sg$9dZddlmZddlZddlmZddlmZmZm Z ddl Z ddl Z ddlmZddlmZmZmZmZmZmZdd lmZmZmZmZmZmZmZmZm Z m!Z!m"Z"dd l#m$Z$m%Z%m&Z&dd l'm(Z(m)Z)m*Z*m+Z+m,Z,m-Z-d d l.m/Z/e,j`e1Z2dZ3dZ4Gdde jjjlZ7GddZ8dZ9Gdde jjjlZ:Gdde jjjlZ;Gdde jjjlZ<Gdde jjjlZ=dHdZ>Gdde jjjlZ?e!Gd d!e jjjlZ@e!Gd"d#e jjjlZAGd$d%e jjjlZBGd&d'e jjjlZCGd(d)e jjjlZDGd*d+eZEeGd,d-e(ZFd.ZGd/ZHe*d0eGGd1d2eEZIe*d3eGGd4d5eEZJe*d6eGGd7d8eEZKe*d9eGGd:d;eEeZLe*deEeZMe*d?eGGd@dAeEeZNe*dBeGGdCdDeEeZOe*dEeGGdFdGeEeZPy)IzTF 2.0 Funnel model.) annotationsN) dataclass)OptionalTupleUnion)get_tf_activation)TFBaseModelOutputTFMaskedLMOutputTFMultipleChoiceModelOutputTFQuestionAnsweringModelOutputTFSequenceClassifierOutputTFTokenClassifierOutput) TFMaskedLanguageModelingLossTFModelInputTypeTFMultipleChoiceLossTFPreTrainedModelTFQuestionAnsweringLossTFSequenceClassificationLossTFTokenClassificationLossget_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) FunnelConfigr%g.Ac2eZdZdZfdZddZddZxZS)TFFunnelEmbeddingszGConstruct the embeddings from word, position and token_type embeddings.c Pt|di|||_|j|_|jdn |j|_t j j|jd|_ t j j|j|_ y)N? layer_normepsilonname)rate) super__init__config hidden_sizeinitializer_stdrlayersLayerNormalizationlayer_norm_eps LayerNormDropouthidden_dropoutdropoutselfr2kwargs __class__s `/var/www/html/venv/lib/python3.12/site-packages/transformers/models/funnel/modeling_tf_funnel.pyr1zTFFunnelEmbeddings.__init__Gs "6" !--&,&<&<&Ds&J`J`88AVAV]i8j||++1F1F+G ctjd5|jd|jj|j gt |j|_ddd|jryd|_ t|ddetj|jj5|jjdd|jjgdddyy#1swYxYw#1swYyxYw)Nword_embeddingsweight)initializer_range)r-shape initializerTr8)tf name_scope add_weightr2 vocab_sizer3rr4rDbuiltgetattrr8r-buildd_modelr= input_shapes r@rNzTFFunnelEmbeddings.buildQs ]], - //{{--t/?/?@+d>R>RS*DK  ::  4d + 7t~~223 H$$dD$++2E2E%FG H H 8   H HsAC363C?3C<?Dc||J||J|@t||jjtj|j |}|j |}|j||}|S)z Applies embedding based on inputs tensor. Returns: final_embeddings (`tf.Tensor`): output embedding tensor. )inputs)rStraining)rr2rKrHgatherrDr8r;)r= input_ids inputs_embedsrTfinal_embeddingss r@callzTFFunnelEmbeddings.call`s%-*?@@)m.GHH  *9dkk6L6L MIIdkk9=M>>>?<2;;8N8N7OWeWkWk+lm--w ;R^^GUW=XY}}Wn55rAc|jdk(r%tjd|d}tjd|jdzd}dd||jdzz zz }tjd||}tj |}|j ||}tj|} |j| |} tj||gd } tj| |gd } tj| | gd } tj| | gd }| | | |fStjd|jdzd}dd||jdzz zz }tj| dz|dzd}|tjdz}tjd||}|j tj ||}|jtj||} tj|| gd }tjd|}|}g}td|jD]-}tjdgd }|dk7rm|j||}d|dz z}|j|||d }tj||j }||z}tj"||d }|}d|z}|j||}tj||j }||z}tj$j'|tj(|dtj"||d }|j+||g0|S)a Create and cache inputs related to relative position encoding. Those are very different depending on whether we are using the factorized or the relative shift attention: For the factorized attention, it returns the matrices (phi, pi, psi, omega) used in the paper, appendix A.2.2, final formula. For the relative shift attention, it returns all possible vectors R used in the paper, appendix A.2.1, final formula. Paper link: https://arxiv.org/abs/2006.03236 factorizedrr)rdr$i'zi,d->idrrraxisg)value)shiftrs)rhrHrangerOeinsumsinrncosroconcatrrifillstride_pool_pos relative_poscastrtrU debugging assert_lessrFappend)r=rurTpos_seqfreq_seqinv_freqsinusoid sin_embed sin_embed_d cos_embed cos_embed_dphipsipiomega rel_pos_id zero_offset pos_embedpos pooled_posposition_embeds_list block_indexposition_embeds_poolingstriderel_posposition_embeds_no_poolings r@rvz.TFFunnelAttentionStructure.get_position_embedssD   , .hhq'3/Gxx4<<1#4c:HEh$,,!2C&DEFHyyGX>Hx(I**9x*HKx(I**9x*HK))[+6R@C))Y 2 >#A#; rAc||}|d|dz }|t|dz}|||zz}|d|dz }tj||dz | S)zV Build the relative positional vector between `pos` and `pooled_pos`. rrr$)rrHr) r=rrrr ref_point num_removemax_distmin_dists r@rz'TFFunnelAttentionStructure.relative_posso  JqMCF* Z 3A66 zF22a=3r7*xx(Q,88rAc|ytttfrD]}j||}|St|ttfrt |fd|DSt t |zjrjr tdddn tddd}tdgz|gz}jr8tdgztddgz}tj|||g}||S)zT Perform pooling by stride slicing the tensor along the given axis. Nc3BK|]}j|ywrZ) stride_pool).0xrr=s r@ z9TFFunnelAttentionStructure.stride_pool.. sJa 0 0D 9Jsrrdr$) isinstancelisttuplertypelenrrjrkslicerHr)r=tensorrax axis_slice enc_slice cls_slices` ` r@rz&TFFunnelAttentionStructure.stride_pools > dT5M * 6))&"5 6M fudm ,4<J6JJ J Jv&''+/+<+<ARARU4Q'X]^bdhjkXl 4[MD(J<7   t ,dA/??IYYy 16:DAFi  rAcytttfrtfdDSjr;j r ddddfn}t jddddf|gdtt}|dk(r dddddfdk(r%t jjdd nad k(r%t jjdd n7d k(r't jj dd  n td |dk(rt jdSS)z3Apply 1D pooling to a tensor of size [B x T (x H)].Nc3FK|]}jyw))moderN) pool_tensor)rrrr=rrs r@rz9TFFunnelAttentionStructure.pool_tensor..3s$cWX 0 0d6 0 Rcs!rr$rrdmeanNWCSAME)strides data_formatpaddingmaxminz0The supported modes are 'mean', 'max' and 'min'.)rrrrrjrkrHrrrnn avg_pool1d max_pool1dNotImplementedErrorsqueeze)r=rrrsuffixndims```` r@rz&TFFunnelAttentionStructure.pool_tensor,sC > fudm ,4<c\bcc c   '+'8'8VAssF^fFYYq"1"u v6Q?F:f%& 19Aq$J'F 6>UU%%fffRWag%hF U]UU%%fffRWag%hF U]ee&&wTYci&jjF%&XY Y(, rzz&!$=v=rAc0|\}}}}|jrl|jdk(r|j|ddd|ddz}|j|d}|j|d}|j||j}n|xj dzc_|jdk(r|j|d}|j|ddg}|j|ddg}|j|d}|j||j}||||f}||fS)zTPool `output` and the proper parts of `attention_inputs` before the attention layer.rNrdrr$rr)rlrhrrrmrp)r=outputattention_inputsr|r}rzr~s r@pre_attention_poolingz0TFFunnelAttentionStructure.pre_attention_poolingHs5DTA   ""l2"&"2"2?2A3F"J_]^]_M`"`!--na@N''!4H%%f43D3D%EF    " ""l2"&"2"2?A"F!--nq!fEN''1a&9H!--n5-IN%%f43D3D%EF+^^XV'''rAc$|\}}}}|jrv|xjdzc_|jdk(r|dd|j|dddz}|j|d}|j|d}|j |d}||||f}|S)zFPool the proper parts of `attention_inputs` after the attention layer.rdrNrr$rr)rlrprhrr)r=rr|r}rzr~s r@post_attention_poolingz1TFFunnelAttentionStructure.post_attention_pooling\sDTA       " ""l2"1"1"58H8HYZY[I\^_8`"`!--na@N''!4H!--n5-IN+^^XVrAr[F)Nr$)rrd)r\r]r^r_rf__annotations__r1rrwrvrrrrrrr/rAr@rcrctsIs ! K6Q(f  9!6>8((  rArcct|\}}}}tj|||||g}|dddd|dddf}tj||||||z g}|dd|f}|S)N.)rrHreshape)positional_attn context_lenr batch_sizen_headru max_rel_lens r@_relative_shift_gatherrjs~/9//J,Jjj:v{T[2\]O%aEFAo6Ojj:vwP[^cPc2deO%ceZdZfdZddZddZddZddZxZS)TFFunnelRelMultiheadAttentionc t|d i||j|_|jx|_}|jx|_}|j x|_}|j |_||_tjj|j|_ tjj|j|_ t|j }tjj||zd|d|_tjj||z|d|_tjj||z|d|_tjj||d|_tjj'|j(d |_d |d zz |_y) NFq_head)use_biaskernel_initializerr-k_headrr-v_head post_projr*r+r)g?r/)r0r1rhrd_headrOrErrr5r9r:attention_dropoutrDenserrrrr6r7r*scale) r=r2rr>rrrOrGr?s r@r1z&TFFunnelRelMultiheadAttention.__init__{sp "6"$33%}}, f%}}, f!'/ w!'!9!9&#ll2263H3HI!&!5!5f6N6N!O%f&>&>? ll(( VOe RZ)  ll((&[_g(h ll((&[_g(h ++G Ze+f,,99&BWBW^j9kFCK( rAc|j|j|j}}}t|j}|j ||f|dd|_|j ||f|dd|_|j |||f|dd|_|j ||f|dd|_ |j d||f|dd|_ |jryd|_ t|d dPtj|jj 5|jj#dd|gdddt|d dPtj|j$j 5|j$j#dd|gdddt|d dPtj|j&j 5|j&j#dd|gdddt|d dStj|j(j 5|j(j#dd||zgdddt|d dQtj|j*j 5|j*j#dd|gdddyy#1swYxYw#1swY3xYw#1swYxYw#1swYxYw#1swYyxYw)NTr_w_biasrFrG trainabler-r_r_biasr_kernelr_s_biasrd seg_embedrrrrr*)rrrOrrErJrrrrrrLrMrHrIrr-rNrrrr*)r=rQrrrOrGs r@rNz#TFFunnelRelMultiheadAttention.builds"&++t{{DLL%d&<&<= 6" tR\(  6" tR\(  FF+PT[e(  6" tR\(  ff%;$U`)  ::  44 ( 4t{{//0 9 !!4w"78 9 44 ( 4t{{//0 9 !!4w"78 9 44 ( 4t{{//0 9 !!4w"78 9 4d + 7t~~223 D$$dD&6/%BC D 4t , 8t334 =%%tT7&;< = = 9 9 9 9 9 9 9 D D = =s<J2/J? K )"K K$2J<?K  KK!