`i/PdZddlmZddlZddlZddlmZddl m Z m Z m Z m Z mZmZddlmZmZmZmZmZmZmZmZmZmZmZmZddlmZmZm Z dd l!m"Z"m#Z#m$Z$m%Z%d d l&m'Z'e%j(e)Z*d Z+d Z,Gddej-j.Z/Gddej-j.Z0Gddej-j.Z1Gddej-j.Z2Gddej-j.Z3Gddej-j.Z4Gddej-j.Z5Gddej-j.Z6Gddej-j.Z7Gd d!ej-j.Z8eGd"d#ej-j.Z9Gd$d%eZ:d&Z;d'Z<e#d(e;Gd)d*e:Z=Gd+d,ej-j.Z>Gd-d.ej-j.Z?e#d/e;Gd0d1e:eZ@Gd2d3ej-j.ZAe#d4e;Gd5d6e:eZBe#d7e;Gd8d9e:eZCe#d:e;Gd;dd?e:eZEgd@ZFdS)AzTF 2.0 ConvBERT model.) annotationsN)get_tf_activation)TFBaseModelOutputTFMaskedLMOutputTFMultipleChoiceModelOutputTFQuestionAnsweringModelOutputTFSequenceClassifierOutputTFTokenClassifierOutput) TFMaskedLanguageModelingLossTFModelInputTypeTFMultipleChoiceLossTFPreTrainedModelTFQuestionAnsweringLossTFSequenceClassificationLossTFSequenceSummaryTFTokenClassificationLossget_initializerkeraskeras_serializable unpack_inputs)check_embeddings_within_bounds shape_liststable_softmax)add_code_sample_docstringsadd_start_docstrings%add_start_docstrings_to_model_forwardlogging)ConvBertConfigzYituTech/conv-bert-baser cBeZdZdZdfd ZddZ dddZxZS)TFConvBertEmbeddingszGConstruct the embeddings from word, position and token_type embeddings.configr c :tjdi|||_|j|_|j|_|j|_t j|j d|_ t j |j |_ dS)N LayerNormepsilonname)rate)super__init__r#embedding_sizemax_position_embeddingsinitializer_rangerlayersLayerNormalizationlayer_norm_epsr%Dropouthidden_dropout_probdropoutselfr#kwargs __class__s /home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/transformers/models/convbert/modeling_tf_convbert.pyr,zTFConvBertEmbeddings.__init__As""6""" $3'-'E$!'!988AV]h8ii|++1K+LL Nctjd5|d|jj|jgt |j|_dddn #1swxYwYtjd5|d|jj |jgt |j|_ dddn #1swxYwYtjd5|d|j |jgt |j|_ dddn #1swxYwY|j rdSd|_ t|dd`tj|jj5|jdd|jjgddddS#1swxYwYdSdS) Nword_embeddingsweight)r(shape initializertoken_type_embeddings embeddingsposition_embeddingsTr%)tf name_scope add_weightr# vocab_sizer-rr/r>type_vocab_sizerAr.rCbuiltgetattrr%r(buildr7 input_shapes r:rKzTFConvBertEmbeddings.buildKs ], - -  //{-t/BC+D,BCC*DK               ]2 3 3  )-!{2D4GH+D,BCC*9**D &               ]0 1 1  '+!3T5HI+D,BCC(7((D $                :  F 4d + + 7t~233 O O$$dD$+2L%MNNN O O O O O O O O O O O O O O O O O O 8 7sIAA""A&)A&ACCC-rfill expand_dimsrangerCrAr%r5) r7rNrPrQrRpast_key_values_lengthrSrMposition_embedstoken_type_embedsfinal_embeddingss r:callzTFConvBertEmbeddings.callis0  !6TUU U  *9dk6L M M MIT[)LLLM //4  !W+Q???N  >5[^Nd=deeelmL)4+C\ZZZIT-GQ_```(?:=NN>>1A>BB<sQQMMMMMMOOOO@'+)-+/*. & & & & & & & & & r;r"c4eZdZfdZdZddZddZxZS) TFConvBertSelfAttentionc ptjdi||j|jzdkr t d|jd|jdt |j|jz }|dkr|j|_d}n|}|j|_||_|j|_|j|jzdkrt d|j|jz|_|j|jz|_ tj |j t|jd|_tj |j t|jd |_tj |j t|jd |_tj |j |jd dtd|jz t|jd |_tj |j|jzddt|j|_tj |j ddt|j|_tj |j|_||_dS)NrzThe hidden size (z6) is not a multiple of the number of attention heads ()rz6hidden_size should be divisible by num_attention_headsquerykernel_initializerr(keyr\samekey_conv_attn_layer)padding activationdepthwise_initializerpointwise_initializerr(conv_kernel_layer)r~r(ryconv_out_layerr*)r+r, hidden_sizenum_attention_headsrbint head_ratioconv_kernel_sizeattention_head_size all_head_sizerr0Denserr/rwrzr\SeparableConv1Dr|rrr3attention_probs_dropout_probr5r#)r7r#r8new_num_attention_headsrr9s r:r,z TFConvBertSelfAttention.