o ZhX:@sdZddlmZddlZddlmZddlmZmZm 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,0e1Z2dZ3dZ4Gddde j5j6Z7GdddZ8ddZ9Gddde j5j6Z:Gddde j5j6Z;Gddde j5j6ZGd"d#d#e j5j6Z?e!Gd$d%d%e j5j6Z@e!Gd&d'd'e j5j6ZAGd(d)d)e j5j6ZBGd*d+d+e j5j6ZCGd,d-d-e j5j6ZDGd.d/d/eZEeGd0d1d1e(ZFd2ZGd3ZHe*d4eGGd5d6d6eEZIe*d7eGGd8d9d9eEZJe*d:eGGd;d<dd?d?eEeZLe*d@eGGdAdBdBeEeZMe*dCeGGdDdEdEeEeZNe*dFeGGdGdHdHeEeZOe*dIeGGdJdKdKeEeZPgdLZQdS)NzTF 2.0 Funnel model.) annotationsN) dataclass)OptionalTupleUnion)get_tf_activation)TFBaseModelOutputTFMaskedLMOutputTFMultipleChoiceModelOutputTFQuestionAnsweringModelOutputTFSequenceClassifierOutputTFTokenClassifierOutput) TFMaskedLanguageModelingLossTFModelInputTypeTFMultipleChoiceLossTFPreTrainedModelTFQuestionAnsweringLossTFSequenceClassificationLossTFTokenClassificationLossget_initializerkeraskeras_serializable unpack_inputs)check_embeddings_within_bounds shape_liststable_softmax) ModelOutputadd_code_sample_docstringsadd_start_docstrings%add_start_docstrings_to_model_forwardloggingreplace_return_docstrings) FunnelConfigr$g.Acs4eZdZdZfddZd ddZd dd ZZS) TFFunnelEmbeddingszGConstruct the embeddings from word, position and token_type embeddings.c s`tjdi|||_|j|_|jdurdn|j|_tjj|jdd|_ tjj |j d|_ dS)N? layer_normepsilonname)Zrate) super__init__config hidden_sizeinitializer_stdrlayersLayerNormalizationlayer_norm_eps LayerNormDropouthidden_dropoutdropoutselfr.kwargs __class__r+\/var/www/auris/lib/python3.10/site-packages/transformers/models/funnel/modeling_tf_funnel.pyr-Gs zTFFunnelEmbeddings.__init__NcCstd|jd|jj|jgt|jdd|_Wdn1s"wY|j r,dSd|_ t |dddur\t|j j |j dd|jjgWddS1sUwYdSdS)NZword_embeddingsweight)initializer_range)r*shape initializerTr4)tf name_scope add_weightr. vocab_sizer/rr0r>builtgetattrr4r*buildd_modelr9 input_shaper+r+r=rHQs    "zTFFunnelEmbeddings.buildFcCsj|dur |dur J|dur|durJ|dur&t||jjt|j|}|j|d}|j||d}|S)z Applies embedding based on inputs tensor. Returns: final_embeddings (`tf.Tensor`): output embedding tensor. N)inputs)rLtraining)rr.rErBgatherr>r4r7)r9 input_ids inputs_embedsrMZfinal_embeddingsr+r+r=call`s zTFFunnelEmbeddings.callNNNF)__name__ __module__ __qualname____doc__r-rHrQ __classcell__r+r+r;r=r%Ds   r%c@sveZdZUdZdZded<ddZdd d Zd d Zdd dZ ddZ dddZ ddZ d ddZ ddZddZdS)!TFFunnelAttentionStructurez@ Contains helpers for `TFFunnelRelMultiheadAttention `. intcls_token_type_idcCsb|j|_|j|_|j|_|j|_|j|_|j|_|j|_tj |j |_ tj |j |_ d|_ dSrR)rIattention_type num_blocks separate_cls truncate_seq pool_q_only pooling_typerr1r5r6 sin_dropout cos_dropout pooling_mult)r9r.r+r+r=r-{s z#TFFunnelAttentionStructure.