+ jyRt^RIt^RIt^RIHt^RIt^RIHt^RIHtH t H t ^RI H t ^RIHt^RIHtHtHt^RIHt^R IHt^R IHtHtHtHt^R IHt^R IH t ^R I!H"t"H#t#H$t$^RI%H&t&]$PN!](4t)Rt*RQt+]#!RR7]!RR]"444t,!RR]PZ4t.!RR]PZ4t/!RR]PZ4t0!RR]PZ4t1!RR]PZ4t2!RR]PZ4t3!R R!]4t4!R"R#]PZ4t5!R$R%]PZ4t6!R&R']PZ4t7!R(R)]PZ4t8!R*R+]PZ4t9]#!R,R-]44t:]#!R.R/]:44t;]#!R0R1]:44t<]#!R2R7!R3R4]:44t=]#!R5R7!R6R7]:44t>!R8R9]?]P4tA!R:R;4tB!R<R=]B4tCRRR>ltDRSR?ltERTR@ltFRAtGRURBltHRVRCltIRWRDltJRXREltKRFtLRGtMRHtNRItORJtPRKtQRLtRRMtSRYRNROlltTRPtU.RZOtVR#)[zPyTorch TAPAS model.N) dataclass)nn)BCEWithLogitsLossCrossEntropyLossMSELoss)initialization)ACT2FN)Cache DynamicCacheEncoderDecoderCache)create_bidirectional_mask)GradientCheckpointingLayer)BaseModelOutputBaseModelOutputWithPoolingMaskedLMOutputSequenceClassifierOutput)PreTrainedModel)apply_chunking_to_forward) ModelOutputauto_docstringlogging TapasConfigg|=z7 Output type of [`TapasForQuestionAnswering`]. ) custom_introcDa]tRt^+toRtRtRtRtRtRt V3Rlt Rt Vt R#)TableQuestionAnsweringOutputa loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` (and possibly `answer`, `aggregation_labels`, `numeric_values` and `numeric_values_scale` are provided)): Total loss as the sum of the hierarchical cell selection log-likelihood loss and (optionally) the semi-supervised regression loss and (optionally) supervised loss for aggregations. logits (`torch.FloatTensor` of shape `(batch_size, sequence_length)`): Prediction scores of the cell selection head, for every token. logits_aggregation (`torch.FloatTensor`, *optional*, of shape `(batch_size, num_aggregation_labels)`): Prediction scores of the aggregation head, for every aggregation operator. Nc <V^8dQh/S[PR,;R&S[PR,;R&S[PR,;R&S[S[P,R,;R&S[S[P,R,;R&#)Nlosslogitslogits_aggregation hidden_states attentions)torch FloatTensortuple)format __classdict__s"i/home/luca/screen-checking/.venv/lib/python3.14/site-packages/transformers/models/tapas/modeling_tapas.py __annotate__)TableQuestionAnsweringOutput.__annotate__+s"   d ")#$    $+%&))D07'(**+d29)*e''(4/6+) __name__ __module__ __qualname____firstlineno____doc__rrr r!r"__annotate_func____static_attributes____classdictcell__r's@r(rr+s-&*D'+F3759M26J+r+rc@aa]tRt^CtoRtV3RltRRltRtVtV;t #)TapasEmbeddingsz Construct the embeddings from word, position and token_type embeddings. Same as BertEmbeddings but with a number of additional token type embeddings to encode tabular structure. c <\SV`4\P!VPVP VP R7Vn\P!VPVP 4Vn \VP4F4wr#RV 2p\W\P!W1P 44K6 \VP4Vn\P!VP VP R7Vn\P"!VP$4VnWnR#)) padding_idxtoken_type_embeddings_epsN)super__init__r Embedding vocab_size hidden_size pad_token_idword_embeddingsmax_position_embeddingsposition_embeddings enumeratetype_vocab_sizessetattrlennumber_of_token_type_embeddings LayerNormlayer_norm_epsDropouthidden_dropout_probdropoutconfig)selfrPirGname __class__s&& r(r>TapasEmbeddings.__init__Is  "||F,=,=v?Q?Q_e_r_rs#%<<0N0NPVPbPb#c #,V-D-D#E A+A3/D D -=?Q?Q R S$F0363J3J/K,f&8&8f>S>STzz&"<"<=  r+c(VeVP4pMVP4RRpV^,pVe VPM VPpVEfq\P!V\PVR7pVP ^4P V4pVPP'Ed \VR,VPP^,^R7p\VR,VPP^,^R7p \W4p \W:4^,p \W4p \P!V\PVR7P ^4p \P!\P!VPP ^, VR7W, 4pVf8\P"!WPP$,\PVR7pVfVP'V4pVP)V4pWN,p\+VP$4F+pRV 2pV\-VV4!VRRV3,4, pK- VP/V4pVP1V4pV#) NdtypedeviceNNN) batch_dimsrYr:)rZrZ)rZrZr)sizerYr#arangelong unsqueezeexpandrPreset_position_index_per_cellIndexMaprGProductIndexMap reduce_mingathermin as_tensorrDzerosrJrCrErangegetattrrKrO)rQ input_idstoken_type_ids position_ids inputs_embeds input_shape seq_lengthrY col_index row_index full_indexfirst_position_per_segmentfirst_positionpositionrE embeddingsrRrSs&&&&& r(forwardTapasEmbeddings.forward\s  #..*K',,.s3K ^ %.%:!! @T@T   << %**VTL'11!4;;KHL{{888$^G%Z>Z[\>]jkl $^G%Z>Z[\>]jkl ,YB -7 -QRS-T*!'(B!O << %**VT^^_`a$yyOODKK$G$G!$KTZ[]e]v    !"[[CCCEJJ_eN   00;M"66|D"8 t;;r{r3r4 __classcell__rTr's@@r(r7r7Cs &11r+r7c@aa]tRt^toRV3RlltRRltRtVtV;t#)TapasSelfAttentionc*<\SV`4VPVP,^8wd8\ VR4'g&\ RVP RVP 24hVPVn\ VPVP, 4VnVPVP,Vn\P!VPVP4Vn \P!VPVP4Vn \P!VPVP4Vn \P!VP4VnVP"VnW nR#)embedding_sizezThe hidden size z4 is not a multiple of the number of attention heads N)r=r>rAnum_attention_headshasattr ValueErrorintattention_head_size all_head_sizerLinearquerykeyvaluerMattention_probs_dropout_probrO is_decoder layer_idxrQrPrrTs&&&r(r>TapasSelfAttention.__init__s/     : : :a ?PVXhHiHi"6#5#5"673346  $*#=#= #&v'9'9F  &/+>??,77;;DNNK %+:+P+P(+:+O+O('6$2D.- /"=*,33DNNCHHI.55dnnELLK055lCMMaQRSI**^499,GQQRSUVWK*)=)D)DY]a]k]k)l& %*_FY*Z*ZAEO..t~~>!<< 5H5HR5PQ+dii8P8P.QQ  %/.@ --//0@b/I,,7 _kB %--aAq9DDF "/"4"4"6s";t?Q?Q>S"S%**,CD 6G=/2mM] ??? 22Gr+) rrrOrrrrrrNNNNF r-r.r/r0r>r{r3r4r}r~s@@r(rrs#(AAr+rcDaa]tRt^toV3RltV3RlRltRtVtV;t#)TapasSelfOutputc0<\SV`4\P!VPVP4Vn\P !VPVPR7Vn\P!VP4Vn R#r;N) r=r>rrrAdenserKrLrMrNrOrQrPrTs&&r(r>TapasSelfOutput.__init__s` YYv1163E3EF f&8&8f>S>STzz&"<"<= r+ch<V^8dQhRS[PRS[PRS[P/#rr! input_tensorreturnr#Tensor)r&r's"r(r)TapasSelfOutput.__annotate__1U\\RWR^R^r+cxVPV4pVPV4pVPW,4pV#rrrOrKrQr!rs&&&r(r{TapasSelfOutput.forward5 =1  ]3 }'CD r+rKrrOrr~s@@r(rr> r+rcLaa]tRt^toRV3RlltRV3RlRlltRtVtV;t#)TapasAttentioncf<\SV`4\WR7Vn\ V4VnR#)rN)r=r>rrQroutputrs&&&r(r>TapasAttention.__init__s& &vC %f- r+c <V^8dQhRS[PRS[PR,RS[PR,RS[R,RS[R,RS[S[P,/#)rr!rNrrrrr#rr$r boolr%)r&r's"r(r)TapasAttention.__annotate__sr||))D0 %0047     $;  u|| r+c VPVVVVVR7pVPV^,V4pV3VR,,p V #)rrrrr^NN)rQr) rQr!rrrrr self_outputsattention_outputrs &&&&&&, r(r{TapasAttention.forwardsTyy )"7+/ !  ;;|A F#% R(88r+)rrQrrrr~s@@r(rrs. r+rcDaa]tRtRtoV3RltV3RlRltRtVtV;t#)TapasIntermediateic"<\SV`4\P!VPVP 4Vn\VP\4'd\VP,Vn R#VPVn R#r) r=r>rrrAintermediate_sizerr hidden_actstrrintermediate_act_fnrs&&r(r>TapasIntermediate.__init__s` YYv1163K3KL f'' - -'-f.?.?'@D $'-'8'8D $r+cN<V^8dQhRS[PRS[P/#rr!rr)r&r's"r(r)TapasIntermediate.