__init__s""6"""   : :a ? ?8F$688 4888  #&f&@6CT&T"U"U "Q & &$8DO"#  "9 $/DO#6 & 7   8 8A = =UVV V#)#59S#S !58PP\''  ?6C[3\3\cj(   <%%  ?6C[3\3\ch&  \''  ?6C[3\3\cj(   $)<#?#?    !"1!d6K2K"L"L"1&2J"K"K&$@$ $  "'!3!3  $t'< <$.v/GHH "4" " $l00  !.v/GHH 1  |++F,OPP  r;cxtj||d|j|jf}tj|gdS)NrZrrrperm)rDreshaperr transpose)r7x batch_sizes r:transpose_for_scoresz,TFConvBertSelfAttention.transpose_for_scoress; Jq:r4+CTE]^ _ _|ALLL1111r;Fct|d||}|}|}} |} t j|} | } t j | dj dg} t| d} t j ddgtj dz dz tj dz dz gddgg} |t j djgt j| dt jfdt%j Dd}t j |djj gt j| t j djgt j| | d }t jt| d|j}|t j|z }|||z}t|d}|| }|||z}t j |djjg}t j|gd }t j||}t j|gd  }t j djjg}t j||gd}t j |djjzf}|r||fn|f}|S) NrrZrr_rCONSTANTc xg|]6}tjd|dgtdjg7S)rr)rDslicerr).0irrmixed_query_layerr7s r: z0TFConvBertSelfAttention.call..sW   !QZL]A^A^_`Aacgcu4vww   r;T) transpose_brSrr)rrwrzr\r|rrDmultiplyrrrrconstantrrrpadstackrfrmatmulcastdtypemathsqrtr5rrconcatr)r7 hidden_statesattention_mask head_maskoutput_attentionsrSmixed_key_layermixed_value_layermixed_key_conv_attn_layer query_layer key_layerconv_attn_layerrpaddingsunfold_conv_out_layerattention_scoresdkattention_probs value_layer context_layerconv_outoutputsrrrs` @@@r:rkzTFConvBertSelfAttention.calls ..q1  JJ}55((=11 JJ}55$($<$<]$K$K!//0A:NN --ozJJ +&?ARSS 22?CCJ'82t?TVW:XYY*+<1EEE;d+a/1455sD3DEENR9K4LMM9    WZ **2.0@0F G G+bgll2.>.>>  %/.@ ))9CCC,,,JJ  - 9Oj  B0H$Jbc  l; ==  /;??  ]FFF :nz2t?WY]Yq.rss =(";Q??  JDOd>P,PQ  7H]=/22mM]r;NcJ|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|dd]tj|j j5|j dd|jjgdddn #1swxYwYt|dd]tj|j j5|j dd|jjgdddn #1swxYwYt|dd]tj|j j5|j dd|jjgdddn #1swxYwYt|ddXtj|j j5|j dd|j gdddn #1swxYwYt|dd`tj|jj5|jdd|jjgddddS#1swxYwYdSdS)NTrwrzr\r|rr)rIrJrDrErwr(rKr#rrzr\r|rrrrLs r:rKzTFConvBertSelfAttention.build)sN :  F 4$ ' ' 3tz// H H   $dk.E!FGGG H H H H H H H H H H H H H H H 4 % % 1tx}-- F FdDK,CDEEE F F F F F F F F F F F F F F F 4$ ' ' 3tz// H H   $dk.E!FGGG H H H H H H H H H H H H H H H 4. 5 5 At7<== V V(..dDK>GG8#H''H+.H+!(JJJFrl)rmrnror,rrkrKrqrrs@r:rtrts}<<<<<|222 QQQQfQQQQQQQQr;rtc.eZdZfdZddZddZxZS)TFConvBertSelfOutputc ltjdi|tj|jt |jd|_tj |j d|_ tj |j |_||_dSNdenserxr%r&r*)r+r,rr0rrrr/rr1r2r%r3r4r5r#r6s r:r,zTFConvBertSelfOutput.__init__Bs""6"""\''  ?6C[3\3\cj(   88AV]h8ii|++F,FGG  r;Fc||}|||}|||z}|SNrrr5r%r7r input_tensorrSs r:rkzTFConvBertSelfOutput.callLD =11  ]X FF }|'CDD r;Nc|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|dd`tj|j j5|j dd|jjgddddS#1swxYwYdSdSNTrr% rIrJrDrErr(rKr#rr%rLs r:rKzTFConvBertSelfOutput.buildS :  F 4$ ' ' 3tz// H H   $dk.E!FGGG H H H H H H H H H H H H H H H 4d + + 7t~233 L L$$dD$+2I%JKKK L L L L L L L L L L L L L L L L L L 8 7$(A44A8;A8.