__init__NFc Csd|_t|d|_}|j||d}|dur||nd}|jr9ttj|d|dg|j dddgddggnd}||||fS)zCReturns the attention inputs associated to the inputs of the model.r#rMNdtyper) rerseq_lenget_position_embedstoken_type_ids_to_matr_rBpadonesrh) r9rPattention_masktoken_type_idsrMriposition_embedstoken_type_matcls_maskr+r+r=init_attention_inputss2 z0TFFunnelAttentionStructure.init_attention_inputscCs`tt|dt|d}t|tj|jg|jd}tt|dt|d}t||S)z-Convert `token_type_ids` to `token_type_mat`.rg)rBequalZ expand_dimsconstantr\rh logical_or)r9rorqZcls_idsZcls_matr+r+r=rks z0TFFunnelAttentionStructure.token_type_ids_to_matcCsh|jdkrktd|d}td|jdd}dd||jd}td||}t|}|j||d}t|} |j| |d} tj ||gd d } tj | |gd d } tj | | gd d } tj | | gd d }| | | |fStd|jdd}dd||jd}t| d|dd}|t d}td||}|jt||d}|jt||d} tj || gd d }td|}|}g}td|j D]j}tj dgd d }|dkr| ||}d|d}|j|||dd }tj||jd}||}tj||dd }|}d|}|||}tj||jd}||}tj|t|dtj||dd }|||gq|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&rZr#i'zi,d->idrfrtaxisg)value)shiftrg)r]rBrangerIeinsumsinrccosrdconcatrwr^fillstride_pool_pos relative_poscastrhrN debuggingZ assert_lessr@append)r9rirMZpos_seqZfreq_seqZinv_freqZsinusoidZ sin_embedZ sin_embed_dZ cos_embedZ cos_embed_dphipsipiomegaZ rel_pos_idZ zero_offsetZ pos_embedpos pooled_posZposition_embeds_list block_indexZposition_embeds_poolingstrideZrel_posZposition_embeds_no_poolingr+r+r=rjsV        z.TFFunnelAttentionStructure.get_position_embedscCsh|jr-tjd| dg|jd}|jr|ddn|dd}t||dddgdS|dddS)zc Pool `pos_id` while keeping the cls token separate (if `self.separate_cls=True`). rZr#rgrtNr)r_rBrwrhr`r)r9Zpos_idrZcls_posZ pooled_pos_idr+r+r=rs z*TFFunnelAttentionStructure.stride_pool_posr#c Cs\|dur|}|d|d}|t|d}|||}|d|d}t||d| S)zV Build the relative positional vector between `pos` and `pooled_pos`. Nrrtr#)rrBr~) r9rrrr}Z ref_pointZ num_removeZmax_distZmin_distr+r+r=rs z'TFFunnelAttentionStructure.relative_poscs|durdStttfrD]}||}q|St|ttfr/t|fdd|DStt|;jrCjrCt dddnt ddd}t dg|g}jrmt dgt ddg}t |||g}||S)zT Perform pooling by stride slicing the tensor along the given axis. Nc3s|] }|VqdSrR) stride_pool.0xr{r9r+r= z9TFFunnelAttentionStructure.stride_pool..rtrZr#) isinstancelisttuplertypelenrr_r`slicerBr)r9tensorr{axZ axis_sliceZ enc_sliceZ cls_slicer+rr=rs$z&TFFunnelAttentionStructure.stride_poolmeancs.durdStttfrtfddDSjr@jr-ddddfn}tjddddf|gddtt }|dkrUdddddfdkretj j d d d n&d krutj j d d d nd krtj j  d d d  nt d|dkrtdSS)z3Apply 1D pooling to a tensor of size [B x T (x H)].Nc3s |] }jdVqdS))moderN) pool_tensorrrr9rrr+r=r3sz9TFFunnelAttentionStructure.pool_tensor..rtr#rzrZrZNWCZSAME)stridesZ data_formatpaddingmaxminz0The supported modes are 'mean', 'max' and 'min'.)rrrrr_r`rBrrrnnZ avg_pool1dZ max_pool1dNotImplementedErrorsqueeze)r9rrrsuffixndimr+rr=r,s$ " z&TFFunnelAttentionStructure.pool_tensorcCs|\}}}}|jr3|jdkr||ddd|dd}||d}||d}|j||jd}n1|jd9_|jdkrE||d}||ddg}||ddg}|j|dd}|j||jd}||||f}||fS)zTPool `output` and the proper parts of `attention_inputs` before the attention layer.ryNrZrr#rr)rar]rrrbre)r9outputattention_inputsrprqrnrrr+r+r=pre_attention_poolingHs        z0TFFunnelAttentionStructure.pre_attention_poolingcCs|\}}}}|jr8|jd9_|jdkr%|dd||ddd}||d}||d}|j|dd}||||f}|S)zFPool the proper parts of `attention_inputs` after the attention