__annotate__s#U\\ellr+cJVPV4pVPV4pV#rrrrQr!s&&r(r{TapasIntermediate.forwards& =1 00? r+rrr~s@@r(rrs9r+rcDaa]tRtRtoV3RltV3RlRltRtVtV;t#) TapasOutputi%c0<\SV`4\P!VPVP 4Vn\P!VP VPR7Vn\P!VP4Vn R#r) r=r>rrrrArrKrLrMrNrOrs&&r(r>TapasOutput.__init__&s` YYv779K9KL f&8&8f>S>STzz&"<"<= r+ch<V^8dQhRS[PRS[PRS[P/#rr)r&r's"r(r)TapasOutput.__annotate__,rr+cxVPV4pVPV4pVPW,4pV#rrrs&&&r(r{TapasOutput.forward,rr+rrr~s@@r(rr%rr+rcRaa]tRtRtoRV3RlltRV3RlRlltRtRtVtV;t #) TapasLayeri3c<\SV`4VPVn^Vn\ WR7VnVP VnVPVnVP'd2VP 'g\V R24h\ WR7Vn \V4Vn \V4Vn R#)r^rz> should be used as a decoder model if cross attention is addedN)r=r>chunk_size_feed_forward seq_len_dimr attentionradd_cross_attentionrcrossattentionr intermediaterrrs&&&r(r>TapasLayer.__init__4s '-'E'E$'D ++#)#=#=  # # #??? D6)g!hii"0"MD -f5!&) r+c<V^8dQhRS[PRS[PR,RS[PR,RS[PR,RS[R,RS[R,RS[S[P,/#) rr!rNrencoder_attention_maskrrrr)r&r's"r(r)TapasLayer.__annotate__Cs))||)))D0) %0047 ) !& 1 1D 8 )  ) $;) u|| )r+c VPVVVVR7pV^,p VR,p VP'dTVeP\VR4'g\RV R24hVP V VVVVR7p V ^,p WR,,p \ VP VPVPV 4p V 3V ,p V #))rrrrrz'If `encoder_hidden_states` are passed, z` has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`r) rrrrrrfeed_forward_chunkrr) rQr!rrrrrrself_attention_outputsrrcross_attention_outputs layer_outputs &&&&&&&, r(r{TapasLayer.forwardCs"& )/+ "0"  2!4(, ???4@4!122 =dVDDD '+&9&9 5&; /"3 ':' # 7q9  ;;G0  # #T%A%A4CSCSUe  /G+r+cJVPV4pVPW!4pV#r)rr)rQrintermediate_outputr s&& r(r TapasLayer.feed_forward_chunkos)"//0@A{{#6I r+)rrrrrrrrr)NNNNF) r-r.r/r0r>r{r r3r4r}r~s@@r(rr3s *))Xr+rc<aa]tRtRtoV3RltRRltRtVtV;t#) TapasEncoderiuc <\SV`4Wn\P!\ VP 4Uu.uFp\WR7NK up4VnRVn R#uupi)rFN) r=r>rPr ModuleListrlnum_hidden_layersrlayergradient_checkpointing)rQrPrRrTs&& r(r>TapasEncoder.__init__vsS  ]]USYSkSkMl#mMlJv$CMl#mn &+#$nsA'c V'd9Vf5\\VPR7\VPR74pV'dRMRp V'dRMRp \VP4FGwrV'd W3,p V!VVVVVVR7pV^,pV'gK7W^,3,p KI V'd W3,p V 'g1\ ;QJd.RWV 34FNK 5#!RWV 344#\ WV R7#)N)rP)rrrc30"TF qfKVxK R#5irr,).0vs& r( 'TapasEncoder.forward..sh$Vq$Vs )last_hidden_stater!r"r,)r r rPrFrr%r)rQr!rrrr use_cacheroutput_hidden_states return_dictrall_hidden_statesall_attentionsrR layer_module layer_outputss&&&&&&&&&&, r(r{TapasEncoder.forward|s 01,dkk2RT`hlhshsTtuO"6BD0d(4OA#$58H$H!(%'= /"3 M*!,M  !/3C2E!E 5 14D D 5h]~$Vh5 h5h]~$Vhh h+Yg  r+)rPrr)NNNNNFFTrr~s@@r(rrus, ) ) r+rcDaa]tRtRtoV3RltV3RlRltRtVtV;t#) TapasPooleric<\SV`4\P!VPVP4Vn\P !4VnR#r)r=r>rrrArTanh activationrs&&r(r>TapasPooler.__init__s9 YYv1163E3EF '')r+cN<V^8dQhRS[PRS[P/#rr)r&r's"r(r)TapasPooler.__annotate__s#U\\ellr+c\VR,pVPV4pVPV4pV#)rZ)rZr)rr-)rQr!first_token_tensor pooled_outputs&& r(r{TapasPooler.forwards2+40 #56  6 r+)r-rrr~s@@r(r*r*s$ r+r*cDaa]tRtRtoV3RltV3RlRltRtVtV;t#)TapasPredictionHeadTransformic<\SV`4\P!VPVP4Vn\ VP\4'd\VP,Vn MVPVn \P!VPVPR7Vn R#r) r=r>rrrArrrrrtransform_act_fnrKrLrs&&r(r>%TapasPredictionHeadTransform.__init__s~ YYv1163E3EF f'' - -$*6+<+<$=D !$*$5$5D !f&8&8f>S>STr+cN<V^8dQhRS[PRS[P/#rr)r&r's"r(r))TapasPredictionHeadTransform.__annotate__s#U\\ellr+clVPV4pVPV4pVPV4pV#r)rr8rKrs&&r(r{$TapasPredictionHeadTransform.forwards4 =1 --m< }5 r+)rKrr8rr~s@@r(r6r6sUr+r6c8aa]tRtRtoV3RltRtRtVtV;t#)TapasLMPredictionHeadic<\SV`4\V4Vn\P !VP VPRR7Vn\P!\P!VP44Vn R#)T)biasN) r=r>r6 transformrrrAr@decoder Parameterr#rkrArs&&r(r>TapasLMPredictionHead.__init__s[ 5f=yy!3!3V5F5FTR LLV->->!?@ r+cJVPV4pVPV4pV#r)rBrCrs&&r(r{TapasLMPredictionHead.forwards$}5  ]3 r+)rArCrBrr~s@@r(r?r?sAr+r?cDaa]tRtRtoV3RltV3RlRltRtVtV;t#)TapasOnlyMLMHeadicD<\SV`4\V4VnR#r)r=r>r? predictionsrs&&r(r>TapasOnlyMLMHead.__init__s 08r+cN<V^8dQhRS[PRS[P/#)rsequence_outputrr)r&r's"r(r)TapasOnlyMLMHead.__annotate__s#!!u||! !r+c(VPV4pV#rrK)rQrNprediction_scoress&& r(r{TapasOnlyMLMHead.forwards ,,_=  r+rQrr~s@@r(rIrIs9!!!r+rIcnaa]tRtRtoRtRt]P!4V3Rl4tV3Rlt Rt Vt V;t #)TapasPreTrainedModelitapasTc <\SV`V4\V\4'd!\P !VP 4\V\4'Ed VPP'dB\P !VP4\P !VP4Ml\P!VPVPPR7\P!VPVPPR7\P !VP4\P !VP4R#R#)zInitialize the weights)stdN)r= _init_weightsrr?initzeros_rATapasForQuestionAnsweringrP#init_cell_selection_weights_to_zerooutput_weightscolumn_output_weightsnormal_initializer_range output_biascolumn_output_bias)rQmodulerTs&&r(rY"TapasPreTrainedModel._init_weightss f% f3 4 4 KK $ f7 8 8}}@@@ F112 F889 V22 8W8WX V99v}}?^?^_ KK** + KK11 2 9r+c&<V^8dQh/S[;R&#rrPr)r&r's"r(r)!TapasPreTrainedModel.__annotate__s r+r,) r-r.r/r0base_model_prefixsupports_gradient_checkpointingr#no_gradrYr2r3r4r}r~s@@r(rUrUs2 &*# ]]_ 3 3r+rUcfaa]tRtRtoRtR V3RlltRtRt]R V3RlRll4t Rt Vt V;t #) TapasModelia This class is a small change compared to [`BertModel`], taking into account the additional token type ids. The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of cross-attention is added between the self-attention layers, following the architecture described in [Attention is all you need](https://huggingface.co/papers/1706.03762) by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin. c <\SV`V4Wn\V4Vn\ V4VnV'd \V4MRVnVP4R#)z^ add_pooling_layer (bool, *optional*, defaults to `True`): Whether to add a pooling layer N) r=r>rPr7rzrencoderr*pooler post_init)rQrPadd_pooling_layerrTs&&&r(r>TapasModel.__init__ sK   )&1#F+ ->k&)D  r+c.VPP#rrzrCrQs&r(get_input_embeddingsTapasModel.get_input_embeddingss...r+c&WPnR#rru)rQrs&&r(set_input_embeddingsTapasModel.set_input_embeddingss */'r+c<V^8dQhRS[PR,RS[PR,RS[PR,RS[PR,RS[PR,RS[PR,RS[PR,R S[R,R S[R,R S[R,R S[S[,/ #) rrnNrrorprqrrrr"r#r)r# LongTensorr$rr%r)r&r's"r(r)TapasModel.__annotate__ sp p ##d*p ))D0p ((4/ p &&- p ((4/ p  %0047p !