$K-c|jdk(r|\}}}}|j|jz} |j} t j d|| z| } | |dddfz} | |dddfz} t j d| |t j d| |z}nt |d|k7rd}||jd}nd}||jd}|j|jz}|j} t j d|| }t j d ||z|}t|||}|||z}|S) z5Relative attention score for the positional encodingsrzbinh,dnh->bindNz bind,jd->bnijr$rdrz td,dnh->tnhzbinh,tnh->bnit) rhrrrrHrrrr)r=r|rrr~rrrruw_r q_r_attentionq_r_attention_1q_r_attention_2rrrvr_heads r@relative_positional_attentionz;TFFunnelRelMultiheadAttention.relative_positional_attentions\   , .#2 CS%  *A--CII&6 CHM+c!T'l:O+bDk9O!ii#NQSQZQZ%RO &!!$ 3#D$4$45a8#D$4$45a8  *A--CYY}a5F ii(8&1*fMO4_kSXYO   x 'OrAc |yt|\}}}|j|jz}tjd||z|j }tj |dddfdt|dddg}tj|dd\} } tj|tj | ddd|gtj | ddd|g} || |z} | S)z/Relative attention score for the token_type_idsNrzbind,snd->bnisr$rdrr) rrrrHrrtilesplitwhere) r=r}rr~rrurrtoken_type_biasdiff_token_typesame_token_typetoken_type_attns r@relative_token_type_attentionz;TFFunnelRelMultiheadAttention.relative_token_type_attentions  !+5n+E( G[==4::-))$4fx6GX4!81j>PQR>SUVXY:Z[,.88OQR+P(((  GGOaA{%; < GGOaA{%; <    x 'OrAc |\}}} } t|\} } } t|d}|j|j}}tj|j || | ||g}tj|j || |||g}tj|j|| |||g}||jz}|j|jz}tjd||z|}|j|||| }|j||| }||z|z}| 8tj| |j} |td| ddddfz zz }t!|d}|j#||}tjd||}|j%tj|| | ||zg}|j'||}|j)||z}|r||fS|fS)Nr$zbind,bjnd->bnijrsrrrrzbnij,bjnd->bind)rrrrHrrrrrrrrrrrtINFrrrr:r*)r=querykeyrroutput_attentionsrTr|r}rzr~rru_rrrrrrr content_scorerr attn_score attn_probattn_vecattn_outrs r@rYz"TFFunnelRelMultiheadAttention.callsEUA!+E!2 GQ oa( dkkDKK.Wff0UVDKK,z;PV.WXDKK.[&RX0YZ$**$==4::- "3Vh5FO <<_fVackl<<^VU]^#_4F   %WW^:;K;KLN#sa.D$2O.O'PQJ#:B7 **9x*H 99. 6B>>"**X GVV\_7]"^_&&x(&C!12&7 "FfYFrArZFF) r\r]r^r1rNrrrYr`ras@r@rrzs)0%=N,\6+GrArc.eZdZfdZddZddZxZS)TFFunnelPositionwiseFFNc lt|di|t|j}tj j |j|d|_t|j|_ tj j|j|_ tj j |j|d|_tj j|j |_tj j%|j&d|_||_y)Nlinear_1rlinear_2r*r+r/)r0r1rrErr5rd_innerr(r hidden_actactivation_functionr9activation_dropoutrOr)r:r;r6r7r*r2r=r2r>rGr?s r@r1z TFFunnelPositionwiseFFN.__init__2s "6"%f&>&>?  **6>>k`j*k #4V5F5F#G "',,"6"6v7P7P"Q **6>>k`j*k ||++F,A,AB ,,99&BWBW^j9k rAc|j|}|j|}|j||}|j|}|j ||}|j ||zSNrr)r(r,r-r)r;r*)r=hiddenrThs r@rYzTFFunnelPositionwiseFFN.call=sh MM& !  $ $Q '  # #A # 9 MM!  LLXL .vz**rAc|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|dddtj|jj 5|jj dd|jjgdddt|ddetj|jj 5|jj dd|jjgdddyy#1swYxYw#1swYxYw#1swYyxYw)NTr(r)r*) rLrMrHrIr(r-rNr2rOr)r*r*rPs r@rNzTFFunnelPositionwiseFFN.buildEsA ::  4T * 6t}}112 G ##T41D1D$EF G 4T * 6t}}112 G ##T41D1D$EF G 4t , 8t334 I%%tT4;;3F3F&GH I I 9  G G G G I Is$ 3E*<3E6-3F*E36E?F rrZr\r]r^r1rYrNr`ras@r@r&r&1s + IrAr&c.eZdZfdZddZddZxZS) TFFunnelLayerc nt|di|t||d|_t |d|_y)N attentionr-ffnr/)r0r1rr8r&r:)r=r2rr>r?s r@r1zTFFunnelLayer.__init__Us2 "6"6v{Q\]*6>rAct|j||||||}|j|d|}|r||dfS|fS)NrrTrrrr$)r8r:) r=rrrrrrTattnrs r@rYzTFFunnelLayer.callZsU~~ 3/CT_g $q'H5$5Q DF9DrAc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYw)NTr8r:)rLrMrHrIr8r-rNr:rPs r@rNzTFFunnelLayer.buildas ::  4d + 7t~~223 +$$T* + 4 % 1txx}}- %t$ % % 2 + + % % C %C CC r$rZr4ras@r@r6r6Ts? E %rAr6c:eZdZfdZ ddZddZxZS)TFFunnelEncoderc rt|di||j|_|j|_|j|_t ||_t|jDcgc]0\}}t|Dcgc]}t||d|d|c}2c}}}|_ ycc}wcc}}}w)Nz blocks_._z_._r9r/) r0r1rjrl block_repeatsrcattention_structure enumerate block_sizesrr6blocks)r=r2r>r block_sizeir?s r@r1zTFFunnelEncoder.__init__ns "6""//!