layer.rZryNrr#rr)rarer]rr)r9rrprqrnrrr+r+r=post_attention_pooling\s      z1TFFunnelAttentionStructure.post_attention_poolingrSF)Nr#)rrZ)rTrUrVrWr\__annotations__r-rsrkrjrrrrrrr+r+r+r=rYts    S   rYcCspt|\}}}}t|||||g}|dddd|dddf}t||||||g}|dd|f}|S)N.)rrBreshape)positional_attn context_lenr} batch_sizen_headriZ max_rel_lenr+r+r=_relative_shift_gatherjs  rcsDeZdZfddZd ddZd ddZd dd Zdd d ZZS)TFFunnelRelMultiheadAttentionc stjd i||j|_|j|_}|j|_}|j|_}|j|_||_tj |j |_ tj |j |_ t |j}tj j||d|dd|_tj j|||dd|_tj j|||dd|_tj j||dd|_tj j|jdd |_d |d |_dS) NFq_head)Zuse_biaskernel_initializerr*k_headrr*v_head post_projr'r(r&g?r+)r,r-r]rd_headrIr?rrr1r5r6attention_dropoutrDenserrrrr2r3r'scale)r9r.rr:rrrIrAr;r+r=r-{s$     z&TFFunnelRelMultiheadAttention.__init__NcCsN|j|j|j}}}t|j}|j||f|ddd|_|j||f|ddd|_|j|||f|ddd|_|j||f|ddd|_ |jd||f|ddd|_ |j rSdSd|_ t |d ddur~t |jj|jdd|gWdn1sywYt |d ddurt |jj|jdd|gWdn1swYt |d ddurt |jj|jdd|gWdn1swYt |d ddurt |jj|jdd||gWdn1swYt |d ddur%t |jj|jdd|gWddS1swYdSdS)NTr_w_biasr@rAZ trainabler*r_r_biasr_kernelr_s_biasrZ seg_embedrrrrr')rrrIrr?rDrrrrrrFrGrBrCrr*rHrrrr')r9rKrrrIrAr+r+r=rHsR    $z#TFFunnelRelMultiheadAttention.buildcCs|jdkr@|\}}}}|j|j} |j} td|| | } | |dddf} | |dddf} td| |td| |}n:t|d|krRd}||jd}n d}||jd}|j|j}|j} td|| }td |||}t|||}|dur||9}|S) z5Relative attention score for the positional encodingsryzbinh,dnh->bindNz bind,jd->bnijr#rZrz td,dnh->tnhzbinh,tnh->bnit) r]rrrrBrrrr)r9rprrrrrrrruZw_rZ q_r_attentionZq_r_attention_1Zq_r_attention_2rr}rvZr_headr+r+r=relative_positional_attentions.     z;TFFunnelRelMultiheadAttention.relative_positional_attentionc Cs|durdSt|\}}}|j|j}td|||j}t|dddfdt|dddg}tj|ddd\} } t|t| ddd|gt| ddd|g} |durZ| |9} | S)z/Relative attention score for the token_type_idsNrzbind,snd->bnisr#rZrtrz) rrrrBrrZtilesplitwhere) r9rqrrrrrirrZtoken_type_biasZdiff_token_typeZsame_token_typetoken_type_attnr+r+r=relative_token_type_attentions (z;TFFunnelRelMultiheadAttention.relative_token_type_attentionFc Cs||\}}} } t|\} } } t|d}|j|j}}t||| | ||g}t||| |||g}t||| |||g}||j}|j |j}t d|||}| |||| }| ||| }|||}| durtj | |jd} |td| ddddf}t|dd}|j||d}t d||}|t|| | ||g}|j||d}|||}|r||fS|fS)Nr#zbind,bjnd->bnijrgrtrzrfzbnij,bjnd->bind)rrrrBrrrrrrrrrrrhINFrrrr6r')r9querykeyr|routput_attentionsrMrprqrnrrrri_rrrrrrrZ content_scorerrZ attn_scoreZ attn_probZattn_vecZattn_outrr+r+r=rQs.      z"TFFunnelRelMultiheadAttention.callrRFF) rTrUrVr-rHrrrQrXr+r+r;r=rzs    ' .rc0eZdZfddZd ddZd ddZZS) TFFunnelPositionwiseFFNc stjdi|t|j}tjj|j|dd|_t |j |_ tj |j |_ tjj|j|dd|_tj |j|_tjj|jdd|_||_dS)Nlinear_1rlinear_2r'r(r+)r,r-rr?rr1rd_innerrr hidden_actactivation_functionr5activation_dropoutrIrr6r7r2r3r'r.r9r.r:rAr;r+r=r-2s   z TFFunnelPositionwiseFFN.__init__FcCsH||}||}|j||d}||}|j||d}|||SNrf)rrrrr7r')r9hiddenrMhr+r+r=rQ=s   zTFFunnelPositionwiseFFN.callNcCs|jrdSd|_t|dddur2t|jj|jdd|jjgWdn1s-wYt|dddur\t|j j|j dd|jj gWdn1sWwYt|dddurt|j j|j dd|jjgWddS1swYdSdS)NTrrr') rFrGrBrCrr*rHr.rIrrr'rJr+r+r=rHEs "zTFFunnelPositionwiseFFN.buildrrRrTrUrVr-rQrHrXr+r+r;r=r1s  rcr) TFFunnelLayerc s4tjdi|t||dd|_t|dd|_dS)N attentionr*ffnr+)r,r-rrrr)r9r.rr:r;r+r=r-UszTFFunnelLayer.