(C##C'*C'rrlrmrnror,rkrKrqrrs@r:rrAsh L L L L L L L Lr;rc4eZdZfdZdZddZddZxZS) TFConvBertAttentionc tjdi|t|d|_t |d|_dS)Nr7r(outputr*)r+r,rtself_attentionr dense_outputr6s r:r,zTFConvBertAttention.__init__`sP""6"""5f6JJJ0hGGGr;ctrlNotImplementedError)r7headss r: prune_headszTFConvBertAttention.prune_headsfs!!r;Fc||||||}||d||}|f|ddz}|SNrrr)rr) r7rrrrrS self_outputsattention_outputrs r:rkzTFConvBertAttention.callisf** .)5FQY+   ,,\!_lU],^^#% QRR(88r;Nc|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|ddStj|jj5|jdddddS#1swxYwYdSdS)NTrr)rIrJrDrErr(rKrrLs r:rKzTFConvBertAttention.buildrsX :  F 4)4 0 0 <t2788 0 0#))$/// 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 4 . . :t0566 . .!''--- . . . . . . . . . . . . . . . . . . ; :$A''A+.A+!C  C C rrl)rmrnror,rrkrKrqrrs@r:rr_stHHHHH """ . . . . . . . .r;rc0eZdZfdZdfd ZdZxZS)GroupedLinearLayerc tjdi|||_||_||_||_|j|jz|_|j|jz|_dSNr*)r+r, input_size output_size num_groupsry group_in_dim group_out_dim)r7rrrryr8r9s r:r,zGroupedLinearLayer.__init__sg""6"""$&$"4 Ot>!-@r;Nc|d|j|j|jg|jd|_|d|jg|j|jd|_t |dS)NkernelT)r?r@ trainablebias)r?r@rr) rFrrrryrrrrr+rKr7rMr9s r:rKzGroupedLinearLayer.buildsoo %t'8$/J/ &   OO 4+,$:QY]Ycos$     k"""""r;ct|d}tjtj|d|j|jggd}tj|tj|jgd}tj|gd}tj||d|jg}tj ||j }|S)NrrZ)rrr)rrrr\r) rrDrrrrrrrnnbias_addr)r7rrrs r:rkzGroupedLinearLayer.calls ..q1 LMBIZ3[\\^g^g^g h h Iadk999== > > LIII & & Jq:r4+;< = = ENNN 3 3r;rl)rmrnror,rKrkrqrrs@r:rr~sjAAAAA # # # # # #r;rc,eZdZfdZdZddZxZS)TFConvBertIntermediatec tjdi||jdkr?tj|jt|jd|_ n:t|j |j|jt|jd|_ t|j trt|j |_n |j |_||_dS)Nrrrxrryr(r*)r+r,rrr0rintermediate_sizerr/rrr isinstance hidden_actstrrintermediate_act_fnr#r6s r:r,zTFConvBertIntermediate.__init__s""6"""   ! !++(_VMe=f=fmt,DJJ,"(!,#263K#L#L DJ f' - - 9'89J'K'KD $ $'-'8D $ r;cZ||}||}|Srl)rrr7rs r:rkzTFConvBertIntermediate.calls, =11 00?? r;Nc|jrdSd|_t|dd`tj|jj5|jdd|jjgddddS#1swxYwYdSdS)NTr) rIrJrDrErr(rKr#rrLs r:rKzTFConvBertIntermediate.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 3s(A55A9<A9rlrrrs@r:rrsc* HHHHHHHHr;rc.eZdZfdZddZddZxZS)TFConvBertOutputc tjdi||jdkr?tj|jt|jd|_ n:t|j |j|jt|jd|_ tj |j d|_tj|j|_||_dS)Nrrrxrr%r&r*)r+r,rrr0rrrr/rrrr1r2r%r3r4r5r#r6s r:r,zTFConvBertOutput.__init__s""6"""   ! !++"vG_7`7`gn,DJJ,("!,#263K#L#L DJ88AV]h8ii|++F,FGG  r;Fc||}|||}|||z}|Srrrs r:rkzTFConvBertOutput.callrr;Nc|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|dd`tj|j j5|j dd|jj gddddS#1swxYwYdSdSNTr%r) rIrJrDrEr%r(rKr#rrrrLs r:rKzTFConvBertOutput.builds :  F 4d + + 7t~233 L L$$dD$+2I%JKKK L L L L L L L L L L L L L L L 4$ ' ' 3tz// N N   $dk.K!LMMM N N N N N N N N N N N N N N N N N N 4 3rrrlrrrs@r:rrsh& N N N N N N N Nr;rc.eZdZfdZddZddZxZS)TFConvBertLayerc tjdi|t|d|_t |d|_t |d|_dS)N attentionr intermediaterr*)r+r,rr rr r bert_outputr6s r:r,zTFConvBertLayer.