& 1 1D 8p  $;p #Tkp D[p  + +p r+c VeTMVPPpV eT MVPPp V eT MVPPp VeVe \ R4hVe#VP W4VP 4p M#VeVP 4RR p M \ R4hVe VPM VPp Vf\P!WR7pVfJ\P!.V O\VPP4N5\PV R7pVPWW5R7p\VPVVR7pVe\VPVVVR7pVP!VVVVVV V R 7pV^,pVP"eVP#V4MRpV 'gVV3VR ,,#\%VVVP&VP(R 7#) a token_type_ids (`torch.LongTensor` of shape `(batch_size, sequence_length, 7)`, *optional*): Token indices that encode tabular structure. Indices can be obtained using [`AutoTokenizer`]. See this class for more info. [What are token type IDs?](../glossary#token-type-ids) position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. If `reset_position_index_per_cell` of [`TapasConfig`] is set to `True`, relative position embeddings will be used. Selected in the range `[0, config.max_position_embeddings - 1]`. [What are position IDs?](../glossary#position-ids) Examples: ```python >>> from transformers import AutoTokenizer, TapasModel >>> import pandas as pd >>> tokenizer = AutoTokenizer.from_pretrained("google/tapas-base") >>> model = TapasModel.from_pretrained("google/tapas-base") >>> data = { ... "Actors": ["Brad Pitt", "Leonardo Di Caprio", "George Clooney"], ... "Age": ["56", "45", "59"], ... "Number of movies": ["87", "53", "69"], ... } >>> table = pd.DataFrame.from_dict(data) >>> queries = ["How many movies has George Clooney played in?", "How old is Brad Pitt?"] >>> inputs = tokenizer(table=table, queries=queries, padding="max_length", return_tensors="pt") >>> outputs = model(**inputs) >>> last_hidden_states = outputs.last_hidden_state ```NzDYou cannot specify both input_ids and inputs_embeds at the same timez5You have to specify either input_ids or inputs_embedsr\rW)rnrprorq)rPrqr)rPrqrr)rrrrr"r#r)r  pooler_outputr!r"r])rPrr"r#r%warn_if_padding_and_no_attention_maskr_rYr#onesrkrIrGrarzr rorprr!r")rQrnrrorprqrrrr"r#rrrrYembedding_outputencoder_outputsrNr3s&&&&&&&&&&&, r(r{TapasModel.forwardsd2C1N-TXT_T_TqTq$8$D $++JjJj &1%cd d  "  6 6y Q#..*K  &',,.s3KTU U%.%:!! @T@T  !"ZZ CN  !"[[A+As4;;#?#?@A\bN ??>+ 3;;*)  " -%>{{.5&; & ",, )"7#9/!5#' *!,8< 8O O4UY #]3ob6II I)-')77&11   r+)rPrzrorp)T) NNNNNNNNNN) r-r.r/r0r1r>rwrzrr{r3r4r}r~s@@r(rmrms; /0p p p p r+rmczaa]tRtRtoRRRR/tRtV3RltRtR t] RV3R lR ll4t V3R lt R t Vt V;t#)TapasForMaskedLMizcls.predictions.decoder.biaszcls.predictions.biaszcls.predictions.decoder.weightz'tapas.embeddings.word_embeddings.weightrVc<\SV`V4\VRR7Vn\ V4VnVP 4R#)F)rrN)r=r>rmrVrIclsrqrs&&r(r>TapasForMaskedLM.__init__s6  %@ #F+ r+cBVPPP#r)rrKrCrvs&r(get_output_embeddings&TapasForMaskedLM.get_output_embeddingssxx##+++r+cWPPnVPVPPnR#r)rrKrCrA)rQnew_embeddingss&&r(set_output_embeddings&TapasForMaskedLM.set_output_embeddingss*'5$$2$7$7!r+c<V^8dQhRS[PR,RS[PR,RS[PR,RS[PR,RS[PR,RS[PR,RS[PR,R S[PR,R S[R,R S[R,R S[R,R S[S[,/ #)rrnNrrorprqrrlabelsrr"r#r)r#r}r$rr%r)r&r's"r(r)TapasForMaskedLM.__annotate__sZ Z ##d*Z ))D0Z ((4/ Z &&- Z ((4/ Z  %0047Z !& 1 1D 8Z   4'Z  $;Z #TkZ D[Z   Z r+c V eT MVPPp VPVVVVVVVV V V R7 p V ^,pVPV4pRpVeG\ 4pV!VP RVPP 4VP R44pV 'g"V3V R,,pVe V3V,#T#\VVV PV PR7#)aZ token_type_ids (`torch.LongTensor` of shape `(batch_size, sequence_length, 7)`, *optional*): Token indices that encode tabular structure. Indices can be obtained using [`AutoTokenizer`]. See this class for more info. [What are token type IDs?](../glossary#token-type-ids) position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. If `reset_position_index_per_cell` of [`TapasConfig`] is set to `True`, relative position embeddings will be used. Selected in the range `[0, config.max_position_embeddings - 1]`. [What are position IDs?](../glossary#position-ids) labels (`torch.LongTensor` 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]` Examples: ```python >>> from transformers import AutoTokenizer, TapasForMaskedLM >>> import pandas as pd >>> tokenizer = AutoTokenizer.from_pretrained("google/tapas-base") >>> model = TapasForMaskedLM.from_pretrained("google/tapas-base") >>> data = { ... "Actors": ["Brad Pitt", "Leonardo Di Caprio", "George Clooney"], ... "Age": ["56", "45", "59"], ... "Number of movies": ["87", "53", "69"], ... } >>> table = pd.DataFrame.from_dict(data) >>> inputs = tokenizer( ... table=table, queries="How many [MASK] has George [MASK] played in?", return_tensors="pt" ... ) >>> labels = tokenizer( ... table=table, queries="How many movies has George Clooney played in?", return_tensors="pt" ... )["input_ids"] >>> outputs = model(**inputs, labels=labels) >>> logits = outputs.logits ```N) rrorprqrrrr"r#rNNrrr!r"r]) rPr#rVrrrr@rr!r")rQrnrrorprqrrrrr"r#rrrNrRmasked_lm_lossloss_fctrs&&&&&&&&&&&&, r(r{TapasForMaskedLM.forwardsv&1%rrrr{r2r3r4r}r~s@@r(rrsU '(>(*S  ,8Z Z Z 5r+ra Tapas Model with a cell selection head and optional aggregation head on top for question-answering tasks on tables (linear layers on top of the hidden-states output to compute `logits` and optional `logits_aggregation`), e.g. for SQA, WTQ or WikiSQL-supervised tasks. c^aa]tRtRtoV3RlV3Rllt]RV3RlRll4tRtVtV;t #)r\i c <V^8dQhRS[/#rgr)r&r's"r(r)&TapasForQuestionAnswering.__annotate__s{r+c<\SV`V4\V4Vn\P !VP 4Vn\P!\P!VP44Vn \P!\P!VP44Vn \P!\P!.44Vn\P!\P!.44VnVP ^8d1\P"!VPVP 4VnVP'4R#)rN)r=r>rmrVrrMrNrOrDr#emptyrAr^r_rbrcnum_aggregation_labelsraggregation_classifierrqrs&&r(r>"TapasForQuestionAnswering.__init__s   ' zz&"<"<= !ll5;;v7I7I+JK%'\\%++f>P>P2Q%R"<< B8"$,,u{{2"?  ( (1 ,*,))F4F4FHeHe*fD ' r+c$<V^8dQhRS[PR,RS[PR,RS[PR,RS[PR,RS[PR,RS[PR,RS[PR,R S[PR,R S[PR,R S[PR,R S[PR,R S[R,RS[R,RS[R,RS[S[,/#)rrnNrrorprq table_maskraggregation_labels float_answernumeric_valuesnumeric_values_scalerr"r#r)r#r}r$rr%r)r&r's"r(r)r)s=q q ##d*q ))D0q ((4/ q &&- q ((4/ q $$t+q   4'q ",,t3q ''$.q ))D0q $//$6q  $;q #Tkq D[q " - -#q r+c VeTMVPPpVPVVVVVV V VR7pV^,pV^,pVPV4pVeVP 4pMVP 4RRpVe VP M VP pVfJ\ P!.VO\VPP4N5\ PVR7p.ROpVRRVPR43,pVRRVPR43,p\\ P!V\ P!VPP^, VP R74VPP^R7p\\ P!V\ P!VPP ^, VP R74VPP ^R7p\#VV4pVeVP 4MVP 4RRpVe VP M VP pVf\ P$!VVR7pVfD\ P&!V^8\ P(!V4\ P*!V44pVP-V\ P.R 7pVP-V\ P.R 7p\1VV4wpp\3VVPP4VP6VP84pRp VPP:'d9\=VVP>VP@VVVPPB4p Rp!VPPD^8dVPGV4p!R p"R p#VEeR p#VPPD^8*;'gVPPH'*p$V$'dRp%MmV e_VPJ^,V PJ^,8XgQR 4h\MV VVPPNVVPF4p%M \QR4hVPPR'd\1VV4wp&p\UV&V4p\ PVPYVR7p'Rp(VPP:'g\ P&!V^8H\ P(!V\ PZR7VPP\\ P(!V\ PZR7,4p)V'P_V4)V),p*\ P`!V*V,^R7\ P`!V^R7\b,, p(M3\eVV VVVV4wp(p\ PVPYVR7p'VPPf'dMQV$'dV"\ Ph!V(4, p"M+V"\ Ph!V(RV%, ,4, p"VPPD^8EdV$'dVe~VPJ^,VPJ^,8XgQR4h\kV!V%VVPPHVPPDVPPl4p+M\QR4h\ P!VPJ^,\ PVP R7p\kV!V%VVPPHVPPDVPPl4p+VPPH'dcV eTV ePV PJV PJ8XgQh\oV V%V'V V VV!VP4wp,p-V+V,, p+V+V-,p+M \QR4hV"\ Ph!V+4, p"M)\ P*!V4p\eVV VVVV4wppV'g'VV!3VR,,p.V#'d V"3V.,#T.