__init__Fc Cs>|j||||||d}|j|d|d}|r||dfS|fS)NrrMrrfr#)rr) r9rrr|rrrMZattnrr+r+r=rQZs  zTFFunnelLayer.callNcC|jrdSd|_t|dddur-t|jj|jdWdn1s(wYt|dddurUt|jj|jdWddS1sNwYdSdS)NTrr)rFrGrBrCrr*rHrrJr+r+r=rHa"zTFFunnelLayer.buildrrRrr+r+r;r=rTs  rc<eZdZfddZ      d ddZd dd ZZS) TFFunnelEncoderc sRtjdi|j|_j|_j|_t|_fddtjD|_ dS)Ncs(g|]\}fddt|DqS)c s&g|]}tdd|dqS)z blocks_._z_._rrri)rr.r+r= us&z7TFFunnelEncoder.__init__...)r~)r block_sizer.)rr=rtsz,TFFunnelEncoder.__init__..r+) r,r-r_ra block_repeatsrYattention_structure enumerate block_sizesblocksr8r;rr=r-ns   zTFFunnelEncoder.__init__NFTc Csj|jj||||d}|} |r|fnd} |rdnd} t|jD]\} } t| d|jr,dndk}|o4| dk}tt| }|rG|j| |\}}t| D]S\}}t |j | D]G}|dkoa|dkoa|}|rp|}|j rk| n|}}n| }}}|||||||d}|d} |r|j |}|r| |dd} |r| | f} qVqKq|st dd| | | fDSt| | | d S) NrnrorMr+r#rZrrcs|] }|dur|VqdSrRr+rrr+r+r=rrz'TFFunnelEncoder.call..last_hidden_state hidden_states attentions)rrsrrrr_rBzerosrr~rrarrr )r9rPrnroroutput_hidden_states return_dictrMrrall_hidden_statesall_attentionsrblockZ pooling_flagZ pooled_hiddenZ layer_indexlayerZ repeat_indexZ do_poolingrrr| layer_outputr+r+r=rQysP       zTFFunnelEncoder.callc Cs`|jrdSd|_|jD]"}|D]}t|j |dWdn1s'wYqq dS)NT)rFrrBrCr*rH)r9rKrrr+r+r=rHs  zTFFunnelEncoder.buildNNFFTFrRrr+r+r;r=rms  9rTFcCs|dkr|S|r|ddddf}|ddddf}tj||dd}|rP|r9t|ddgd|dgddgg}|ddd|df}tj||gdd}|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)Zrepeatsr{rrz)rBrepeatrlr)rr target_lenr_r`clsrr+r+r=upsamples"r cr) TFFunnelDecoderc s^tjdi|j|_j|_dtjd|_t|_fddt j D|_ dS)NrZr#cs g|] }tdd|dqS)rz layers_._rrrrr+r=rs z,TFFunnelDecoder.__init__..r+) r,r-r_r`rrrrYrr~Znum_decoder_layersr1r8r;rr=r-s  zTFFunnelDecoder.__init__NFTc  Cst||jt|d|j|jd} | |} |r| fnd} |rdnd} |jj| |||d} |jD]!}|| | | | ||d}|d} |rH| |dd} |rO| | f} q.|s^tdd| | | fDSt | | | d S) Nr#)rr r_r`r+rrrcsrrRr+rr+r+r=rrz'TFFunnelDecoder.call..r) r rrr_r`rrsr1rr )r9 final_hiddenfirst_block_hiddenrnrorrrrMZupsampled_hiddenrrrrrrr+r+r=rQs:      zTFFunnelDecoder.callc Csj|jrdSd|_t|dddur1|jD]}t|j |dWdn1s+wYqdSdS)NTr1)rFrGr1rBrCr*rH)r9rKrr+r+r=rHs  zTFFunnelDecoder.buildr rRrr+r+r;r=rs   -rcdeZdZdZeZfddZddZddZdd Z e dd d Z dddZ Z S)TFFunnelBaseLayerzBase model without decoderc sPtjdi|||_|j|_|j|_|j|_t|dd|_t |dd|_ dS)N embeddingsrencoderr+) r,r-r.rruse_return_dictrr%rrrr8r;r+r=r-szTFFunnelBaseLayer.__init__cC|jSrRrr9r+r+r=get_input_embeddings!z&TFFunnelBaseLayer.get_input_embeddingscC||j_t|d|j_dSNrrr>rrEr9r|r+r+r=set_input_embeddings$z&TFFunnelBaseLayer.set_input_embeddingscCtrRrr9Zheads_to_pruner+r+r= _prune_heads(zTFFunnelBaseLayer._prune_headsNFc Cs|dur |dur td|durt|} n|dur"t|dd} ntd|dur0t| d}|dur:t| d}|durE|j||d}|j|||||||d} | S)NDYou cannot specify both input_ids and inputs_embeds at the same timert5You have to specify either input_ids or inputs_embedsr#rrfrnrorrrrM) ValueErrorrrBrrr) r9rOrnrorPrrrrMrKencoder_outputsr+r+r=rQ+s.     zTFFunnelBaseLayer.callcCr)NTrr)rFrGrBrCrr*rHrrJr+r+r=rHUrzTFFunnelBaseLayer.buildNNNNNNNFrRrTrUrVrWr$ config_classr-rrr$rrQrHrXr+r+r;r=rs$   )rcr)TFFunnelMainLayerzBase model with decoderc sftjdi|||_|j|_|j|_|j|_|j|_t|dd|_ t |dd|_ t |dd|_ dS)Nrrrdecoderr+)r,r-r.rrrrrr%rrrrr/r8r;r+r=r-gszTFFunnelMainLayer.