--#11#=f#E ,5V5G5G+H  ' Z`eeo_p qZ[]6;y SQRPS5T U q  q s-B2B-B2-B2c |jj||||}|} |r|fnd} |rdnd} t|jD] \} } t | d|j rdndkD}|xr| dkD}t jt | }|r|jj| |\}}t| D]\}}t|j| D]r}|dk(xr |dk(xr|}|r|}|jr| n|x}}n| x}x}}|||||||}|d} |r|jj|}|r| |ddz} |sm| | fz} t|std| | | fDSt| | | S) Nrzr{rTr/r$rdrr<c3&K|] }|| ywrZr/rr s r@rz'TFFunnelEncoder.call..aqSTS`alast_hidden_state hidden_states attentions)rDrrErGrrjrHzerosrrrCrlrrr )r=rWrzr{routput_hidden_states return_dictrTrr1all_hidden_statesall_attentionsrblock pooling_flag pooled_hidden layer_indexlayer repeat_index do_poolingrrr layer_outputs r@rYzTFFunnelEncoder.callys 33II )) J  0D],$0d"+DKK"8 J K%f-a09J9JAPQRL';K!OLHHZ%78M262J2J2`2`,3/ /'0&6 J" U$)$*<*<[*I$JJL".!"3!\+:J!\P\J! -040@0@fmSe.444e#(sE+;O`ks$L*!_F!+/+C+C+Z+Z[k+l(()7,qr:J)J+,= ,I)#J J J<aV->$Oaa a 6IZguvvrAc|jryd|_|jD]B}|D];}tj|j5|j dddd=Dy#1swYJxYw)NT)rLrGrHrIr-rN)r=rQrYr]s r@rNzTFFunnelEncoder.buildsk ::  [[ &E &]]5::.&KK%&& & &&&s A''A0 NNFFTFrZr4ras@r@rArAms(  "7wr&rArAc(|dk(r|S|r|ddddf}|ddddf}tj||d}|rM|r"tj|ddgd|dz gddgg}|ddd|dz f}tj|gd}|S|ddd|f}|S)z{ Upsample tensor `x` to match `target_len` by repeating the tokens `stride` time on the sequence length dimension. r$N)repeatsrrr)rHrepeatrxr)rr target_lenrjrkclsrs r@upsamplerhs{2A2h aeH YYq&q 1F VVFaVa!_q!f$EFF+Z!^++,C=q1 M;J;' MrAc:eZdZfdZ ddZddZxZS)TFFunnelDecoderc @t|di||j|_|j|_dt |j dz z|_t||_t|jDcgc]}t|dd|c}|_ ycc}w)Nrdr$rz layers_._r9r/) r0r1rjrkrrFrrcrDrnum_decoder_layersr6r5)r=r2r>rIr?s r@r1zTFFunnelDecoder.__init__s "6""//"//C 2 23a78 #=f#E OTU[UnUnOop!}VQy_Ep ps:Bc t||jt|d|j|j} | |z} |r| fnd} |rdnd} |j j | |||} |jD](}|| | | | ||}|d} |r| |ddz} |s#| | fz} *|std| | | fDSt| | | S) Nr$)rrfrjrkr/rKr<rc3&K|] }|| ywrZr/rMs r@rz'TFFunnelDecoder.call..rNrOrP) rhrrrjrkrDrr5rr )r= final_hiddenfirst_block_hiddenrzr{rrUrVrTupsampled_hiddenr1rWrXrr]r`s r@rYzTFFunnelDecoder.calls$ ;;!"45a8****  "$66)=VI40d33II )) J [[ BE (8L]hpL"!_F !/,qr2B!B#$5 $A! BaV->$Oaa a 6IZguvvrAc|jryd|_t|ddK|jD];}tj|j 5|j dddd=yy#1swYIxYw)NTr5)rLrMr5rHrIr-rN)r=rQr]s r@rNzTFFunnelDecoder.buildsp ::  44 ( 4 &]]5::.&KK%&& & 5&&s A..A7 rbrZr4ras@r@rjrjs)q"+wZ&rArjcbeZdZdZeZfdZdZdZdZ e ddZ d dZ xZ S) TFFunnelBaseLayerzBase model without decoderc t|di|||_|j|_|j|_|j |_t|d|_t|d|_ y)N embeddingsr9encoderr/) r0r1r2rrUuse_return_dictrVr'rvrArwr<s r@r1zTFFunnelBaseLayer.__init__s` "6" !'!9!9$*$?$?!!11,V,G&vI> rAc|jSrZrvr=s r@get_input_embeddingsz&TFFunnelBaseLayer.get_input_embeddings! rAc`||j_t|d|j_yNrrvrDrrKr=rs r@set_input_embeddingsz&TFFunnelBaseLayer.set_input_embeddings$$!&%/%6q%9"rActrZrr=heads_to_prunes r@ _prune_headszTFFunnelBaseLayer._prune_heads(!!rAc 2| | td| t|} n|t|dd} n td|tj| d}|tj| d}||j ||}|j |||||||} | S)NDYou cannot specify both input_ids and inputs_embeds at the same timer5You have to specify either input_ids or inputs_embedsr$rrrrzr{rrUrVrT) ValueErrorrrHrrvrw) r=rVrzr{rWrrUrVrTrQencoder_outputss r@rYzTFFunnelBaseLayer.call+s  ]%>cd d  "$Y/K  &$]3CR8KTU U  !WW[!4N  !WW[!4N   OOIOIM,, ))/!5#' rAc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYw)NTrvrw)rLrMrHrIrvr-rNrwrPs r@rNzTFFunnelBaseLayer.buildUs ::  4t , 8t334 ,%%d+ , 4D ) 5t||001 ) ""4( ) ) 6 , , ) )r?NNNNNNNFrZr\r]r^r_r% config_classr1r|rrrrYrNr`ras@r@rtrtsS$L ?:"!''R )rArtcbeZdZdZeZfdZdZdZdZ e ddZ d dZ xZ S) TFFunnelMainLayerzBase model with decoderc &t|di|||_|j|_|j|_|j |_|j |_t|d|_ t|d|_ t|d|_ y)Nrvr9rwdecoderr/)r0r1r2rFrrUrxrVr'rvrArwrjrr<s r@r1zTFFunnelMainLayer.