#\qV#'dT"MRVV!VPrVPtR7#)a token_type_ids (`torch.LongTensor` of shape `(batch_size, sequence_length, 7)`, *optional*): Token indices that encode tabular structure. Indices can be obtained using [`AutoTokenizer`]. See this class for more info. [What are token type IDs?](../glossary#token-type-ids) position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. If `reset_position_index_per_cell` of [`TapasConfig`] is set to `True`, relative position embeddings will be used. Selected in the range `[0, config.max_position_embeddings - 1]`. [What are position IDs?](../glossary#position-ids) table_mask (`torch.LongTensor` of shape `(batch_size, seq_length)`, *optional*): Mask for the table. Indicates which tokens belong to the table (1). Question tokens, table headers and padding are 0. labels (`torch.LongTensor` of shape `(batch_size, seq_length)`, *optional*): Labels per token for computing the hierarchical cell selection loss. This encodes the positions of the answer appearing in the table. Can be obtained using [`AutoTokenizer`]. - 1 for tokens that are **part of the answer**, - 0 for tokens that are **not part of the answer**. aggregation_labels (`torch.LongTensor` of shape `(batch_size, )`, *optional*): Aggregation function index for every example in the batch for computing the aggregation loss. Indices should be in `[0, ..., config.num_aggregation_labels - 1]`. Only required in case of strong supervision for aggregation (WikiSQL-supervised). float_answer (`torch.FloatTensor` of shape `(batch_size, )`, *optional*): Float answer for every example in the batch. Set to *float('nan')* for cell selection questions. Only required in case of weak supervision (WTQ) to calculate the aggregate mask and regression loss. numeric_values (`torch.FloatTensor` of shape `(batch_size, seq_length)`, *optional*): Numeric values of every token, NaN for tokens which are not numeric values. Can be obtained using [`AutoTokenizer`]. Only required in case of weak supervision for aggregation (WTQ) to calculate the regression loss. numeric_values_scale (`torch.FloatTensor` of shape `(batch_size, seq_length)`, *optional*): Scale of the numeric values of every token. Can be obtained using [`AutoTokenizer`]. Only required in case of weak supervision for aggregation (WTQ) to calculate the regression loss. Examples: ```python >>> from transformers import AutoTokenizer, TapasForQuestionAnswering >>> import pandas as pd >>> tokenizer = AutoTokenizer.from_pretrained("google/tapas-base-finetuned-wtq") >>> model = TapasForQuestionAnswering.from_pretrained("google/tapas-base-finetuned-wtq") >>> data = { ... "Actors": ["Brad Pitt", "Leonardo Di Caprio", "George Clooney"], ... "Age": ["56", "45", "59"], ... "Number of movies": ["87", "53", "69"], ... } >>> table = pd.DataFrame.from_dict(data) >>> queries = ["How many movies has George Clooney played in?", "How old is Brad Pitt?"] >>> inputs = tokenizer(table=table, queries=queries, padding="max_length", return_tensors="pt") >>> outputs = model(**inputs) >>> logits = outputs.logits >>> logits_aggregation = outputs.logits_aggregation ```Nrrorprqrr"r#rW column_idsrow_idsrZr\indices num_segmentsr[)rYrXgFTz>Make sure the answers are a FloatTensor of shape (batch_size,)zJYou have to specify float answers in order to calculate the aggregate maskrrXr?zHMake sure the aggregation labels are a LongTensor of shape (batch_size,)zQYou have to specify aggregation labels in order to calculate the aggregation losszeYou have to specify numeric values and numeric values scale in order to calculate the regression lossr)rrr r!r"r]) segment_idsrr prev_labels column_ranksinv_column_ranksnumeric_relations);rPr#rVrOr_rYr#rkrIrGraindexrerirj max_num_rowsmax_num_columnsrfrwhere ones_like zeros_liketofloat reduce_meancompute_token_logits temperaturer^rbselect_one_columncompute_column_logitsr_rcallow_empty_column_selectionrruse_answer_as_supervisionr_calculate_aggregate_maskcell_selection_preferenceraverage_logits_per_cellrh distributions Bernoullifloat32positive_label_weightlog_probsumEPSILON_ZERO_DIVISION"_single_column_cell_selection_lossdisable_per_token_lossmean_calculate_aggregation_lossaggregation_loss_weight_calculate_regression_lossrr!r")/rQrnrrorprqrrrrrrrr"r#rrrNr3rrrY token_typesrrrurt cell_indexinput_mask_floattable_mask_float cell_mask_r column_logitsr total_losscalculate_loss is_supervisedaggregate_masklogits_per_celldist_per_tokenselection_loss_per_exampleweightselection_loss_per_tokenper_example_additional_loss answer_losslarge_answer_loss_maskrs/&&&&&&&&&&&&&&&, r(r{!TapasForQuestionAnswering.forward(s\&1%!>f!>!M{{111ejj)CDD ejj)Cs^G[)\]] {{11A5 )5%||A2D2J2J12MMfM7R.*. KKAA KK>> KK?? 73)o */V\\!_EJJ_e_l_l)m&2M*&* == :: ;; 3/;;888%16J6V-337K7Q7QQQQ>X(***0,. KK ?; %;4{B337MM3(/ ejj)DEE %%f-F: vz9iIAv01GBK?F/=ZMF* I6 I+-41!//))   r+)rrcr_rOrbr^rV)NNNNNNNNNNNNNN r-r.r/r0r>rr{r3r4r}r~s@@r(r\r\ s1,q q q q r+r\z Tapas Model with a sequence classification head on top (a linear layer on top of the pooled output), e.g. for table entailment tasks, such as TabFact (Chen et al., 2020). cRaa]tRtRtoV3Rlt]RV3RlRll4tRtVtV;t #)TapasForSequenceClassificationi]c2<\SV`V4VPVn\V4Vn\ P !VP4Vn\ P!VPVP4Vn VP4R#r) r=r> num_labelsrmrVrrMrNrOrrA classifierrqrs&&r(r>'TapasForSequenceClassification.