrrwrrwrrwrrwrrrrurdrurvrrs)rxryrzr2rrqrSr|r}s@rArr8suG ""  )  !* )  : .rBrc0eZdZdfd ZddZddZxZS)TFRemBertIntermediatec Tt|di|tjj |j t |jd|_t|jtr"t|j|_ ||_y|j|_ ||_y)Nrrr0)r1r2rr7rintermediate_sizerr6r isinstance hidden_actstrr intermediate_act_fnr3r=s rAr2zTFRemBertIntermediate.__init__ms "6"\\''**vOgOg?hov(   f'' -'89J9J'KD $ (.'8'8D $ rBcL|j|}|j|}|S)Nra)rr)r>rs rArqzTFRemBertIntermediate.callzs( - 8 00? rBc(|jryd|_t|ddetj|jj 5|jj dd|jjgdddyy#1swYyxYwNTr rQrRrLrMrr-rSr3rrTs rArSzTFRemBertIntermediate.build} ::  4$ ' 3tzz/ H   $dkk.E.E!FG H H 4 H H 3BBrrrrwrvrwrsrr}s@rArrls  HrBrc2eZdZdfd ZdddZddZxZS)TFRemBertOutputc xt|di|tjj |j t |jd|_tjj|jd|_ tjj|j|_||_yrrr=s rAr2zTFRemBertOutput.__init__rrBcz|j|}|j||}|j||z}|Srrrs rArqzTFRemBertOutput.callrrBc"|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYwr) rQrRrLrMrr-rSr3rr*rrTs rArSzTFRemBertOutput.builds ::  4$ ' 3tzz/ N   $dkk.K.K!LM N 4d + 7t~~223 L$$dD$++2I2I%JK L L 8 N N L Lrrrrrrsrr}s@rArrrrBrcVeZdZdfd Z d ddZddZxZS)TFRemBertLayerc Dt|di|t|d|_|j|_|j |_|j r,|jst |dt|d|_t|d|_ t|d|_ y)N attentionrz> should be used as a decoder model if cross attention is addedcrossattention intermediaterr0) r1r2rrradd_cross_attentionrrrrr bert_outputr=s rAr2zTFRemBertLayer.__init__s "6"+FE ++#)#=#=  # #?? D6)g!hii"4VBR"SD 1&~N*6ArBc ||ddnd} |j|||dd| ||} | d} |jr | dd} | d} n| dd} d}|jrV|Tt|dstd|d||d dnd}|j | |||||||}|d} | |ddz} |d} |z} |j | }|j || | }|f| z} |jr|  fz} | S) Nr)rrrrrrrrdrr%rYrz'If `encoder_hidden_states` are passed, z` has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`rr)rrhasattrrrrr)r>rrrrrrrrdself_attn_past_key_valueself_attention_outputsrrpresent_key_valuecross_attn_present_key_valuecross_attn_past_key_valuecross_attention_outputsintermediate_output layer_outputs rArqzTFRemBertLayer.calls:H9S>"1#5Y] !%&)"&#'3/"0 " 2!4 ??,Qr2G 6r : ,QR0G'+$ ??4@4!12 =dVDDD @N?Yrs(;_c %&*&9&9--#&;'=8"3!': ' # 7q9  7" ==G,C2+F ( 14P P "//>N/O''-r3r?ir@s rAr2zTFRemBertEncoder.__init__s "6" +0<<+=+=$$.v/G/GH.,>, ( TYY_YqYqSrsanV,2E2Ea2HIs ss3#Bc |j|}| rdnd} |rdnd} |r|jjrdnd}|rdnd}t|jD]h\}}| r| |fz} |||nd}|||||||||| }|d}|r ||dfz }|s=| |dfz} |jjs]|`||dfz}j| r| |fz} | st d|| | |fDSt ||| | | S) Nrar0rrrYr%rc3&K|] }|| ywrsr0).0vs rA z(TFRemBertEncoder.call..Osghgts)last_hidden_statepast_key_valuesr attentionscross_attentions)rr3r enumeratertupler )r>rrrrrr use_cacheroutput_hidden_states return_dictrdall_hidden_statesall_attentionsall_cross_attentionsnext_decoder_cacher layer_moduler layer_outputss rArqzTFRemBertEncoder.callsi88 8N "6BD0d%64;;;Z;Zr`d#,R$(4 VOA|#$58H$H!3B3N_Q/TXN(+-#A,&;'=-"3! M*!,M"}R'8&::" !