__init__cCrrRrrr+r+r=rtrz&TFFunnelMainLayer.get_input_embeddingscCrrrrr+r+r=rwr z&TFFunnelMainLayer.set_input_embeddingscCr!rRr"r#r+r+r=r${r%zTFFunnelMainLayer._prune_headsNFc  Cs^|dur |dur td|durt|} n|dur"t|dd} ntd|dur0t| d}|dur:t| d}|durE|j||d}|j||||d||d} |j| d| d|jd||||||d } |sd} | df} |r| d7} | | d| | f} |r| d7} | | d | | f} | St| d|r| j | j nd|r| j | j d Sdd S) Nr&rtr'r#rrfTr()rrrnrorrrrMrZr) r)rrBrrrr/rr rr)r9rOrnrorPrrrrMrKr*Zdecoder_outputsidxoutputsr+r+r=rQ~sf       zTFFunnelMainLayer.callcCs|jrdSd|_t|dddur-t|jj|jdWdn1s(wYt|dddurRt|jj|jdWdn1sMwYt|dddurzt|jj|jdWddS1sswYdSdS)NTrrr/) rFrGrBrCrr*rHrr/rJr+r+r=rHs "zTFFunnelMainLayer.buildr+rRr,r+r+r;r=r.as$   Er.cs2eZdZdZfddZddZd ddZZS) TFFunnelDiscriminatorPredictionszEPrediction module for the discriminator, made up of two dense layers.c s\tjdi|t|j}tjj|j|dd|_t |j |_ tjjd|dd|_ ||_ dS)Ndenserr#dense_predictionr+)r,r-rr?rr1rrIr3rrrr4r.rr;r+r=r-s    z)TFFunnelDiscriminatorPredictions.__init__cCs(||}||}t||}|SrR)r3rrBrr4)r9discriminator_hidden_statesrlogitsr+r+r=rQs  z%TFFunnelDiscriminatorPredictions.callNcC|jrdSd|_t|dddur2t|jj|jdd|jjgWdn1s-wYt|dddur_t|j j|j dd|jjgWddS1sXwYdSdS)NTr3r4) rFrGrBrCr3r*rHr.rIr4rJr+r+r=rH"z&TFFunnelDiscriminatorPredictions.buildrR)rTrUrVrWr-rQrHrXr+r+r;r=r2s  r2csReZdZfddZfddZddZddZd d Zd d ZdddZ Z S)TFFunnelMaskedLMHeadc s*tjdi|||_|j|_||_dS)Nr+)r,r-r.r/input_embeddings)r9r.r:r:r;r+r=r-s zTFFunnelMaskedLMHead.__init__cs*|j|jjfdddd|_t|dS)NrTbiasr)rDr.rEr;r,rHrJr;r+r=rHszTFFunnelMaskedLMHead.buildcCrrR)r:rr+r+r=get_output_embeddingsrz*TFFunnelMaskedLMHead.get_output_embeddingscCrr)r:r>rrErr+r+r=set_output_embeddingsr z*TFFunnelMaskedLMHead.set_output_embeddingscCs d|jiS)Nr;)r;rr+r+r=get_biass zTFFunnelMaskedLMHead.get_biascCs"|d|_t|dd|j_dS)Nr;r)r;rr.rErr+r+r=set_biass zTFFunnelMaskedLMHead.set_biasFcCsdt|dd}tj|d|jgd}tj||jjdd}tj|d||jjgd}tj j ||j d}|S)N)rr#rt)rr@T)abZ transpose_b)r|r;) rrBrr/matmulr:r>r.rErZbias_addr;)r9rrM seq_lengthr+r+r=rQ s zTFFunnelMaskedLMHead.callr) rTrUrVr-rHr<r=r>r?rQrXr+r+r;r=r9s  r9cr) TFFunnelClassificationHeadc s`tjdi|t|j}tjj|j|dd|_tj |j |_ tjj||dd|_ ||_ dS)N linear_hiddenr linear_outr+)r,r-rr?rr1rrIrEr5r6r7rFr.)r9r.Zn_labelsr:rAr;r+r=r-s   z#TFFunnelClassificationHead.__init__FcCs.