__init__dsi   ++' zz&"<"<= ))F$6$68I8IJ r+c<V^8dQhRS[PR,RS[PR,RS[PR,RS[PR,RS[PR,RS[PR,RS[R,R S[R,R S[R,R S[S[P ,S[,/ #) rrnNrrorprqrrr"r#r)r#r}r$rr%rr)r&r's"r(r)+TapasForSequenceClassification.__annotate__psl l ##d*l ))D0l ((4/ l &&- l ((4/ l   4'l  $;l #Tkl D[l  u|| 7 7l r+c V eT MVPPp VPVVVVVVVV R7p V ^,p VPV 4p VP V 4p RpVEeVPP fVP ^8XdRVPnMrVP ^8dQVP\P8Xg VP\P8XdRVPnMRVPnVPP R8XdK\4pVP ^8Xd'V!V P4VP44pMV!W4pMVPP R8Xd>\4pV!V PRVP 4VPR44pM-VPP R8Xd\4pV!W4pV 'g"V 3V R,,pVe V3V,#T#\!VV V P"V P$R7#) a. token_type_ids (`torch.LongTensor` of shape `(batch_size, sequence_length, 7)`, *optional*): Token indices that encode tabular structure. Indices can be obtained using [`AutoTokenizer`]. See this class for more info. [What are token type IDs?](../glossary#token-type-ids) position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. If `reset_position_index_per_cell` of [`TapasConfig`] is set to `True`, relative position embeddings will be used. Selected in the range `[0, config.max_position_embeddings - 1]`. [What are position IDs?](../glossary#position-ids) labels (`torch.LongTensor` 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). Note: this is called "classification_class_index" in the original implementation. Examples: ```python >>> from transformers import AutoTokenizer, TapasForSequenceClassification >>> import torch >>> import pandas as pd >>> tokenizer = AutoTokenizer.from_pretrained("google/tapas-base-finetuned-tabfact") >>> model = TapasForSequenceClassification.from_pretrained("google/tapas-base-finetuned-tabfact") >>> data = { ... "Actors": ["Brad Pitt", "Leonardo Di Caprio", "George Clooney"], ... "Age": ["56", "45", "59"], ... "Number of movies": ["87", "53", "69"], ... } >>> table = pd.DataFrame.from_dict(data) >>> queries = [ ... "There is only one actor who is 45 years old", ... "There are 3 actors which played in more than 60 movies", ... ] >>> inputs = tokenizer(table=table, queries=queries, padding="max_length", return_tensors="pt") >>> labels = torch.tensor([1, 0]) # 1 means entailed, 0 means refuted >>> outputs = model(**inputs, labels=labels) >>> loss = outputs.loss >>> logits = outputs.logits ```Nr regressionsingle_label_classificationmulti_label_classificationrrr])rPr#rVrOr problem_typerrXr#rarrsqueezerrrrr!r")rQrnrrorprqrrr"r#rrr3rrrrs&&&&&&&&&&, r(r{&TapasForSequenceClassification.forwardosx&1%IndexMap.__init__s2w~~F !OOLP$r+cPVPP4RVP#r)rr_r[rvs&r( batch_shapeIndexMap.batch_shapes!||  "#4T__55r+)r[rrN)r) r-r.r/r0r1r>r r3r4r5s@r(reres1%&66r+recBaa]tRtRtoRtV3RltRtRtRtVt V;t #)rfizThe product of two indices.c 2<VPVP8wd \R4h\SV` VPVPVP ,,VP VP ,VPR7WnW nR#)av Combines indices i and j into pairs (i, j). The result is an index where each segment (i, j) is the intersection of segments i and j. For example if the inputs represent table cells indexed by respectively rows and columns the output will be a table indexed by (row, column) pairs, i.e. by cell. The implementation combines indices {0, .., n - 1} and {0, .., m - 1} into {0, .., nm - 1}. The output has *num_segments* equal to *outer_index.num_segments* * *inner_index.num_segments* Args: outer_index (`IndexMap`): IndexMap. inner_index (`IndexMap`): IndexMap, must have the same shape as *outer_index*. zCouter_index.batch_dims and inner_index.batch_dims must be the same.rN)r[rr=r>rr outer_index inner_index)rQr rrTs&&&r(r>ProductIndexMap.__init__s  ! ![%;%; ;bc c  ((;+>+>AYAY+YY$11K4L4LL"--  '&r+c \P!VPVPPRR7P \P 4p\W PPVPR7#)zDProjects an index with the same index set onto the outer components.floor) rounding_moder) r#divrrrtyperarer r[)rQrrs&& r( project_outerProductIndexMap.project_outer!sW))EMM4+;+;+H+HX_`eefkfpfpq6F6F6S6S`e`p`pqqr+c P\\P!VPVPP 4P \P4P4P \P4VPP VPR7#)zDProjects an index with the same index set onto the inner components.r) rer#fmodrrrrrrrar[)rQrs&&r( project_innerProductIndexMap.project_inner&shJJu}}d.>.>.K.KL T%++  UW T%** ))66''   r+)rr ) r-r.r/r0r1r>rrr3r4r}r~s@@r(rfrfs%'2r   r+rfcVPp\VPVPR4^8de\P !VVPVP VP4^,R44P VP44#VPR4PVP4p\P !WPV4#)a Gathers from *values* using the index map. For each element in the domain of the index map this operation looks up a value for that index in *values*. Two elements from the same segment always get assigned the same value. Args: values (`torch.Tensor` of shape (B1, ..., Bn, num_segments, V1, ...)): Tensor with segment values. index (`IndexMap` of shape (B1, ..., Bn, I1, ..., Ik)): IndexMap. name (`str`, *optional*, defaults to 'segmented_gather'): Name for the operation. Currently not used Returns: `tuple(torch.Tensor)`: Tensor of shape (B1, ..., Bn, I1, ..., Ik, V1, ...) with the gathered values. Nr]) rrIrr[r#rhrr_rbrc)rrrSrs&&& r(rhrh2s mmG 6<<((* +,q0||     LL a "  $w||~   ##B'..v||<||F$4$4g>>r+cZ\P!\P!\VP 4444p\P !^W P PR7VP ,pVPVP 44p\VP\VPP444FpVPR4pK W0P,p\VPR4VP V,^R7#)a: Flattens a batched index map (which is typically of shape batch_size, seq_length) to a 1d index map. This operation relabels the segments to keep batch elements distinct. The k-th batch element will have indices shifted by *num_segments* * (k - 1). The result is a tensor with *num_segments* multiplied by the number of elements in the batch. Args: index (`IndexMap`): IndexMap to flatten. name (`str`, *optional*, defaults to 'segmented_flatten'): Name for the operation. Currently not used Returns: (`IndexMap`): The flattened IndexMap. startendrYrr])r#prodtensorlistr r`rrYrrlr[rIrr_rbre)rrS batch_sizeoffsetrrs&& r(flattenr%Ss"ELLe.?.?.A)BCDJ\\z:L:L:S:S TW\WiWi iF [[**, -F 5##S););)=%> ?!!"%@}}$G GLL,5;M;MPZ;Zgh iir+c\P!V4'd VPMRp\P!V\PVR7p\ VP 44^8XgQh\P!WR7p\ VP 44^8XgQh\P!^WPR7p\P!\P!V\PVPR7VP^R7.^R7pVP4Uu.uFp\V4NK ppVPV4p\P!V\P!^.VR7.