__init__sd""6""",V+FFF26OOO+FBBBr;Fc||||||}|d}||}||||} | f|ddz} | Sr)r r r ) r7rrrrrSattention_outputsrintermediate_output layer_outputrs r:rkzTFConvBertLayer.calls NN >96GRZ+  -Q/"//0@AA''(;=MX`'aa /$5abb$99r;Ncr|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|ddPtj|jj5|jddddn #1swxYwYt|ddStj|jj5|jdddddS#1swxYwYdSdS)NTr r r ) rIrJrDrEr r(rKr r rLs r:rKzTFConvBertLayer.builds :  F 4d + + 7t~233 + +$$T*** + + + + + + + + + + + + + + + 4 . . :t0566 . .!''--- . . . . . . . . . . . . . . . 4 - - 9t/455 - - &&t,,, - - - - - - - - - - - - - - - - - - : 96A''A+.A+!CC C D**D.1D.rrlrrrs@r:r r seCCCCC     - - - - - - - -r;r c0eZdZfdZ ddZddZxZS)TFConvBertEncoderc tjdi|fdtjD|_dS)Nc8g|]}td|S)zlayer_._r)r )rrr#s r:rz.TFConvBertEncoder.__init__..s,lllqof>a>>BBBlllr;r*)r+r,rfnum_hidden_layerslayerr6s ` r:r,zTFConvBertEncoder.__init__sL""6"""llllERXRjLkLklll r;Fc&|rdnd}|rdnd} t|jD]9\} } |r||fz}| |||| ||} | d}|r | | dfz} :|r||fz}|std||| fDSt||| S)Nr*rrrc3K|]}||V dSrlr*)rvs r: z)TFConvBertEncoder.call..1s(hhqZ[ZgZgZgZgZghhr;)last_hidden_stater attentions) enumeratertupler) r7rrrroutput_hidden_states return_dictrSall_hidden_statesall_attentionsr layer_module layer_outputss r:rkzTFConvBertEncoder.calls #7@BBD0:d(44 F FOA|# I$58H$H!(L~y|=NYaM*!,M  F!/=3C2E!E  E 1]4D D  ihh]4E~$Vhhhhh h +;LYg    r;Nc|jrdSd|_t|ddP|jD]J}tj|j5|ddddn #1swxYwYIdSdS)NTr)rIrJrrDrEr(rK)r7rMrs r:rKzTFConvBertEncoder.build7s :  F 4$ ' ' 3 & &]5:..&&KK%%%&&&&&&&&&&&&&&& 4 3 & &sA&&A* -A* rrlrrrs@r:rr skmmmmm" " " " H&&&&&&&&r;rc,eZdZfdZdZddZxZS)!TFConvBertPredictionHeadTransformc tjdi|tj|jt |jd|_t|j trt|j |_ n |j |_ tj|jd|_||_dSr)r+r,rr0rr-rr/rrrrrtransform_act_fnr1r2r%r#r6s r:r,z*TFConvBertPredictionHeadTransform.__init__Bs""6"""\''  !ofF^6_6_fm(    f' - - 6$5f6G$H$HD ! !$*$5D !88AV]h8ii r;c||}||}||}|Srl)rr,r%rs r:rkz&TFConvBertPredictionHeadTransform.callQs= =11 --m<< }55 r;Nc|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|dd`tj|j j5|j dd|jjgddddS#1swxYwYdSdSrrrLs r:rKz'TFConvBertPredictionHeadTransform.buildXrrrlrrrs@r:r*r*Asc      L L L L L L L Lr;r*cteZdZeZfdZdZdZdZdZ dZ e d d Z d d Z xZS) TFConvBertMainLayerc tjdi|t|d|_|j|jkr+t j|jd|_ t|d|_ ||_ dS)NrBrembeddings_projectencoderr*) r+r,r"rBr-rrr0rr2rr3r#r6s r:r,zTFConvBertMainLayer.__init__hs""6""".vLIII  F$6 6 6&+l&8&89KRf&8&g&gD #(i@@@  r;c|jSrl)rBr7s r:get_input_embeddingsz(TFConvBertMainLayer.get_input_embeddingsss r;cL||j_|jd|j_dSNr)rBr>r?rGr7r\s r:set_input_embeddingsz(TFConvBertMainLayer.set_input_embeddingsvs"!&%*[^"""r;ct)z Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base class PreTrainedModel r)r7heads_to_prunes r: _prune_headsz TFConvBertMainLayer._prune_headszs "!r;c|tj|d}tj||ddd|df}tj||}d|z dz}|S)Nrrg?g)rDrdrr)r7rrMrextended_attention_masks r:get_extended_attention_maskz/TFConvBertMainLayer.get_extended_attention_masksn  !W[!44N#%*^k!naQRT_`aTb=c"d"d#%'*A5"I"I#&)@#@H"L&&r;c8|tdg|jjz}|Srl)rr#r)r7rs r: get_head_maskz!TFConvBertMainLayer.get_head_masks%  % %!