/=3C2E!E;;227L7X+?=QRCSBU+U(1 V6 1]4D D )+rUrs rArSzTFRemBertEncoder.build[s ::  46 = It??DDE g0066dDKKDdDd7ef g 4$ ' 3 &]]5::.&KK%&& & 4 g g&&s 3CC+C(+C4 rrr)rrwrrwrrwrrwrrwrzTuple[Tuple[tf.Tensor]]rrurrur rur!rurdrurvzDUnion[TFBaseModelOutputWithPastAndCrossAttentions, Tuple[tf.Tensor]]rsrr}s@rAr r s t.=  = "=  = ) = !* = 1= =  = #= = =  N= ~ &rBr c0eZdZdfd ZddZddZxZS)TFRemBertPoolerc t|di|tjj |j t |jdd|_||_ y)Ntanhr)rr activationr-r0) r1r2rr7rrrr6rr3r=s rAr2zTFRemBertPooler.__init__jsT "6"\\''$$.v/G/GH (   rBc<|dddf}|j|}|S)Nrra)r)r>rfirst_token_tensor pooled_outputs rArqzTFRemBertPooler.callus*+1a40 *< = rBc(|jryd|_t|ddetj|jj 5|jj dd|jjgdddyy#1swYyxYwrrrTs rArSzTFRemBertPooler.build}rrrrrrsrr}s@rAr*r*is HrBr*cNeZdZdfd Zd dZd dZdZd dZd dZd dZ xZ S)TFRemBertLMPredictionHeadc t|di|||_|j|_|j|_t j j|jt|jd|_ t|jtrt|j|_n|j|_t j j|j d|_y)Nr)rr-r*r+r0)r1r2r3r6output_embedding_sizerr7rrrrrrr r-r8r9r*r>r3input_embeddingsr?r@s rAr2z"TFRemBertLMPredictionHead.__init__s "6" !'!9!9%+%A%A"\\''  ( (_TMcMc=dkr(   f'' -/0A0ABDO$//DO88AVAV]h8irBc|jd|jj|jgt |j |_|j|jjfddd|_|jryd|_t|dddtj|jj5|jjdd|jjgdddt|dddtj|j j5|j jd|jjgdddyy#1swY{xYw#1swYyxYw) Nzdecoder/weightrFzerosTz decoder/bias)rGrH trainabler-rr*)rNr3rOr5rr6decoder decoder_biasrQrRrLrMrr-rSrr*rTs rArSzTFRemBertLMPredictionHead.buildsK!;;))4+E+EF'(>(>?'  !OO;;))+DWe,  ::  4$ ' 3tzz/ H   $dkk.E.E!FG H 4d + 7t~~223 P$$dDKK,M,M%NO P P 8 H H P Ps3E372E?3E<?Fc|Srsr0r>s rAget_output_embeddingsz/TFRemBertLMPredictionHead.get_output_embeddingss rBcL||_t|d|j_yNr)r;rrOr>r\s rAset_output_embeddingsz/TFRemBertLMPredictionHead.set_output_embeddingss ",U"3A"6 rBcd|jiS)Nr<)r<r>s rAget_biasz"TFRemBertLMPredictionHead.get_biass 1 122rBcX|d|_t|dd|j_y)Nr<r)r<rr3rOrBs rAset_biasz"TFRemBertLMPredictionHead.set_biass*!.1!+E.,A!B1!E rBc|j|}|j|}t|d}tj|d|j g}|j |}tj||jd}tj|d||jjg}tjj||j}|S) Nra)rr%rYrT)abr)r\bias)rr-rrLrr5r*rr;r3rOnnbias_addr<)r>r seq_lengths rArqzTFRemBertLMPredictionHead.calls - 8  6 }5a8  -DD^D^?_` }5  MT\\tT  -JPTP[P[PfPf?gh ]ARARS rBr3r&r7keras.layers.LayerrsrvrP)rvzDict[str, tf.Variable]r\z tf.Variabler) rxryrzr2rSr?rCrErGrqr|r}s@rAr3r3s) jP(73F rBr3c0eZdZdfd ZddZddZxZS)TFRemBertMLMHeadc Jt|di|t||d|_y)N predictionsrr0)r1r2r3rVr6s rAr2zTFRemBertMLMHead.