^R7pVPVP44p\WA\!VP 44^,R7#uupi)a} Constructs an index map equal to range(num_segments). Args: batch_shape (`torch.Size`): Batch shape num_segments (`int`): Number of segments name (`str`, *optional*, defaults to 'range_index_map'): Name for the operation. Currently not used Returns: (`IndexMap`): IndexMap of shape batch_shape with elements equal to range(num_segments). cpurWr\rrr)r# is_tensorrYrjrarIr_r`catrrbtolistrrrepeatrer") r rrSrYr new_tensorx new_shape multipless &&& r(range_index_mapr0psv%*OOL$A$A\ uF//5::fK {! "a '' '??L |  " #q (( (ll\*=*=G EJJ|?R?R SUaUkUkpqUkUrs J ",!2!2!45!4AQ!4I5ll9%G ;F(KLRSTInnY--/0G G4P[P`P`PbKcdeKf gg6s6Gc \V4pVP4\VPP44Rp\P !\P !R.\PR7\P !V\PR7.^R7pVPVP44p\P!\VP4\PVPR7pVP^VPP4VP4VRR7p VPPp \P !\P !VP!4\PV R7\P !VP\PV R7P#^R7\P !V\PV R7.^R7p V P%4P'V P44P)VP*4p \-VP!4VP4p W3#)a Applies a segment reduction segment-wise. Args: values (`torch.Tensor`): Tensor with segment values. index (`IndexMap`): IndexMap. segment_reduce_fn (`str`): Name for the reduce operation. One of "sum", "mean", "max" or "min". name (`str`): Name for the operation. Currently not used Returns: (`IndexMap`): IndexMap of shape batch_shape with elements equal to range(num_segments). NrrrWF)rrsrcreduce include_selfr])r%r_rIrr#r)rjrareshaper*rkrrrrYscatter_reducer rbclonerrrXr0)rrsegment_reduce_fnrS flat_index vector_shapeflattened_shape flat_valuesout segment_meansrYr. output_values output_indexs&&&& r(_segment_reducerAs(J;;=U]]%7%7%9!:!<=Lii "UZZ 0%//,V[V`V`2abhiO..!7!7!9:K ++c*112%++kN`N` aC&& Z'',,.K4E4E4GPapu'M   & &F OOE--/uzz& Q OOE..ejj P Z Z_` Z a OOL 6 J  I"'')..y/?/?/ABEEfllSM"5#4#4#68J8JKL  &&r+c\WRV4#)a: Sums a tensor over its segments. Outputs 0 for empty segments. This operations computes the sum over segments, with support for: - Batching using the first dimensions [B1, B2, ..., Bn]. Each element in a batch can have different indices. - Vectorization using the last dimension [V1, V2, ...]. If they are present, the output will be a sum of vectors rather than scalars. Only the middle dimensions [I1, ..., Ik] are reduced by the operation. Args: values (`torch.Tensor` of shape [B1, B2, ..., Bn, I1, .., Ik, V1, V2, ..]): Tensor containing the values of which the sum must be taken segment-wise. index (`IndexMap`, indices are of shape [B1, B2, ..., Bn, I1, .., Ik].): Index defining the segments. name (`str`, *optional*, defaults to 'segmented_reduce_sum'): Name for the operation. Currently not used Returns: output_values (`torch.Tensor`of shape [B1, B2, ..., Bn, num_segments, V1, V2, ..]): Tensor containing the output values. output_index (`IndexMap`): IndexMap with shape [B1, B2, ..., Bn, num_segments]. . rrArrrSs&&&r( reduce_sumrEs0 6% 66r+c\WRV4#)a@ Averages a tensor over its segments. Outputs 0 for empty segments. This operations computes the mean over segments, with support for: - Batching using the first dimensions [B1, B2, ..., Bn]. Each element in a batch can have different indices. - Vectorization using the last dimension [V1, V2, ...]. If they are present, the output will be a mean of vectors rather than scalars. Only the middle dimensions [I1, ..., Ik] are reduced by the operation. Args: values (`torch.Tensor` of shape [B1, B2, ..., Bn, I1, .., Ik, V1, V2, ..]): Tensor containing the values of which the mean must be taken segment-wise. index (`IndexMap`, indices are of shape [B1, B2, ..., Bn, I1, .., Ik].): Index defining the segments. name (`str`, *optional*, defaults to 'segmented_reduce_sum'): Name for the operation. Currently not used Returns: output_values (`torch.Tensor`of shape [B1, B2, ..., Bn, num_segments, V1, V2, ..]): Tensor containing the output values. output_index (`IndexMap`): IndexMap with shape [B1, B2, ..., Bn, num_segments]. rrCrDs&&&r(rrs4 6&$ 77r+c\WRV4#)a2 Computes the maximum over segments. This operation computes the maximum over segments, with support for: - Batching using the first dimensions [B1, B2, ..., Bn]. Each element in a batch can have different indices. - Vectorization using the last dimension [V1, V2, ...]. If they are present, the output will be an element-wise maximum of vectors rather than scalars. Only the middle dimensions [I1, ..., Ik] are reduced by the operation. Args: values (`torch.Tensor` of shape [B1, B2, ..., Bn, I1, .., Ik, V1, V2, ..]): Tensor containing the values of which the max must be taken segment-wise. index (`IndexMap`, indices are of shape [B1, B2, ..., Bn, I1, .., Ik].): Index defining the segments. name (`str`, *optional*, defaults to 'segmented_reduce_sum'): Name for the operation. Currently not used Returns: output_values (`torch.Tensor`of shape [B1, B2, ..., Bn, num_segments, V1, V2, ..]): Tensor containing the output values. output_index (`IndexMap`): IndexMap with shape [B1, B2, ..., Bn, num_segments]. amaxrCrDs&&&r( reduce_maxrI0 6&$ 77r+c\WRV4#)a3 Computes the minimum over segments. This operations computes the minimum over segments, with support for: - Batching using the first dimensions [B1, B2, ..., Bn]. Each element in a batch can have different indices. - Vectorization using the last dimension [V1, V2, ...]. If they are present, the output will be an element-wise minimum of vectors rather than scalars. Only the middle dimensions [I1, ..., Ik] are reduced by the operation. Args: values (`torch.Tensor` of shape [B1, B2, ..., Bn, I1, .., Ik, V1, V2, ..]): Tensor containing the values of which the min must be taken segment-wise. index (`IndexMap`, indices are of shape [B1, B2, ..., Bn, I1, .., Ik].): Index defining the segments. name (`str`, *optional*, defaults to 'segmented_reduce_sum'): Name for the operation. Currently not used Returns: output_values (`torch.Tensor`of shape [B1, B2, ..., Bn, num_segments, V1, V2, ..]): Tensor containing the output values. output_index (`IndexMap`): IndexMap with shape [B1, B2, ..., Bn, num_segments]. aminrCrDs&&&r(rgrg rJr+c\P!RW4V,p\Wc4wrxVPV4p \ Wt,V 4wr\ WI4wrW\ ,,p \P !V R8\P!V P^4(4pV \\P!V\PVPR7,, p V'gmV \\P!