>>Ir;NFc ||td|t|} n)|t|dd} ntd|tj| d}|tj| d}|||||| } ||| | j} ||}t|dr| | | } | | | |||| | } | S)NzDYou cannot specify both input_ids and inputs_embeds at the same timerZz5You have to specify either input_ids or inputs_embedsrrrr2) rbrrDrdrBr@rrBhasattrr2r3)r7rNrrQrPrrRrr"r#rSrMrr?s r:rkzTFConvBertMainLayer.calls<  ]%>cdd d  "$Y//KK  &$]33CRC8KKTUU U  !W[!44N  !W[!44N <Q^iqrr "&"B"B>S^`m`s"t"t&&y11 4- . . V 33MH3UUM  #   %  r;c|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|ddPtj|jj5|jddddn #1swxYwYt|dd`tj|jj5|jdd|j j gddddS#1swxYwYdSdS)NTrBr3r2) rIrJrDrErBr(rKr3r2r#r-rLs r:rKzTFConvBertMainLayer.builds  :  F 4t , , 8t344 , ,%%d+++ , , , , , , , , , , , , , , , 4D ) ) 5t|011 ) ) ""4((( ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) 4-t 4 4 @t6;<< X X'--tT4;;U.VWWW X X X X X X X X X X X X X X X X X X A @6A''A+.A+!CC C (D77D;>D; NNNNNNNNNFrl)rmrnror config_classr,r6r:r=r@rBrrkrKrqrrs@r:r0r0ds!L     444"""'''*!---]-^ X X X X X X X Xr;r0ceZdZdZeZdZdS)TFConvBertPreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. convbertN)rmrnrorpr rHbase_model_prefixr*r;r:rJrJs' "L"r;rJax 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 ([`ConvBertConfig`]): 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). zbThe bare ConvBERT Model transformer outputting raw hidden-states without any specific head on top.ceZdZfdZeeedee e e  dddZ ddZ xZS)TFConvBertModelcntj|g|Ri|t|d|_dS)NrKr)r+r,r0rKr7r#rar8r9s r:r,zTFConvBertModel.__init__NsB3&333F333+FDDD r;batch_size, sequence_length checkpoint output_typerHNFrNTFModelInputType | Nonernp.array | tf.Tensor | NonerQrPrrRrOr bool | Noner"r#rSrTrU$TFBaseModelOutput | tuple[tf.Tensor]c D|||||||||| |  } | S)N rNrrQrPrrRrr"r#rS)rK) r7rNrrQrPrrRrr"r#rSrs r:rkzTFConvBertModel.callSsA(--))%'/!5#   r;c|jrdSd|_t|ddStj|jj5|jdddddS#1swxYwYdSdS)NTrK)rIrJrDrErKr(rKrLs r:rKzTFConvBertModel.buildvs :  F 4T * * 6t}122 * * ##D))) * * * * * * * * * * * * * * * * * * 7 6sA((A,/A,rG)rNrUrrVrQrVrPrVrrVrRrOrrWr"rWr#rWrSrTrUrXrl)rmrnror,rrCONVBERT_INPUTS_DOCSTRINGformatr_CHECKPOINT_FOR_DOCr_CONFIG_FOR_DOCrkrKrqrrs@r:rNrNIs EEEEE **+D+K+KLi+j+jkk&%$.26:6:4815*.)-,0#' lk]8********r;rNcFeZdZfdZfdZdZdZdZdZdZ xZ S)TFConvBertMaskedLMHeadc ntjdi|||_|j|_||_dSr)r+r,r#r-input_embeddings)r7r#rcr8r9s r:r,zTFConvBertMaskedLMHead.__init__s@""6""" $3 0r;c||jjfddd|_t |dS)NzerosTr)r?r@rr()rFr#rGrr+rKrs r:rKzTFConvBertMaskedLMHead.buildsEOO4;+A*CQXdhouOvv   k"""""r;c|jSrl)rcr5s r:get_output_embeddingsz,TFConvBertMaskedLMHead.get_output_embeddingss $$r;c\||j_t|d|j_dSr8)rcr>rrGr9s r:set_output_embeddingsz,TFConvBertMaskedLMHead.set_output_embeddingss+',$+5e+<+r#rGrrr)r7r seq_lengths r:rkzTFConvBertMaskedLMHead.calls}555a8  -DDW?XYYY  MT5J5Q_cddd  -JPTP[Pf?ghhh ]KK r;) rmrnror,rKrgrirkrmrkrqrrs@r:raras11111##### %%%@@@###>>>r;rac.eZdZfdZddZddZxZS)TFConvBertGeneratorPredictionsc tjdi|tj|jd|_tj|jd|_ ||_ dS)Nr%r&rrr*) r+r,rr0r1r2r%rr-rr#r6s r:r,z'TFConvBertGeneratorPredictions.