__init__s& "6"4V=MTabrBc*|j|}|S)Nr)rV)r>sequence_outputprediction_scoress rArqzTFRemBertMLMHead.calls ,,?,K  rBc|jryd|_t|ddNtj|jj 5|jj ddddyy#1swYyxYw)NTrV)rQrRrLrMrVr-rSrTs rArSzTFRemBertMLMHead.buildsm ::  4 - 9t//445 -  &&t, - - : - - A11A:rO)rXrwrvrwrsrr}s@rArTrTsc ! -rBrTceZdZeZddfd Zd dZd dZdZe d d dZ d dZ xZ S)TFRemBertMainLayerc t|di|||_|j|_t |d|_t |d|_|rt|d|_ yd|_ y)NrJrencoderpoolerr0) r1r2r3rr(rJr r_r*r`)r>r3add_pooling_layerr?r@s rAr2zTFRemBertMainLayer.__init__sZ "6"  ++-f<H'Y? @Qof8< W[ rBc|jSrs)rJr>s rAget_input_embeddingsz'TFRemBertMainLayer.get_input_embeddingss rBc`||j_t|d|j_yrA)rJrErrOrBs rAset_input_embeddingsz'TFRemBertMainLayer.set_input_embeddingss$!&%/%6q%9"rBct)z Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base class PreTrainedModel r)r>heads_to_prunes rA _prune_headszTFRemBertMainLayer._prune_headss "!rBc |jjsd} | | td| t|}n|t|dd}n td|\}}| &d}dgt |j j z} nt| ddd}|tj|||zfd}|tj|d}|j|||||| }t|}||z}|jrtj|}tjtj|ddddf||df|ddddf}tj||j }||dddddfz}t|}tj||dd|d|d f}| d3|dddd| dddf}n tj||ddd|df}tj||j }tj d |j }tj d |j }tj"tj$|||}|jrf|dtj||j }t t|}|dk(r|dddddddf}|d k(r |ddddddf}d z d z}nd}|t&dg|jj(z}|j |||||| | | | | | }|d}|j*|j+|nd}| s ||f|ddzSt-|||j.|j0|j2|j4S)NFzDYou cannot specify both input_ids and inputs_embeds at the same timerYz5You have to specify either input_ids or inputs_embedsrrr%rZ)rirjrkrlrmrdrrg?gr) rrrrrrrrr r!rdr)r pooler_outputrrrr)r3rrrlenr_rrLrfrJrh less_equaltilerrrconstantrsubtractrrr`r rrrr) r>rirrkrjrrlrrrrrr r!rdrUrrNrmembedding_outputattention_mask_shapemask_seq_lengthseq_ids causal_maskextended_attention_maskone_cstten_thousand_cstnum_dims_encoder_attention_maskencoder_extended_attention_maskencoder_outputsrXr0s rArqzTFRemBertMainLayer.calls&{{%%I  ]%>cd d  "$Y/K  &$]3CR8KTU U!, J  "%& "#fs4<<+=+='>>O%/0B10E%Fr%J "  !WW:zDZ7Z*[cdeN  !WW+Q?N??%)'#9 +  *.9$'== ??hh/G--dA ._a0PQa &K''+^5I5IJK&1N1dA:4N&N ##-.E#F &(jj'*>q*A1FZ[\F]_stu_v)w' #q!-*A!Q VWBW*X'&(jj!5a!8!Q@TUV@W X' ##%''*AIYI_I_"`++c)9)?)?@;;x7G7M7MN"$++bkk'CZ.[]m"n ??5A&(WW-CKbKhKh%i ".1*=S2T.U +.!32HDRSUV2W/.!32HDRVXYIY2Z/035T/TX`.` +.2 +  % %!>!>>I,,*2"7#B+/!5#'  *!,FJkkF] / Bcg  #$ $ >-'+;;)77&11,==   rBc|jryd|_t|ddMtj|jj 5|jj ddddt|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYxYw#1swYqxYw#1swYyxYw)NTrJr_r`) rQrRrLrMrJr-rSr_r`rTs rArSzTFRemBertMainLayer.builds ::  4t , 8t334 ,%%d+ , 4D ) 5t||001 ) ""4( ) 44 ( 4t{{//0 ( !!$' ( ( 5  , , ) ) ( (s$ D%%D1?D=%D.1D:=E)T)r3r&rarurQrRNNNNNNNNNNNNNF)riTFModelInputType | Nonernp.ndarray | tf.Tensor | Nonerkr~rjr~rr~rlr~rr~rr~r4Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]]rOptional[bool]rrr rr!rrdrurvGUnion[TFBaseModelOutputWithPoolingAndCrossAttentions, Tuple[tf.Tensor]]rs) rxryrzr& config_classr2rcrerhrrqrSr|r}s@rAr]r]s  L\:".28<8<6:377;?C@DPT$(,0/3&*[ *[ 6[ 6 [ 4 [ 1 [ 5[  =[ !