__init__gs| "6" !--!'!9!9$*$?$?!!11,V,G&vI> &vI> rAc|jSrZrzr{s r@r|z&TFFunnelMainLayer.get_input_embeddingstr}rAc`||j_t|d|j_yrrrs r@rz&TFFunnelMainLayer.set_input_embeddingswrrActrZrrs r@rzTFFunnelMainLayer._prune_heads{rrAc | | td| t|} n|t|dd} n td|tj| d}|tj| d}||j ||}|j ||||d||} |j | d| d|jd|||||| } |s6d} | df} |r| dz } | | d| | zfz} |r| dz } | | d | | zfz} | St| d|r| j| jznd|r| j| jz Sd S) Nrrrr$rrrTr)rorprzr{rrUrVrTrdrP) rrrHrrvrwrrFr rRrS)r=rVrzr{rWrrUrVrTrQrdecoder_outputsidxoutputss r@rYzTFFunnelMainLayer.call~s  ]%>cd d  "$Y/K  &$]3CR8KTU U  !WW[!4N  !WW[!4N   OOIOIM,, ))/!%#' ,,(+.q1$2B2B12EF))/!5#'  C&q)+G#q!_Q%7/#:N%N$PP q!_Q%7/#:N%N$PPN -a0#+88?;X;XXTe22_5O5OO   lp   rAc|jryd|_t|ddMtj|jj 5|jj ddddt|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYxYw#1swYqxYw#1swYyxYw)NTrvrwr) rLrMrHrIrvr-rNrwrrPs r@rNzTFFunnelMainLayer.builds ::  4t , 8t334 ,%%d+ , 4D ) 5t||001 ) ""4( ) 4D ) 5t||001 ) ""4( ) ) 6  , , ) ) ) )s$ D%%D1?D=%D.1D:=ErrZrras@r@rrasU!L ?:"!C C J )rArc0eZdZdZfdZdZddZxZS) TFFunnelDiscriminatorPredictionszEPrediction module for the discriminator, made up of two dense layers.c @t|di|t|j}tj j |j|d|_t|j|_ tj j d|d|_ ||_ y)Ndenserr$dense_predictionr/)r0r1rrErr5rrOrr r+r,rr2r.s r@r1z)TFFunnelDiscriminatorPredictions.__init__s~ "6"%f&>&>? \\'';]d'e #4V5F5F#G % 2 21 2 n rAc|j|}|j|}tj|j |}|SrZ)rr,rHrr)r=discriminator_hidden_statesrRlogitss r@rYz%TFFunnelDiscriminatorPredictions.calls? #>? 00? D11-@A rAc"|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYw)NTrr) rLrMrHrIrr-rNr2rOrrPs r@rNz&TFFunnelDiscriminatorPredictions.builds ::  4$ ' 3tzz/ D   $dkk.A.A!BC D 4+T 2 >t4499: O%%++T49L9L,MN O O ? D D O O 3C9<3D9DDrZ)r\r]r^r_r1rYrNr`ras@r@rrsO OrArcHeZdZfdZfdZdZdZdZdZddZ xZ S) TFFunnelMaskedLMHeadc bt|di|||_|j|_||_y)Nr/)r0r1r2r3input_embeddings)r=r2rr>r?s r@r1zTFFunnelMaskedLMHead.__init__s0 "6" !-- 0rAc|j|jjfddd|_t||y)NrTTbiasr)rJr2rKrr0rN)r=rQr?s r@rNzTFFunnelMaskedLMHead.builds7OO4;;+A+A*CQXdhouOv   k"rAc|jSrZ)rr{s r@get_output_embeddingsz*TFFunnelMaskedLMHead.get_output_embeddingss$$$rAc`||j_t|d|j_yr)rrDrrKrs r@set_output_embeddingsz*TFFunnelMaskedLMHead.set_output_embeddingss(',$+5e+rArc.eZdZfdZddZddZxZS)TFFunnelClassificationHeadc ht|di|t|j}tj j |j|d|_tj j|j|_ tj j ||d|_ ||_ y)N linear_hiddenr linear_outr/)r0r1rrErr5rrOrr9r:r;rr2)r=r2n_labelsr>rGr?s r@r1z#TFFunnelClassificationHead.__init__s "6"%f&>&>? "\\//S^et/u||++F,A,AB ,,,,X+\h,i rAc|j|}tjj|}|j ||}|j |Sr0)rr activationstanhr;r)r=r1rTs r@rYzTFFunnelClassificationHead.callsI##F+""''/fx8v&&rAc"|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYw)NTrr) rLrMrHrIrr-rNr2rOrrPs r@rNz TFFunnelClassificationHead.build#s ::  4$ / ;t11667 L""(($dkk6I6I)JK L 4t , 8t334 I%%tT4;;3F3F&GH I I 9 L L I IrrrZr4ras@r@rrs' IrArc(eZdZdZeZdZedZy)TFFunnelPreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. funnelcPdtjdtjiS)NrV)r$rrsrHryint32r{s r@ dummy_inputsz$TFFunnelPreTrainedModel.dummy_inputs8sRWWV288<==rAN) r\r]r^r_r%rbase_model_prefixpropertyrr/rAr@rr/s&  L  >>rArc<eZdZUdZdZded<dZded<dZded<y)TFFunnelForPreTrainingOutputa& Output type of [`FunnelForPreTraining`]. Args: logits (`tf.Tensor` of shape `(batch_size, sequence_length)`): Prediction scores of the head (scores for each token 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. Nz tf.TensorrzTuple[tf.Tensor] | NonerRrS)r\r]r^r_rrrRrSr/rAr@rr>s)&FI-1M*1*.J'.rAraa The Funnel Transformer model was proposed in [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le. 