||}tj|}|j||d}||Sr)rErZ activationstanhr7rF)r9rrMr+r+r=rQs   zTFFunnelClassificationHead.callNcCr7)NTrErF) rFrGrBrCrEr*rHr.rIrFrJr+r+r=rH#r8z TFFunnelClassificationHead.buildrrRrr+r+r;r=rDs  rDc@s$eZdZdZeZdZeddZdS)TFFunnelPreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. funnelcCdtjdtjdiS)NrO)r#rrgrBrmZint32rr+r+r= dummy_inputs8z$TFFunnelPreTrainedModel.dummy_inputsN) rTrUrVrWr$r-Zbase_model_prefixpropertyrLr+r+r+r=rH/s rHc@s6eZdZUdZdZded<dZded<dZded<dS)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. NzOptional[tf.Tensor]r6zTuple[tf.Tensor] | Nonerr)rTrUrVrWr6rrrr+r+r+r=rO>s   rOaa 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. csleZdZdfdd Zeededee d e d d!ddZ ddZ d"ddZ ZS)#TFFunnelBaseModelr.r$returnNonec,tj|g|Ri|t|dd|_dSNrIr)r,r-rrIr9r.rLr:r;r+r=r-zTFFunnelBaseModel.__init__batch_size, sequence_lengthfunnel-transformer/small-base checkpoint output_typer-NFrOTFModelInputType | Nonernnp.ndarray | tf.Tensor | NonerorPrOptional[bool]rrrMbool*Union[Tuple[tf.Tensor], TFBaseModelOutput]c C|j||||||||dSN)rOrnrorPrrrrMrI r9rOrnrorPrrrrMr+r+r=rQzTFFunnelBaseModel.callcCt|j|j|jdSNrr rrrr9rr+r+r=serving_output z TFFunnelBaseModel.serving_outputcCd|jrdSd|_t|dddur0t|jj|jdWddS1s)wYdSdSNTrIrFrGrBrCrIr*rHrJr+r+r=rH"zTFFunnelBaseModel.buildr.r$rQrRr+rOr\rnr]ror]rPr]rr^rr^rr^rMr_rQr`rR)rTrUrVr-r FUNNEL_INPUTS_DOCSTRINGformatrr _CONFIG_FOR_DOCrrQrjrHrXr+r+r;r=rPs(  rPzlThe bare Funnel Transformer Model transformer outputting raw hidden-states without any specific head on top.csleZdZdfdd Zeeedede e d  d d!ddZ ddZ d"ddZ ZS)# TFFunnelModelr.r$rQrRcrSrT)r,r-r.rIrUr;r+r=r-rVzTFFunnelModel.__init__rWfunnel-transformer/smallrYNFrOr\rnr]rorPrr^rrrMr_r`c Crarbrcrdr+r+r=rQrezTFFunnelModel.callcCrfrgrhrir+r+r=rjrkzTFFunnelModel.serving_outputcCrlrmrnrJr+r+r=rHrozTFFunnelModel.buildrpr+rqrR)rTrUrVr-rr rrrsrr rtrQrjrHrXr+r+r;r=rus(  ruz| Funnel model with a binary classification head on top as used during pretraining for identifying generated tokens. csjeZdZdfdd Zeeedee e d dd ddZ ddZ d!ddZ ZS)"TFFunnelForPreTrainingr.r$rQrRc s4tj|fi|t|dd|_t|dd|_dS)NrIrdiscriminator_predictions)r,r-r.rIr2rxr8r;r+r=r-,szTFFunnelForPreTraining.__init__rW)r[r-NFrOr\rnr]rorPrr^rrrMr_5Union[Tuple[tf.Tensor], TFFunnelForPreTrainingOutput]c KsT|j||||||||d} | d} || } |s!| f| ddSt| | j| jdS)a Returns: Examples: ```python >>> from transformers import AutoTokenizer, TFFunnelForPreTraining >>> import torch >>> tokenizer = AutoTokenizer.from_pretrained("funnel-transformer/small") >>> model = TFFunnelForPreTraining.from_pretrained("funnel-transformer/small") >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf") >>> logits = model(inputs).logits ```rrMrr#Nr6rr)rIrxrOrr) r9rOrnrorPrrrrMr:r5Zdiscriminator_sequence_outputr6r+r+r=rQ2s& zTFFunnelForPreTraining.callcCrfNr{)rOr6rrrir+r+r=rjf z%TFFunnelForPreTraining.serving_outputcCr)NTrIrx)rFrGrBrCrIr*rHrxrJr+r+r=rHmrzTFFunnelForPreTraining.buildrpr+)rOr\rnr]ror]rPr]rr^rr^rr^rMr_rQryrR)rTrUrVr-rr rrrsr"rOrtrQrjrHrXr+r+r;r=rw%s   1rwz4Funnel Model with a `language modeling` head on top.cseZdZd(fdd Zd)dd Zd*d d Zeee d e de e d         d+d,d d!Z d-d$d%Zd.d&d'ZZS)/TFFunnelForMaskedLMr.r$rQrRcs@tj|g|Ri|t|dd|_t||jjdd|_dS)NrIrlm_head)r,r-r.rIr9rrrUr;r+r=r-{szTFFunnelForMaskedLM.