>[ N[ "[ *[ -[ $[ [ Q![ [ z (rBr]ceZdZdZeZdZy)TFRemBertPreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. rembertN)rxryrzr{r&rbase_model_prefixr0rBrArrs !L!rBraw 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 ([`RemBertConfig`]): 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 (`np.ndarray`, `tf.Tensor`, `List[tf.Tensor]` ``Dict[str, tf.Tensor]` or `Dict[str, np.ndarray]` and each example must have the 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 (`np.ndarray` 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 (`np.ndarray` 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 (`np.ndarray` 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 (`np.ndarray` 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 (`np.ndarray` or `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). z`The bare RemBERT Model transformer outputing raw hidden-states without any specific head on top.ceZdZdfd Zeeejdede e  d d dZ d dZ xZ S) TFRemBertModelcPt||g|i|t|d|_y)Nrr)r1r2r]rr>r3rbr?r@s rAr2zTFRemBertModel.__init__s( 3&3F3)&yA rBbatch_size, sequence_lengthgoogle/rembert checkpoint output_typercD|j||||||||| | | | | |}|S)a encoder_hidden_states (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if the model is configured as a decoder. encoder_attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. past_key_values (`Tuple[Tuple[tf.Tensor]]` of length `config.n_layers`) contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding. If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`. use_cache (`bool`, *optional*, defaults to `True`): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). Set to `False` during training, `True` during generation rirrkrjrrlrrrrrr r!rd)r)r>rirrkrjrrlrrrrrr r!rdrs rArqzTFRemBertModel.callsIX,,))%'"7#9+/!5# "rBc|jryd|_t|ddNtj|jj 5|jj ddddyy#1swYyxYw)NTr)rQrRrLrMrr-rSrTs rArSzTFRemBertModel.buildXsi ::  4D ) 5t||001 ) ""4( ) ) 6 ) )r[rrr|)rir}rr~rkr~rjr~rr~rlr~rr~rr~rrrrrrr rr!rrdrrvrrs)rxryrzr2rr#REMBERT_INPUTS_DOCSTRINGrr!r _CONFIG_FOR_DOCrqrSr|r}s@rArrs B *+C+J+JKh+ij#B$.28<8<6:377;?C@DPT$(,0/3&*#(6*6666 6 4 6 1 656 =6!>6N6"6*6-6$6!6 Q!6 k6p)rBrz5RemBERT Model with a `language modeling` head on top.ceZdZdfd Zd dZeeejde de e  d d dZ d dZ xZS) TFRemBertForMaskedLMct||g|i||jrtj dt |dd|_t||j jd|_ y)NzpIf you want to use `TFRemBertForMaskedLM` make sure `config.is_decoder=False` for bi-directional self-attention.rFr-ra mlm___clsr7r- r1r2rloggerwarningr]rrTrJmlmrs rAr2zTFRemBertForMaskedLM.