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 ([`XxxConfig`]): 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) 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). z The base Funnel Transformer Model transformer outputting raw hidden-states without upsampling head (also called decoder) or any task-specific head on top. ceZdZdfd Zeej dedee e d d dZ dZ d dZ xZS) TFFunnelBaseModelcPt||g|i|t|d|_yNrr9)r0r1rtrr=r2rSr>r?s r@r1zTFFunnelBaseModel.__init__( 3&3F3'X> rAbatch_size, sequence_lengthfunnel-transformer/small-base checkpoint output_typerc 4|j||||||||SN)rVrzr{rWrrUrVrTr r=rVrzr{rWrrUrVrTs r@rYzTFFunnelBaseModel.call2${{))'/!5#  rAcZt|j|j|jSNrPr rQrRrSr=rs r@serving_outputz TFFunnelBaseModel.serving_output,!$66 ..((  rAc|jryd|_t|ddNtj|jj 5|jj ddddyy#1swYyxYwNTrrLrMrHrIrr-rNrPs r@rNzTFFunnelBaseModel.buildi ::  44 ( 4t{{//0 ( !!$' ( ( 5 ( ( A11A:r2r%returnNonerrVTFModelInputType | Nonerznp.ndarray | tf.Tensor | Noner{rrWrrOptional[bool]rUrrVrrTboolrz*Union[Tuple[tf.Tensor], TFBaseModelOutput]rZ)r\r]r^r1r!FUNNEL_INPUTS_DOCSTRINGformatrr _CONFIG_FOR_DOCrrYrrNr`ras@r@rrs?++B+I+IJg+hi2%$ .28<8<7;,0/3&* * 6 6  5  *  - $   4  j , (rArzlThe bare Funnel Transformer Model transformer outputting raw hidden-states without any specific head on top.ceZdZdfd Zeeejdede e  d d dZ dZ d dZ xZS) TFFunnelModelcPt||g|i|t|d|_yr)r0r1rrrs r@r1zTFFunnelModel.__init__rrArfunnel-transformer/smallrc 4|j||||||||Srrrs r@rYzTFFunnelModel.callrrAcZt|j|j|jSrrrs r@rzTFFunnelModel.serving_outputrrAc|jryd|_t|ddNtj|jj 5|jj ddddyy#1swYyxYwrrrPs r@rNzTFFunnelModel.buildrrrrrrZ)r\r]r^r1rr!rrrr rrYrrNr`ras@r@rrs ?*+B+I+IJg+hi-%$.28<8<7;,0/3&* * 6 6  5  *  - $   4 j , (rArz| Funnel model with a binary classification head on top as used during pretraining for identifying generated tokens. ceZdZdfd Zeeejdee e  d d dZ dZ d dZ xZS) TFFunnelForPreTrainingc nt||fi|t|d|_t |d|_y)Nrr9discriminator_predictions)r0r1rrrrr<s r@r1zTFFunnelForPreTraining.__init__,s3 *6*'X> )I&Wr)s&rAr)rrc |j||||||||} | d} |j| } |s | f| ddzSt| | j| jS)a Returns: Examples: ```python >>> from transformers import AutoTokenizer, TFFunnelForPreTraining >>> import torch >>> tokenizer = AutoTokenizer.from_pretrained("funnel-transformer/small") >>> model = TFFunnelForPreTraining.from_pretrained("funnel-transformer/small") >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf") >>> logits = model(inputs).logits ```rVrTrr$NrrRrS)rrrrRrS) r=rVrzr{rWrrUrVrTr>rdiscriminator_sequence_outputrs r@rYzTFFunnelForPreTraining.call2s<'+kk      #'2 ' #)DA(F%//0MN9:12>> >+5CC2==  rAcZt|j|j|jSNr)rrrRrSrs r@rz%TFFunnelForPreTraining.serving_outputfs),==0D0DQWQbQb  rAc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYw)NTrr)rLrMrHrIrr-rNrrPs r@rNzTFFunnelForPreTraining.buildms ::  44 ( 4t{{//0 ( !!$' ( 44d ; Gt==BBC ;..44T: ; ; H ( ( ; ;r?rr)rVrrzrr{rrWrrrrUrrVrrTrrz5Union[Tuple[tf.Tensor], TFFunnelForPreTrainingOutput]rZ)r\r]r^r1rr!rrr#rrrYrrNr`ras@r@rr%st *+B+I+IJg+hi+GVef.28<8<7;,0/3&*/ */ 6/ 6 / 5 / * / -/ $/ /  ?/ gj/ b  ;rArz4Funnel Model with a `language modeling` head on top.ceZdZd fd Zd dZd dZeeejde de e  d ddZ ddZdd ZxZS)TFFunnelForMaskedLMct||g|i|t|d|_t ||jj d|_y)Nrr9lm_head)r0r1rrrrvrrs r@r1zTFFunnelForMaskedLM.