\P!V P^4\PV PPR7,, p V #)a Computes the column logits. Args: sequence_output (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`): Also known as last_hidden_state. Sequence of hidden-states at the output of the last layer of the model. column_output_weights (`torch.FloatTensor` of shape `(hidden_size)`): Weights of the linear layer for column selection. column_output_bias (`torch.FloatTensor` of shape `()`): Bias of the linear layer for column selection. cell_index (`ProductIndexMap`): Index that groups tokens into cells. cell_mask (`torch.FloatTensor` of shape `(batch_size, max_num_rows * max_num_cols)`): Mask for cells that exist in the table (i.e. that are not padding). allow_empty_column_selection (`bool`): Whether to allow not to select any column Returns: column_logits (`torch.FloatTensor`of shape `(batch_size, max_num_cols)`): Tensor containing the column logits for every example in the batch. bsj,j->bs?rW) r#einsumrrrEr logical_andeqrCLOSE_ENOUGH_TO_LOG_ZEROrjrrY)rNr_rcrrr token_logits cell_logitscell_logits_index column_indexr out_index cell_countr is_paddings&&&&&& r(rr>s4<< _TWiiL&1%J"K++,=>L)+*A<PMy7MJ"777M"":#3ehhy?P?PRS6T5TUJ-%-- 0A0A1M (1EOO HHY&& *%-- HYHYH`H`5    r+c \\P!V\PVPR7V4wrg\P !VRR7p\P !\P!VRR7^,^4p \P!V PVP44\P!V4V4p\PPVR7p V PV4)p \W4wr\!\P!V\P"VPR7V4wrVP%V4P&p\P!\P !V\P(!VRR74\PVPR7p\PP+V R7pVPV P-\P44p\P.!VV,V,^R7)pV\P.!VV,^R7\0,,pT pV\P!V PVP44\P!V4V4, p\P!\P !VRR7\P"VPR7p\P!\P !V\P(!VRR74\PVPR7p\P!\P !V^4PVP44\P!V4V4pV \2RVV,, ,,p\5VV4pVV3#)aR Computes the loss for cell selection constrained to a single column. The loss is a hierarchical log-likelihood. The model first predicts a column and then selects cells within that column (conditioned on the column). Cells outside the selected column are never selected. Args: token_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length)`): Tensor containing the logits per token. column_logits (`torch.FloatTensor` of shape `(batch_size, max_num_cols)`): Tensor containing the logits per column. labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Labels per token. cell_index (`ProductIndexMap`): Index that groups tokens into cells. col_index (`IndexMap`): Index that groups tokens into columns. cell_mask (`torch.FloatTensor` of shape `(batch_size, max_num_rows * max_num_cols)`): Mask for cells that exist in the table (i.e. that are not padding). Returns: selection_loss_per_example (`torch.FloatTensor` of shape `(batch_size,)`): Loss for each example. logits (`torch.FloatTensor` of shape `(batch_size, sequence_length)`): New logits which are only allowed to select cells in a single column. Logits outside of the most likely column according to *column_logits* will be set to a very low value (such that the probabilities are 0). rWrrrr])rEr#rjrrYargmaxrRmaxrrr_rr CategoricalrrrIrarrrbrrrrrSrh)rTrrrrtrlabels_per_columnr column_labelno_cell_selected column_distcolumn_loss_per_examplerlabels_per_cell labels_indexcolumn_id_for_cells column_mask cell_dist cell_log_prob cell_lossrselected_column_idselected_column_masknew_logits_per_cellrs&&&&&& r(rrrs:&eoofEMMZ`ZgZg&hjst<< 1r:Lxx #,Q/;;l//12E4D4D\4RT`L%%111GK*33LAA %\>O$. ejjG%!O %22<@HH// $eool&KLmmK##--_-EI&&';';EMM'JKM=;6BJJI;2:=RRRI!8%++8==?@ 34# ]+5::mFZFZ !?? $eoo6Hb&QRmm!((!;; $a(--.B.G.G.IJ -. *,DiZnNnHn,oo ' 4F %v --r+cP\P!RW4V,V, pV#)a Computes logits per token Args: sequence_output (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`): Also known as last_hidden_state. Sequence of hidden-states at the output of the last layer of the model. temperature (`float`): Temperature for the Bernoulli distribution. output_weights (`torch.FloatTensor` of shape `(hidden_size,)`): Weights of the linear layer for cell selection. output_bias (`torch.FloatTensor` of shape `()`): Bias of the linear layer for cell selection Returns: logits (`torch.FloatTensor` of shape `(batch_size, sequence_length)`): Logits per token. rN)r#rP)rNrr^rbrs&&&& r(rrs$"ll;H;VZe eF Mr+c\P!\P!V44P\P4P VP 4pV!V4p\PPPVR7p\P!VPR,^R7pW8*p \P!V^R7^8p \P!\P!W4PVP44\P !V\P"R7V4p V P%4p V #)az Finds examples where the model should select cells with no aggregation. Returns a mask that determines for which examples should the model select answers directly from the table, without any aggregation function. If the answer is a piece of text the case is unambiguous as aggregation functions only apply to numbers. If the answer is a number but does not appear in the table then we must use some aggregation case. The ambiguous case is when the answer is a number that also appears in the table. In this case we use the aggregation function probabilities predicted by the model to decide whether to select or aggregate. The threshold for this is a hyperparameter *cell_selection_preference* Args: answer (`torch.FloatTensor` of shape `(batch_size, )`): Answer for every example in the batch. Nan if there is no scalar answer. pooled_output (`torch.FloatTensor` of shape `(batch_size, hidden_size)`): Output of the pooler (BertPooler) on top of the encoder layer. cell_selection_preference (`float`): Preference for cell selection in ambiguous cases. labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Labels per token. aggregation_classifier (`torch.nn.Linear`): Aggregation head Returns: aggregate_mask (`torch.FloatTensor` of shape `(batch_size,)`): A mask set to 1 for examples that should use aggregation functions. rrrrZr)r# logical_notisnanrr$rrYr categoricalr^rprobsrrQrr_rrdetach) answerr3rrraggregate_mask_initr dist_aggregationaggregation_ops_total_massis_pred_cell_selectionis_cell_supervision_availablers &&&&& r(rrs 4 ++EKK,?@EEeFWFWX[[\b\i\ij/ >**66BBJ\B]!