__init__csb 3&3F3    NN1  *&yTYZ #FT\\=T=T[fgrBc.|jjSrsrrVr>s rA get_lm_headz TFRemBertForMaskedLM.get_lm_headoxx###rBrrrc |j||||||||| |  } | d} |j| | }| dn|j| |}| s|f| ddz}||f|zS|St||| j| j S)a labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]` rirrkrjrrlrr r!rdrrXrdNlabelsrrlossrrr)rrhf_compute_lossr rr)r>rirrkrjrrlrr r!rrdrrXrYrrs rArqzTFRemBertForMaskedLM.callrs6,,))%'/!5#  "!* HH_xHX~t4+?+?vVg+?+h')GABK7F)-)9TGf$ Ev E$!//))   rBc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYwNTrrrQrRrLrMrr-rSrrTs rArSzTFRemBertForMaskedLM.build ::  4D ) 5t||001 ) ""4( ) 4 % 1txx}}- %t$ % % 2 ) ) % %rrrrQ NNNNNNNNNNF)rir}rr~rkr~rjr~rr~rlr~rrr rr!rrr~rdrrvz)Union[TFMaskedLMOutput, Tuple[tf.Tensor]]rs)rxryrzr2rrr#rrr!r rrqrSr|r}s@rArras h$*+C+J+JKh+ij#$$.28<8<6:377;,0/3&*04#(- *- 6- 6 - 4 - 1 - 5- *- -- $- .- !-  3- k- ^ %rBrzIRemBERT Model with a `language modeling` head on top for CLM fine-tuning.ceZdZdfd Zd dZd dZeedee  d d dZ d dZ xZ S)TFRemBertForCausalLMct||g|i||jstj dt |dd|_t||j jd|_ y)NzQIf you want to use `TFRemBertForCausalLM` as a standalone, add `is_decoder=True.`rFrrrrrs rAr2zTFRemBertForCausalLM.__init__s\ 3&3F3  NNn o)&yTYZ #FT\\=T=T[fgrBc.|jjSrsrr>s rArz TFRemBertForCausalLM.get_lm_headrrBc n|j}|tj|}| |ddddf}|||dS)NrY)rirr)rGrLones)r>rirr model_kwargsrUs rAprepare_inputs_for_generationz2TFRemBertForCausalLM.prepare_inputs_for_generationsEoo  !WW[1N  &!!RS&)I&.]lmmrBrrcf|j||||||||| | | | | |}|d}|j||}d}|)|ddddf}|ddddf}|j||}| s|f|ddz}||f|zS|St|||j|j |j |j S) a encoder_hidden_states (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if the model is configured as a decoder. encoder_attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. past_key_values (`Tuple[Tuple[tf.Tensor]]` of length `config.n_layers`) contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding. If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`. use_cache (`bool`, *optional*, defaults to `True`): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). Set to `False` during training, `True` during generation labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the cross entropy classification loss. Indices should be in `[0, ..., config.vocab_size - 1]`. rrrNrYr%rr)rrrrrr)rrrr rrrr)r>rirrkrjrrlrrrrrr r!rrdrrXrrshifted_logitsrs rArqzTFRemBertForCausalLM.calls^,,))%'"7#9+/!5#  "!*/HM  #AssF^NAqrE]F''vn'MDY,F)-)9TGf$ Ev E2#33!