__init__r9cCrrR)rrr+r+r= get_lm_headrzTFFunnelForMaskedLM.get_lm_headstrcCstdt|jd|jjS)NzMThe method get_prefix_bias_name is deprecated. Please use `get_bias` instead./)warningswarn FutureWarningr*rrr+r+r=get_prefix_bias_names z(TFFunnelForMaskedLM.get_prefix_bias_namerWrvrYNFrOr\rnr]rorPrr^rrlabelsrMr_)Union[Tuple[tf.Tensor], TFMaskedLMOutput]c  Cs|j|||||||| d} | d} |j| | d} |durdn||| } |s:| f| dd}| dur8| f|S|St| | | j| jdS)a labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]` rzrrfNr#lossr6rr)rIrhf_compute_lossr rr)r9rOrnrorPrrrrrMr1sequence_outputZprediction_scoresrrr+r+r=rQs, zTFFunnelForMaskedLM.callrr cCrfr|)r r6rrrir+r+r=rjrMz"TFFunnelForMaskedLM.serving_outputcCr)NTrIr)rFrGrBrCrIr*rHrrJr+r+r=rHrzTFFunnelForMaskedLM.buildrp)rQr9)rQr NNNNNNNNF)rOr\rnr]ror]rPr]rr^rr^rr^rr]rMr_rQr)rr rQr rR)rTrUrVr-rrrr rrrsrr rtrQrjrHrXr+r+r;r=r~ys.    ,r~z Funnel Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled output) e.g. for GLUE tasks. cpeZdZd"fdd Zeeedede e d  d#d$ddZ d%ddZ d&d d!Z ZS)'!TFFunnelForSequenceClassificationr.r$rQrRcsFtj|g|Ri||j|_t|dd|_t||jdd|_dS)NrIr classifier)r,r- num_labelsrrIrDrrUr;r+r=r-sz*TFFunnelForSequenceClassification.__init__rWrXrYNFrOr\rnr]rorPrr^rrrrMr_3Union[Tuple[tf.Tensor], TFSequenceClassifierOutput]c  Cs|j|||||||| d} | d} | dddf} |j| | d} |dur&dn||| }|sB| f| dd}|dur@|f|S|St|| | j| jdS)a labels (`tf.Tensor` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). rzrNrfr#r)rIrrr rr)r9rOrnrorPrrrrrMr1r pooled_outputr6rrr+r+r=rQs. z&TFFunnelForSequenceClassification.callrr cCrfr|)r r6rrrir+r+r=rjr}z0TFFunnelForSequenceClassification.serving_outputcCrNTrIrrFrGrBrCrIr*rHrrJr+r+r=rHrz'TFFunnelForSequenceClassification.buildrpr)rOr\rnr]ror]rPr]rr^rr^rr^rr]rMr_rQr)rr rQr rR)rTrUrVr-rr rrrsrr rtrQrjrHrXr+r+r;r=rs*  -rz Funnel Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a softmax) e.g. for RocStories/SWAG tasks. cs|eZdZd$fdd ZeddZeee d e d e e d  d%d&ddZ d'd d!Zd(d"d#ZZS))TFFunnelForMultipleChoicer.r$rQrRcs<tj|g|Ri|t|dd|_t|ddd|_dS)NrIrr#r)r,r-rrIrDrrUr;r+r=r-*sz"TFFunnelForMultipleChoice.__init__cCrJ)NrO)rrrgrKrr+r+r=rL0sz&TFFunnelForMultipleChoice.dummy_inputsz(batch_size, num_choices, sequence_lengthrXrYNFrOr\rnr]rorPrr^rrrrMr_4Union[Tuple[tf.Tensor], TFMultipleChoiceModelOutput]c  Cs\|durt|d} t|d} n t|d} t|d} |dur)t|d| fnd} |dur7t|d| fnd} |durEt|d| fnd}|durXt|d| t|dfnd}|j| | |||||| d}|d}|dddf}|j|| d}t|d| f}|durdn|||}|s|f|dd}|dur|f|S|St|||j|jd S) a5 labels (`tf.Tensor` of shape `(batch_size,)`, *optional*): Labels for computing the multiple choice classification loss. Indices should be in `[0, ..., num_choices]` where `num_choices` is the size of the second dimension of the input tensors. (See `input_ids` above) Nr#rZrtr)rnrorPrrrrMrrfr) rrBrrIrrr rr)r9rOrnrorPrrrrrMZ num_choicesrCZflat_input_idsZflat_attention_maskZflat_token_type_idsZflat_inputs_embedsr1rrr6Zreshaped_logitsrrr+r+r=rQ4sH    zTFFunnelForMultipleChoice.callrr cCrfr|)r r6rrrir+r+r=rjyr}z(TFFunnelForMultipleChoice.serving_outputcCrrrrJr+r+r=rHrzTFFunnelForMultipleChoice.buildrpr)rOr\rnr]ror]rPr]rr^rr^rr^rr]rMr_rQr)rr rQr rR)rTrUrVr-rNrLrr rrrsrr rtrQrjrHrXr+r+r;r=r"s.   >rz Funnel Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. cr)'TFFunnelForTokenClassificationr.r$rQrRcsftj|g|Ri||j|_t|dd|_tj|j|_ tjj |jt |j dd|_ ||_dS)NrIrrr)r,r-rr.rIrr1r5r6r7rrr?rr.rUr;r+r=r-s z'TFFunnelForTokenClassification.__init__rWrvrYNFrOr\rnr]rorPrr^rrrrMr_0Union[Tuple[tf.Tensor], TFTokenClassifierOutput]c  Cs|j|||||||| d} | d} |j| | d} || } |dur#dn||| } |s?| f| dd}| dur=| f|S|St| | | j| jdS)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]`. rzrrfNr#r)rIr7rrrrr)r9rOrnrorPrrrrrMr1rr6rrr+r+r=rQs.  z#TFFunnelForTokenClassification.callrrcCrfr|)rr6rrrir+r+r=rjr}z-TFFunnelForTokenClassification.serving_outputcC|jrdSd|_t|dddur-t|jj|jdWdn1s(wYt|dddurZt|jj|jdd|jj gWddS1sSwYdSdSr) rFrGrBrCrIr*rHrr.r/rJr+r+r=rH"z$TFFunnelForTokenClassification.buildrpr)rOr\rnr]ror]rPr]rr^rr^rr^rr]rMr_rQr)rrrQrrR)rTrUrVr-rr rrrsrrrtrQrjrHrXr+r+r;r=rs*  ,rz Funnel Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of the hidden-states output to compute `span start logits` and `span end logits`). csreZdZd#fdd Zeeedede e d  d$d%ddZ d&dd Z d'd!d"Z ZS)(TFFunnelForQuestionAnsweringr.r$rQrRcsVtj|g|Ri||j|_t|dd|_tjj|jt|j dd|_ ||_ dS)NrIr qa_outputsr) r,r-rr.rIrr1rrr?rr.rUr;r+r=r-s z%TFFunnelForQuestionAnswering.__init__rWrvrYNFrOr\rnr]rorPrr^rrstart_positions end_positionsrMr_7Union[Tuple[tf.Tensor], TFQuestionAnsweringModelOutput]c  Cs|j|||||||| d} | d} || } tj| ddd\}}tj|dd}tj|dd}d}|durE| durE|| d}||||f}|s\||f| dd}|durZ|f|S|St|||| j| jd S) a start_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*): Labels for position (index) of the start of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. end_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*): Labels for position (index) of the end of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. rzrrZrtrzN)Zstart_positionZ end_positionr#)r start_logits end_logitsrr) rIrrBrrrr rr)r9rOrnrorPrrrrrrMr1rr6rrrrrr+r+r=rQs:  z!TFFunnelForQuestionAnswering.callrr cCst|j|j|j|jdS)N)rrrr)r rrrrrir+r+r=rj8s z+TFFunnelForQuestionAnswering.serving_outputcCr)NTrIr) rFrGrBrCrIr*rHrr.r/rJr+r+r=rHBrz"TFFunnelForQuestionAnswering.buildrp) NNNNNNNNNF)rOr\rnr]ror]rPr]rr^rr^rr^rr]rr]rMr_rQr)rr rQr rR)rTrUrVr-rr rrrsrr rtrQrjrHrXr+r+r;r=rs,  : r) rPr~rrwrrrrurH)TF)RrW __future__rr dataclassesrtypingrrrnumpynpZ tensorflowrBZactivations_tfrZmodeling_tf_outputsr r r r r rZmodeling_tf_utilsrrrrrrrrrrrZtf_utilsrrrutilsrrrr r!r"Zconfiguration_funnelr$Z get_loggerrTloggerrtrr1ZLayerr%rYrrrrrr rrr.r2r9rDrHrOZFUNNEL_START_DOCSTRINGrrrPrurwr~rrrr__all__r+r+r+r=s    4   0w8# O@Pq$--44NROcR b