__init__sj""6"""88AV]h8ii\''(=G'LL  r;Fc||}td|}||}|S)Ngelu)rrr%)r7generator_hidden_statesrSrs r:rkz#TFConvBertGeneratorPredictions.callsC #:;; 1)&11-@@ }55 r;Nc|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|dd`tj|j j5|j dd|jj gddddS#1swxYwYdSdSr) rIrJrDrEr%r(rKr#r-rrrLs r:rKz$TFConvBertGeneratorPredictions.builds :  F 4d + + 7t~233 O O$$dD$+2L%MNNN O O O O O O O O O O O O O O O 4$ ' ' 3tz// H H   $dk.E!FGGG H H H H H H H H H H H H H H H H H H 4 3rrrlrrrs@r:rtrtsh H H H H H H H Hr;rtz6ConvBERT Model with a `language modeling` head on top.ceZdZfdZdZdZeee de e e e  dddZddZxZS)TFConvBertForMaskedLMcdtj|fi|||_t|d|_t |d|_t|jtrt|j|_ n |j|_ t||jj d|_dS)NrKrgenerator_predictionsgenerator_lm_head)r+r,r#r0rKrtr}rrrrr~rarBr~rPs r:r,zTFConvBertForMaskedLM.__init__s**6*** +FDDD %CFQh%i%i%i" f' - - 0/0ABBDOO$/DO!7 @X_r!s!s!sr;c|jSrl)r~r5s r: get_lm_headz!TFConvBertForMaskedLM.get_lm_heads %%r;c0|jdz|jjzS)N/)r(r~r5s r:get_prefix_bias_namez*TFConvBertForMaskedLM.get_prefix_bias_namesy3!7!<<! 667P[c6dd 223Dx2XX~tt4+?+?HY+Z+Z F'),CABB,GGF)-)9TGf$$v E$1?.9     r;cr|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|ddPtj|jj5|jddddn #1swxYwYt|ddStj|jj5|jdddddS#1swxYwYdSdS)NTrKr}r~) rIrJrDrErKr(rKr}r~rLs r:rKzTFConvBertForMaskedLM.build s :  F 4T * * 6t}122 * * ##D))) * * * * * * * * * * * * * * * 40$ 7 7 Ct9>?? 7 7*00666 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 4,d 3 3 ?t5:;; 3 3&,,T222 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 @ ?r NNNNNNNNNNF)rNrUrrrQrrPrrrrRrOrrWr"rWr#rWrrOrSrWrUrrl)rmrnror,rrrrr\r]rr^rr_rkrKrqrrs@r:r{r{s t t t t t&&&===**+D+K+KLi+j+jkk&$$.28<8<6:37*.)-,0#'#' %/ / / / lk]/ b 3 3 3 3 3 3 3 3r;r{c0eZdZdZfdZdZddZxZS)TFConvBertClassificationHeadz-Head for sentence-level classification tasks.c tjdi|tj|jt |jd|_|j |j n|j }tj ||_ tj|j t |jd|_||_dS)Nrrxout_projr*)r+r,rr0rrrr/rclassifier_dropoutr4r3r5 num_labelsrr#)r7r#r8rr9s r:r,z%TFConvBertClassificationHead.__init__s""6"""\''  ?6C[3\3\cj(   *0)B)NF % %TZTn |++,>??  **  /&BZ2[2[bl+    r;c |dddddf}||}||}t|jj|}||}||}|Sr8)r5rrr#rr)r7rr8rs r:rkz!TFConvBertClassificationHead.call/sz !!!Q' " LLOO JJqMM 5 dk4 5 5a 8 8 LLOO MM!  r;Nc|jrdSd|_t|dd]tj|jj5|jdd|jjgdddn #1swxYwYt|dd`tj|j j5|j dd|jjgddddS#1swxYwYdSdS)NTrr) rIrJrDrErr(rKr#rrrLs r:rKz"TFConvBertClassificationHead.build9s :  F 4$ ' ' 3tz// H H   $dk.E!FGGG H H H H H H H H H H H H H H H 4T * * 6t}122 K K ##T41H$IJJJ K K K K K K K K K K K K K K K K K K 7 6rrl)rmrnrorpr,rkrKrqrrs@r:rrsi77  K K K K K K K Kr;rzp ConvBERT Model transformer with a sequence classification/regression head on top e.g., for GLUE tasks. ceZdZfdZeeedee e e  dddZ ddZ xZS)#TFConvBertForSequenceClassificationctj|g|Ri||j|_t|d|_t |d|_dS)NrKr classifier)r+r,rr0rKrrrPs r:r,z,TFConvBertForSequenceClassification.