//))$55   rBc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYwrrrTs rArSzTFRemBertForCausalLM.build'rrrrrQ)NN)NNNNNNNNNNNNNNF) rir}rr~rkr~rjr~rr~rlr~rr~rr~rrrrrrr rr!rrr~rdrrvzN NN "N *N -N $N .N !!N " F#N  N ` %rBrzo RemBERT Model transformer with a sequence classification/regression head on top e.g., for GLUE tasks. ceZdZdfd Zeeejdede e  d d dZ d dZ xZ S) "TFRemBertForSequenceClassificationcft||g|i||j|_t|d|_t j j|j|_ t j j|jt|jd|_ ||_y)Nrrr. classifierr)r1r2 num_labelsr]rrr7r:classifier_dropout_probr<rrr6rr3rs rAr2z+TFRemBertForSequenceClassification.__init__:s 3&3F3 ++)&yA ||++1O1O+P ,,,,##.v/G/GH-   rBrrrc *|j||||||||| |  } | d} |j| | } |j| }| dn|j| |}| s|f| ddz}||f|zS|St ||| j | j S) a labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). rr%rcraNrrr)rr<rrrrr)r>rirrkrjrrlrr r!rrdrr0rrrs rArqz'TFRemBertForSequenceClassification.callHs6,,))%'/!5#     MH M  6~t4+?+?vV\+?+]Y,F)-)9TGf$ Ev E)!//))   rBc|jryd|_t|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYwNTrr rQrRrLrMrr-rSrr3rrTs rArSz(TFRemBertForSequenceClassification.build ::  4D ) 5t||001 ) ""4( ) 4t , 8t334 M%%tT4;;3J3J&KL M M 9 ) ) M M C"%3C."C+.C7rrr)rir}rr~rkr~rjr~rr~rlr~rrr rr!rrr~rdrrvz3Union[TFSequenceClassifierOutput, Tuple[tf.Tensor]]rs)rxryrzr2rr#rrr!rrrqrSr|r}s@rArr3s *+C+J+JKh+ij#.$.28<8<6:377;,0/3&*04#(. *. 6. 6 . 4 . 1 . 5. *. -. $. .. !.  =. k. ` MrBrz RemBERT 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. ceZdZdfd Zeeejdede e  d d dZ d dZ xZ S) TFRemBertForMultipleChoicec0t||g|i|t|d|_tj j |j|_tj jdt|jd|_ ||_ y)Nrrr.r%rr)r1r2r]rrr7r:rr<rrr6rr3rs rAr2z#TFRemBertForMultipleChoice.__init__s~ 3&3F3)&yA ||++1O1O+P ,,,,8P8P(QXd-  rBz(batch_size, num_choices, sequence_lengthrrc |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|j S) aE labels (`tf.Tensor` or `np.ndarray` 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%rrYrrrrcrarr) rrLrrr<rrrrr)r>rirrkrjrrlrr r!rrd num_choicesrNflat_input_idsflat_attention_maskflat_token_type_idsflat_position_idsflat_inputs_embedsrr0rreshaped_logitsrrs rArqzTFRemBertForMultipleChoice.calls6  $Y/2K#I.q1J$]3A6K#M215JQZQf9RE xE N MrBrz RemBERT 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. ceZdZdfd Zeeejdede e  d d dZ d dZ xZ S) TFRemBertForTokenClassificationcht||g|i||j|_t|dd|_t j j|j|_ t j j|jt|jd|_ ||_y)NrFrr.rr)r1r2rr]rrr7r:r;r<rrr6rr3rs rAr2z(TFRemBertForTokenClassification.