&+;+A+A%+Ha!P8T%*IIf!$7IJJ33r+c\PPPVR7p\P!VP R,^R7p\P !V4)V,#)a Calculates aggregation loss in the case of answer supervision. Args: logits_aggregation (`torch.FloatTensor` of shape `(batch_size, num_aggregation_labels)`): Logits per aggregation operation. aggregate_mask (`torch.FloatTensor` of shape `(batch_size, )`): A mask set to 1 for examples that should use aggregation functions Returns: aggregation_loss_unknown (`torch.FloatTensor` of shape `(batch_size,)`): Aggregation loss (in case of answer supervision) per example. rrrp)r#rrsr^rrtlog)r rrxrys&& r(#_calculate_aggregation_loss_unknownrRs\**66BBJ\B]!&+;+A+A%+Ha!P II0 1 1N BBr+c`\WW#V4pV'dV\W4, pWV,#)a Calculates the aggregation loss per example. Args: logits_aggregation (`torch.FloatTensor` of shape `(batch_size, num_aggregation_labels)`): Logits per aggregation operation. aggregate_mask (`torch.FloatTensor` of shape `(batch_size, )`): A mask set to 1 for examples that should use aggregation functions. aggregation_labels (`torch.LongTensor` of shape `(batch_size, )`): Aggregation function id for every example in the batch. use_answer_as_supervision (`bool`, *optional*): Whether to use the answer as the only supervision for aggregation examples. num_aggregation_labels (`int`, *optional*, defaults to 0): The number of aggregation operators to predict. aggregation_loss_weight (`float`, *optional*, defaults to 1.0): Importance weight for the aggregation loss. Returns: aggregation_loss (`torch.FloatTensor` of shape `(batch_size,)`): Aggregation loss per example. )rr)r rrrrrper_example_aggregation_losss&&&&&& r(rrjs98$E,>[q$ !$(KL^(oo$ " AAr+c VP'dX\PPVPVP VP,R7pVP 4pM VPpWr, V,p\P!V^R7p\P!\P!V4\P!V4V4p \P!Wy,^R7p VPp V \P8XdW\,, p EM7V \P 8XdN\P!V^RR7V, ^,p \P!W,V , ^R7p MV \P"8Xd\P!V^RR7V, ^,p V^V, ,p\P!V^RR7V, pV\P$!V 4, ^,V , p\P!W,V,^R7p M\'RVP 24hVP('dD\PP+VP,VR,R7pVP 4pM9\.P0P3VR,VP,, RR7p\P4!\P6!V ^R7\P6!V ^R7\P6!V^R7.^R7p\P!VV,^R7pV#) a Calculates the expected result given cell and aggregation probabilities. Args: dist_per_cell (`torch.distributions.Bernoulli`): Cell selection distribution for each cell. numeric_values (`torch.FloatTensor` of shape `(batch_size, seq_length)`): Numeric values of every token. Nan for tokens which are not numeric values. numeric_values_scale (`torch.FloatTensor` of shape `(batch_size, seq_length)`): Scale of the numeric values of every token. input_mask_float (`torch.FloatTensor` of shape `(batch_size, seq_length)`): Mask for the table, without question tokens and table headers. logits_aggregation (`torch.FloatTensor` of shape `(batch_size, num_aggregation_labels)`): Logits per aggregation operation. config ([`TapasConfig`]): Model configuration class with all the hyperparameters of the model Returns: expected_result (`torch.FloatTensor` of shape `(batch_size,)`): The expected result per example. )rrrT)rkeepdimz(Invalid average_approximation_function: rrpr])use_gumbel_for_cellsr#rRelaxedBernoullirrsamplertrrrrraverage_approximation_functionrrrrrsquareruse_gumbel_for_aggregationRelaxedOneHotCategoricalaggregation_temperaturerrrr)rb) dist_per_cellrrrr rP gumbel_distscaled_probability_per_cell count_resultnumeric_values_masked sum_resultavg_approximationaverage_resultex pointwise_varvar multiplieraggregation_op_only_probs all_resultsexpected_results&&&&&& r(_calculate_expected_resultrs."""))::** ''&*<*<< ; '2&8&8&:#&3&9&9#$?#UYi"i998a@L!KK N#U%5%5n%E~6NTUVJ==8>>>#6K'KL :FF F YY24 HKf fij j#8#VY[#[abc :GG G YY24 HKf fij j3q;V7VW ii 1d;mKELL,,q0B6 #8#VYc#cijkCFDiDiCjkll ((())BB  * *3Ee3LC %0$6$6$8!%'MM$9$9 u %(F(F FB%:% !)) OOJA . OON 2 OOLa 0  Kii .G GQOO r+c$V^8dQhR\/#)rdelta)r)r&s"r(r)r)s[[U[r+c\P!W, 4p\P!W28RV^,,W2,RV^,,, 4#)rO)r#absr)inputtargetrerrorss&&& r( huber_lossrsA YYu~ &F ;;v~sVQY#PUWXPX.8Y ZZr+c \W#WEWg4p\P!\P!V4\P!V4V4p VP 'd\P !\P!V4\P!V 44\,P4p W, p W, p \W,W,4p M$\W,W,VPR7p VPf(\P!V \PR7pMl\P!WP8\P!V \PR7\P!V \PR74pVPW,,pW3#)a Calculates the regression loss per example. Args: answer (`torch.FloatTensor` of shape `(batch_size,)`): Answer for every example in the batch. Nan if there is no scalar answer. aggregate_mask (`torch.FloatTensor` of shape `(batch_size,)`): A mask set to 1 for examples that should use aggregation functions. dist_per_cell (`torch.distributions.Bernoulli`): Cell selection distribution for each cell. numeric_values (`torch.FloatTensor` of shape `(batch_size, seq_length)`): Numeric values of every token. Nan for tokens which are not numeric values. numeric_values_scale (`torch.FloatTensor` of shape `(batch_size, seq_length)`): Scale of the numeric values of every token. input_mask_float (`torch.FloatTensor` of shape `(batch_size, seq_length)`): Mask for the table, without question tokens and table headers. logits_aggregation (`torch.FloatTensor` of shape `(batch_size, num_aggregation_labels)`): Logits per aggregation operation. config ([`TapasConfig`]): Model configuration class with all the parameters of the model Returns: per_example_answer_loss_scaled (`torch.FloatTensor` of shape `(batch_size,)`): Scales answer loss for each example in the batch. large_answer_loss_mask (`torch.FloatTensor` of shape `(batch_size,)`): A mask which is 1 for examples for which their answer loss is larger than the answer_loss_cutoff. )rr)rr#rrrruse_normalized_answer_lossr]rrrurhuber_loss_deltaanswer_loss_cutoffrranswer_loss_importance)rvrrrrrr rPr answer_masked normalizernormalized_answer_maskednormalized_expected_resultper_example_answer_lossrper_example_answer_loss_scaleds&&&&&&&& r(rrsSJ1';OaO KK F 3U5E5Ef5MvVM (((ii / :EIImrs ! AA&!CC69uu-699,   H %#  7;7  7$JbiiJZVVt bii RYY:   ")) ?+?D0 2990 h "))  299$ BII "!ryy!3?33,R %R R js +s s lI 4I I X  y %9y  y x"3 "664* h* Z?Bj:)hX.'b768:868<1hh.V,.b+4\C0#BLTp[ DBN r+