__init__{sC 3&3F3'X> +FDKK4J4JQZ[ rAc|jSrZ)rr{s r@ get_lm_headzTFFunnelForMaskedLM.get_lm_heads ||rActjdt|jdz|jjzS)NzMThe method get_prefix_bias_name is deprecated. Please use `get_bias` instead./)warningswarn FutureWarningr-rr{s r@get_prefix_bias_namez(TFFunnelForMaskedLM.get_prefix_bias_names/ egtuyy3!2!222rArrrc |j|||||||| } | d} |j| | } |dn|j|| } |s| f| ddz}| | f|zS|St| | | j| j S)a labels (`tf.Tensor` 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]` rrrrNr$lossrrRrS)rrhf_compute_lossr rRrS)r=rVrzr{rWrrUrVlabelsrTrsequence_outputprediction_scoresrrs r@rYzTFFunnelForMaskedLM.calls2++      #  "!* LL8LL~t4+?+?HY+Z')GABK7F)-)9TGf$ Ev E$!//))   rAcZt|j|j|jSr )r rrRrSrs r@rz"TFFunnelForMaskedLM.serving_outputs$ v}}FDXDXekevevwwrAc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYw)NTrr)rLrMrHrIrr-rNrrPs r@rNzTFFunnelForMaskedLM.builds ::  44 ( 4t{{//0 ( !!$' ( 4D ) 5t||001 ) ""4( ) ) 6 ( ( ) )r?r)rr)rstr NNNNNNNNF)rVrrzrr{rrWrrrrUrrVrrrrTrrz)Union[Tuple[tf.Tensor], TFMaskedLMOutput])rr rr rZ)r\r]r^r1rrrr!rrrr rrYrrNr`ras@r@r r ys\ 3*+B+I+IJg+hi-$$.28<8<7;,0/3&*04* ** 6* 6 * 5 * * * -* $* .* *  3* j* Xx )rAr z Funnel Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled output) e.g. for GLUE tasks. ceZdZdfd Zeeejdede e  d d dZ d dZ d dZ xZS) !TFFunnelForSequenceClassificationct||g|i||j|_t|d|_t ||jd|_y)Nrr9 classifier)r0r1 num_labelsrtrrr&rs r@r1z*TFFunnelForSequenceClassification.__init__sL 3&3F3 ++'X> 4VV=N=NUabrArrrc |j|||||||| } | d} | dddf} |j| | } |dn|j|| }|s| f| ddz}||f|zS|St|| | j| j S)a labels (`tf.Tensor` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). rrNrrr$r)rr&rrrRrS)r=rVrzr{rWrrUrVrrTrrQ pooled_outputrrrs r@rYz&TFFunnelForSequenceClassification.calls2++      #  $AJ)!Q$/ B~t4+?+?+OY,F)-)9TGf$ Ev E)!//))   rAcZt|j|j|jSr )rrrRrSrs r@rz0TFFunnelForSequenceClassification.serving_outputs)*==0D0DQWQbQb  rAc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYwNTrr&rLrMrHrIrr-rNr&rPs r@rNz'TFFunnelForSequenceClassification.build ::  44 ( 4t{{//0 ( !!$' ( 4t , 8t334 ,%%d+ , , 9 ( ( , ,r?rr")rVrrzrr{rrWrrrrUrrVrrrrTrrz3Union[Tuple[tf.Tensor], TFSequenceClassifierOutput])rrrrrZ)r\r]r^r1rr!rrrrrrYrrNr`ras@r@r$r$sc*+B+I+IJg+hi2.$.28<8<7;,0/3&*04+ *+ 6+ 6 + 5 + * + -+ $+ .+ +  =+ j+ Z  ,rAr$z Funnel 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. ceZdZd fd ZedZeeejde de e  d d dZ d dZd dZxZS)TFFunnelForMultipleChoicecvt||g|i|t|d|_t |dd|_y)Nrr9r$r&)r0r1rtrrr&rs r@r1z"TFFunnelForMultipleChoice.__init__*s8 3&3F3'X> 4VQ\RrAcPdtjdtjiS)NrV)rrrsrr{s r@rz&TFFunnelForMultipleChoice.dummy_inputs0sRWWYbhh?@@rAz(batch_size, num_choices, sequence_lengthrrc |t|d} t|d} nt|d} t|d} |tj|d| fnd} |tj|d| fnd} |tj|d| fnd}|%tj|d| t|dfnd}|j| | |||||| }|d}|dddf}|j || }tj|d| f}|dn|j ||}|s|f|ddz}||f|zS|St |||j|j S) a5 labels (`tf.Tensor` of shape `(batch_size,)`, *optional*): Labels for computing the multiple choice classification loss. Indices should be in `[0, ..., num_choices]` where `num_choices` is the size of the second dimension of the input tensors. (See `input_ids` above) Nr$rdrr)rzr{rWrrUrVrTrrrr) rrHrrr&rr rRrS)r=rVrzr{rWrrUrVrrT num_choicesrflat_input_idsflat_attention_maskflat_token_type_idsflat_inputs_embedsrrQr)rreshaped_logitsrrs r@rYzTFFunnelForMultipleChoice.call4s0  $Y/2K#I.q1J$]3A6K#M215JDMDYIJ/?@_cN\Nhbjj"j9IJnrN\Nhbjj"j9IJnr( JJ}r:z-7PQR7S&T U  ++ ..,/!5#  $AJ)!Q$/ B**Vb+->?~t4+?+?+X%''!"+5F)-)9TGf$ Ev E*"!