__init__Ls`3&333F333 ++FDDD 6vLQQQr;rQrRNFrNrUrrrQrPrrRrOrrWr"r#rrSrU"tuple | TFSequenceClassifierOutputc |||||||||| |  } || 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). rrQrPrrRrr"r#rSrrNrr)rKrrr rr)r7rNrrQrPrrRrr"r#rrSrrrrs r:rkz(TFConvBertForSequenceClassification.callRs6-- ))%'/!5#   h??~tt4+?+?+O+O FY,F)-)9TGf$$v E)!/)     r;c|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|ddStj|jj5|jdddddS#1swxYwYdSdSNTrKr)rIrJrDrErKr(rKrrLs r:rKz)TFConvBertForSequenceClassification.buildsS :  F 4T * * 6t}122 * * ##D))) * * * * * * * * * * * * * * * 4t , , 8t344 , ,%%d+++ , , , , , , , , , , , , , , , , , , 9 8rr)rNrUrrrQrrPrrrrRrOrrWr"rWr#rWrrOrSrWrUrrl)rmrnror,rrr\r]rr^r r_rkrKrqrrs@r:rrEsRRRRR **+D+K+KLi+j+jkk&.$.28<8<6:37*.)-,0#'#' %- - - - lk]- ^ , , , , , , , ,r;rz ConvBERT Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a softmax) e.g. for RocStories/SWAG tasks. ceZdZfdZeeedee e e  dddZ ddZ xZS)TFConvBertForMultipleChoicec&tj|g|Ri|t|d|_t ||jd|_tj dt|jd|_ ||_ dS)NrKrsequence_summary)r/r(rrrx) r+r,r0rKrr/rrr0rrrr#rPs r:r,z$TFConvBertForMultipleChoice.__init__s3&333F333+FDDD 1 f&>EW! ! !  ,,, /&2J"K"KR^-   r;z(batch_size, num_choices, sequence_lengthrRNFrNrUrrrQrPrrRrOrrWr"r#rrSrU#tuple | TFMultipleChoiceModelOutputc X|+t|d} t|d} n*t|d} t|d} |tj|d| fnd}|tj|d| fnd}|tj|d| fnd}|tj|d| fnd}|+tj|d| t|dfnd}|||||||||| |  }||d| }||}tj|d| f}| dn|| |}| s|f|ddz}||f|zn|St|||j|j S) a5 labels (`tf.Tensor` of shape `(batch_size,)`, *optional*): Labels for computing the multiple choice classification loss. Indices should be in `[0, ..., num_choices]` where `num_choices` is the size of the second dimension of the input tensors. (See `input_ids` above) NrrrZr)r#rSrrr) rrDrrKrrrrrr)r7rNrrQrPrrRrr"r#rrS num_choicesrrflat_input_idsflat_attention_maskflat_token_type_idsflat_position_idsflat_inputs_embedsrrreshaped_logitsrrs r:rkz TFConvBertForMultipleChoice.calls 8  $Y//2K#I..q1JJ$]33A6K#M2215JDMDYIJ/?@@@_cN\Nhbj"j9IJJJnrN\Nhbj"j9IJJJnrJVJbBJ|b*5EFFFhl( J}r:z-7P7PQR7S&T U U U  --        #   &&wqzH&EE((*Vb+->??~tt4+?+?+X+X F%''!""+5F)-)9TGf$$v E*"!/)     r;c|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|ddPtj|jj5|jddddn #1swxYwYt|dd`tj|jj5|jdd|j j gddddS#1swxYwYdSdS)NTrKrr) rIrJrDrErKr(rKrrr#rrLs r:rKz!TFConvBertForMultipleChoice.builds :  F 4T * * 6t}122 * * ##D))) * * * * * * * * * * * * * * * 4+T 2 2 >t49:: 2 2%++D111 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 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 8rFr)rNrUrrrQrrPrrrrRrOrrWr"rWr#rWrrOrSrWrUrrl)rmrnror,rrr\r]rr^rr_rkrKrqrrs@r:rrs     **!(()STT &/$.28<8<6:37*.)-,0#'#' %> > > > ]> @ M M M M M M M Mr;rz ConvBERT 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. ceZdZfdZeeedee e e  dddZ ddZ xZS) TFConvBertForTokenClassificationctj|g|Ri||j|_t|d|_|j|jn|j}tj ||_ tj |jt|j d|_||_dS)NrKrrrx)r+r,rr0rKrr4rr0r3r5rrr/rr#)r7r#rar8rr9s r:r,z)TFConvBertForTokenClassification.__init__s3&333F333 ++FDDD )/)B)NF % %TZTn |++,>?? ,,,  /&BZ2[2[bn-   r;rQrRNFrNrUrrrQrPrrRrOrrWr"r#rrSrUtuple | 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)rKr5rrr rr)r7rNrrQrPrrRrr"r#rrSrsequence_outputrrrs r:rkz%TFConvBertForTokenClassification.calls2-- ))%'/!5#   "!*,,,JJ11~tt4+?+?+O+O FY,F)-)9TGf$$v E&!/)     r;c|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|dd`tj|jj5|jdd|jj gddddS#1swxYwYdSdSr) rIrJrDrErKr(rKrr#rrLs r:rKz&TFConvBertForTokenClassification.buildKs :  F 4T * * 6t}122 * * ##D))) * * * * * * * * * * * * * * * 4t , , 8t344 M M%%tT4;3J&KLLL M M M M M M M M M M M M M M M M M M 9 8$A''A+.A+!(CCCr)rNrUrrrQrrPrrrrRrOrrWr"rWr#rWrrOrSrWrUrrl)rmrnror,rrr\r]rr^r r_rkrKrqrrs@r:rrs     **+D+K+KLi+j+jkk&+$.28<8<6:37*.)-,0#'#' %, , , , lk], \ M M M M M M M Mr;rz ConvBERT 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`). ceZdZfdZeeedee e e  dddZ ddZ xZS)TFConvBertForQuestionAnsweringctj|g|Ri||j|_t|d|_t j|jt|j d|_ ||_ dS)NrKr qa_outputsrx) r+r,rr0rKrr0rrr/rr#rPs r:r,z'TFConvBertForQuestionAnswering.__init___s3&333F333 ++FDDD ,,,  /&BZ2[2[bn-   r;rQrRNFrNrUrrrQrPrrRrOrrWr"r#start_positions end_positionsrSrU&tuple | TFQuestionAnsweringModelOutputc |||||||||| |  } | d}||}tj|dd\}}tj|d}tj|d}d}| #| !d| i}| |d<||||f}| s||f| d dz}||f|zn|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. rrrrZr_Nstart_position end_positionr)r start_logits end_logitsrr) rKrrDsplitsqueezerr rr)r7rNrrQrPrrRrr"r#rrrSrrrrrrrrs r:rkz#TFConvBertForQuestionAnswering.callis;@-- ))%'/!5#   "!*11#%8FAB#?#?#? jz,R888 Z 444   &=+D&8F%2F> "''z0JKKD F"J/'!""+=F)-)9TGf$$v E-%!!/)     r;c|jrdSd|_t|ddPtj|jj5|jddddn #1swxYwYt|dd`tj|jj5|jdd|jj gddddS#1swxYwYdSdS)NTrKr) rIrJrDrErKr(rKrr#rrLs r:rKz$TFConvBertForQuestionAnswering.buildrr) NNNNNNNNNNNF)rNrUrrrQrrPrrrrRrOrrWr"rWr#rWrrOrrOrSrWrUrrl)rmrnror,rrr\r]rr^r r_rkrKrqrrs@r:rrWs**+D+K+KLi+j+jkk&2$.28<8<6:37*.)-,0#',0*. %; ; ; ; lk]; z M M M M M M M Mr;r)r{rrrrr rNrJ)Grp __future__rnumpynp tensorflowrDactivations_tfrmodeling_tf_outputsrrrr r r modeling_tf_utilsr r rrrrrrrrrrtf_utilsrrrutilsrrrrconfiguration_convbertr get_loggerrmloggerr^r_r0Layerr"rtrrrrrr rr*r0rJCONVBERT_START_DOCSTRINGr\rNrartr{rrrrr__all__r*r;r:rs7""""""//////                            SRRRRRRRRR 322222  H % %/"Q Q Q Q Q 5<-Q Q Q hlQlQlQlQlQel0lQlQlQ^LLLLL5<-LLL<.....%,,...>+B"H"H"H"H"HU\/"H"H"HJ$N$N$N$N$Nu|)$N$N$NN-----el(---D1&1&1&1&1& *1&1&1&h L L L L L (: L L LFuXuXuXuXuX%,,uXuXuXp##### 1###(T5ph/*/*/*/*/*//*/* /*d"""""U\/"""JHHHHHU\%7HHH6RTlmmY3Y3Y3Y3Y357SY3Y3nmY3x&K&K&K&K&K5<#5&K&K&KR  F,F,F,F,F,*CEaF,F,  F,R bMbMbMbMbM";=QbMbMbMJ MMMMMMMMMM'@B[MMMMMM` XMXMXMXMXM%>@WXMXMXMv   r;