__init__s 3&3F3 ++)&yTYZ ||++1K1K+L ,,,,##H`H`8aht-  rBrrrc *|j||||||||| |  } | d} |j| | } |j| }| dn|j| |}| s|f| ddz}||f|zS|St ||| j | j S) z 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]`. rrrcraNrr%r)rr<rrrrr)r>rirrkrjrrlrr r!rrdrrXrrrs rArqz$TFRemBertForTokenClassification.call s2,,))%'/!5#  "!*,,o,Q8~t4+?+?vV\+?+]Y,F)-)9TGf$ Ev E&!//))   rBc|jryd|_t|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYwrrrTs rArSz%TFRemBertForTokenClassification.build@rrrrr)rir}rr~rkr~rjr~rr~rlr~rrr rr!rrr~rdrrvz0Union[TFTokenClassifierOutput, Tuple[tf.Tensor]]rs)rxryrzr2rr#rrr!rrrqrSr|r}s@rArrs *+C+J+JKh+ij#+$.28<8<6:377;,0/3&*04#(, *, 6, 6 , 4 , 1 , 5, *, -, $, ., !,  :, k, \ MrBrz RemBERT Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`). ceZdZdfd Zeeejdede e  d d dZ d dZ xZ S) TFRemBertForQuestionAnsweringc t||g|i||j|_t|dd|_t j j|jt|jd|_ ||_ y)NFr)rar- qa_outputsr) r1r2rr]rrr7rrr6rr3rs rAr2z&TFRemBertForQuestionAnswering.__init__Tsu 3&3F3 ++)&EPYZ ,,,,##H`H`8aht-  rBrrrc |j||||||||| |  } | d}|j|}tj|dd\}}tj|d}tj|d}d}| | d | i}| |d <|j |||f }| s||f| ddz}||f|zS|St |||| j| j S) a start_positions (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*): Labels for position (index) of the start of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. end_positions (`tf.Tensor` or `np.ndarray` of shape `(batch_size,)`, *optional*): Labels for position (index) of the end of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. rrrarrY)r\num_or_size_splitsr`)inputr`Nstart_position end_positionr)r start_logits end_logitsrr) rrrLsplitsqueezerrrr)r>rirrkrjrrlrr r!start_positions end_positionsrdrrXrrrrrrs rArqz"TFRemBertForQuestionAnswering.call_s@,,))%'/!5#  "!*8#%88&QUW#X jzz 2> ZZjr:   &=+D&8F%2F> "''v|Z>X'YD"J/'!"+=F)-)9TGf$ Ev E-%!!//))   rBc|jryd|_t|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYw)NTrr) rQrRrLrMrr-rSrr3rrTs rArSz#TFRemBertForQuestionAnswering.buildrrrr) NNNNNNNNNNNF)rir}rr~rkr~rjr~rr~rlr~rrr rr!rrr~rr~rdrrvz7Union[TFQuestionAnsweringModelOutput, Tuple[tf.Tensor]]rs)rxryrzr2rr#rrr!rrrqrSr|r}s@rArrLs *+C+J+JKh+ij#2$.28<8<6:377;,0/3&*9=7;#(; *; 6; 6 ; 4 ; 1 ; 5; *; -; $; 7; 5; !;  A; k; z MrBr) rrrrrrrrr)Mr{ __future__rrtypingrrrrnumpynp tensorflowrLactivations_tfr modeling_tf_outputsr r r r rrrrmodeling_tf_utilsrrrrrrrrrrrrtf_utilsrrr utilsr!r"r#r$configuration_rembertr& get_loggerrxrrr7Layerr(rrrrrrr r*r3rTr]rREMBERT_START_DOCSTRINGrrrrrrrr__all__r0rBrArs&" ///       SR 1   H %!O %,,,,O fAHU\\//AHJL%,,,,L>0.++0.hHELL..H<Lell((L>d0U\\''d0NU&u||))U&rHell((H:; 2 2;~-u||))-(F(++F(F(R"0"(T5pfK)-K) K)\QSjkO%35QO%lO%dSUly%35Qy%y%x  OM)AC_OM  OMd bM!9;ObMbMJ KM&>@YKMKM\ YM$<>UYMYMx rB