//))   rAcZt|j|j|jSr )r rrRrSrs r@rz(TFFunnelForMultipleChoice.serving_outputys)+==0D0DQWQbQb  rAc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYwr,r-rPs r@rNzTFFunnelForMultipleChoice.buildr.r?rr")rVrrzrr{rrWrrrrUrrVrrrrTrrz4Union[Tuple[tf.Tensor], TFMultipleChoiceModelOutput])rr rr rZ)r\r]r^r1rrrr!rrrr rrYrrNr`ras@r@r0r0"sS AA*+B+I+IJt+uv2/$.28<8<7;,0/3&*04< *< 6< 6 < 5 < * < -< $< .< <  >< w< |  ,rAr0z Funnel Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. ceZdZdfd Zeeejdede e  d d dZ d dZ d dZ xZS) TFFunnelForTokenClassificationcdt||g|i||j|_t|d|_t j j|j|_ t j j|jt|jd|_ ||_y)Nrr9r&r)r0r1r'rrrr5r9r:r;rrrEr&r2rs r@r1z'TFFunnelForTokenClassification.__init__s 3&3F3 ++'X> ||++F,A,AB ,,,,   /&BZBZ2[bn-  rArrrc "|j|||||||| } | d} |j| | } |j| } |dn|j|| } |s| f| ddz}| | f|zS|St | | | j | j S)z labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`. rrrrNr$r)rr;r&rrrRrS)r=rVrzr{rWrrUrVrrTrrrrrs r@rYz#TFFunnelForTokenClassification.calls.++      #  "!*,,,J1~t4+?+?+OY,F)-)9TGf$ Ev E&!//))   rAcZt|j|j|jSr )rrrRrSrs r@rz-TFFunnelForTokenClassification.serving_outputs)'==0D0DQWQbQb  rAc|jryd|_t|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYwr,) rLrMrHrIrr-rNr&r2r3rPs r@rNz$TFFunnelForTokenClassification.build ::  44 ( 4t{{//0 ( !!$' ( 4t , 8t334 M%%tT4;;3J3J&KL M M 9 ( ( M M C"%3C."C+.C7rr")rVrrzrr{rrWrrrrUrrVrrrrTrrz0Union[Tuple[tf.Tensor], TFTokenClassifierOutput])rrrrrZ)r\r]r^r1rr!rrrrrrYrrNr`ras@r@r>r>s *+B+I+IJg+hi-+$.28<8<7;,0/3&*04* ** 6* 6 * 5 * * * -* $* .* *  :* j* X  MrAr>z Funnel Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of the hidden-states output to compute `span start logits` and `span end logits`). ceZdZdfd Zeeejdede e  d d dZ d dZ d dZ xZS) TFFunnelForQuestionAnsweringct||g|i||j|_t|d|_t j j|jt|jd|_ ||_ y)Nrr9 qa_outputsr) r0r1r'rrrr5rrrErHr2rs r@r1z%TFFunnelForQuestionAnswering.__init__sr 3&3F3 ++'X> ,,,,   /&BZBZ2[bn-  rArrrc |j|||||||| } | d} |j| } tj| dd\}}tj|d}tj|d}d}|| || d}|j |||f}|s||f| ddz}||f|zS|St |||| j| j S) a start_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*): Labels for position (index) of the start of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. end_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*): Labels for position (index) of the end of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. rrrdrrN)start_position end_positionr$)r start_logits end_logitsrRrS) rrHrHrrrr rRrS)r=rVrzr{rWrrUrVstart_positions end_positionsrTrrrrLrMrrrs r@rYz!TFFunnelForQuestionAnswering.calls>++      #  "!*1#%88FAB#? jzz,R8 ZZ 4   &=+D(7WF''z0JKD"J/'!"+=F)-)9TGf$ Ev E-%!!//))   rAcpt|j|j|j|jS)N)rLrMrRrS)r rLrMrRrSrs r@rz+TFFunnelForQuestionAnswering.serving_output8s5.,,(( ..((   rAc|jryd|_t|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYw)NTrrH) rLrMrHrIrr-rNrHr2r3rPs r@rNz"TFFunnelForQuestionAnswering.buildBrCrDr) NNNNNNNNNF)rVrrzrr{rrWrrrrUrrVrrNrrOrrTrrz7Union[Tuple[tf.Tensor], TFQuestionAnsweringModelOutput])rr rr rZ)r\r]r^r1rr!rrrr rrYrrNr`ras@r@rFrFs*+B+I+IJg+hi-2$.28<8<7;,0/3&*9=7;8 *8 68 6 8 5 8 * 8 -8 $8 78 58 8  A8 j8 t  MrArF)TF)Qr_ __future__rr dataclassesrtypingrrrnumpynp tensorflowrHactivations_tfr modeling_tf_outputsr r r r rrmodeling_tf_utilsrrrrrrrrrrrtf_utilsrrrutilsrrr r!r"r#configuration_funnelr% get_loggerr\loggerrrr5Layerr'rcrrr&r6rArhrjrtrrrrrrFUNNEL_START_DOCSTRINGrrrrr r$r0r>rFr/rAr@rbsp"!))/    SR/   H %  - ++- `s s l  tGELL$6$6tGn Iell00 IF%ELL&&%2L&ell((L&^(=&ell((=&@M) **M)M)`n) **n)n)bOu||'9'9O:!5<<--!HI!3!3I6 >/ > /;/ /2+Z*Z 1(/1(1(hr1(+1( 1(h  K;4K;  K;\PRhiO)13OO)jO)d L,(?A]L,L,^ `, 79M`,`,F OM%<>WOMOMd _M#: