S^(h_dZddlmZddlZddlmZddlmZmZm Z m Z ddl Z ddl ZddlmZddlmZmZmZmZmZmZmZdd lmZmZmZmZmZmZm Z m!Z!m"Z"m#Z#m$Z$dd l%m&Z&m'Z'm(Z(dd l)m*Z*m+Z+m,Z,m-Z-m.Z.m/Z/d d l0m1Z1e.jde3Z4dZ5dZ6GddZ7Gdde"jpjrZ:Gdde"jpjrZ;Gdde"jpjrZ<Gdde"jpjrZ=Gdde"jpjrZ>GddeZ?Gdde"jpjrZ@e#Gd d!e"jpjrZAeGd"d#e*ZBd$ZCd%ZDe,d&eCGd'd(e?ZEe,d)eCGd*d+e?e7ZFGd,d-e"jpjrZGe,d.eCGd/d0e?eZHe,d1eCGd2d3e?eZIe,d4eCGd5d6e?e ZJe,d7eCGd8d9e?eZKe,d:eCGd;dzTF 2.0 ALBERT model.) annotationsN) dataclass)DictOptionalTupleUnion)get_tf_activation)TFBaseModelOutputTFBaseModelOutputWithPoolingTFMaskedLMOutputTFMultipleChoiceModelOutputTFQuestionAnsweringModelOutputTFSequenceClassifierOutputTFTokenClassifierOutput) 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) AlbertConfigzalbert/albert-base-v2r'ceZdZdZddZy)TFAlbertPreTrainingLossz Loss function suitable for ALBERT pretraining, that is, the task of pretraining a language model by combining SOP + MLM. .. note:: Any label of -100 will be ignored (along with the corresponding logits) in the loss computation. c  tjjdtjjj}|j j r~tjtj|ddd}tjtj|ddt|dd f| }tjtj|dd| }tjtj|d dd}tjtj|d d | }tjtj|d d| } |||} || |} tj| dt| df} tj| d} | | zS|tjj|d|d} tj|ddk7| j } | | z}tj"|tj"| z }tj|d d }|tjj|d |}tj|d dk7|j }||z}tj"|tj"|z }tj||zdS)NT) from_logits reductionlabels)tensorshapeirr.)r0masksentence_order_labelr&)r.r2)y_truey_pred) input_tensoraxisdtype)r&)rlossesSparseCategoricalCrossentropy ReductionNONEconfigtf_legacy_losstf not_equalreshape boolean_maskr reduce_meannnrelucastr: reduce_sum)selfr-logitsloss_fnmasked_lm_active_lossmasked_lm_reduced_logitsmasked_lm_labelssentence_order_active_losssentence_order_reduced_logitsr4masked_lm_losssentence_order_lossunmasked_lm_losses lm_loss_maskmasked_lm_lossesreduced_masked_lm_loss sop_logitsunmasked_sop_loss sop_loss_maskmasked_sop_lossreduced_masked_sop_losss k/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/transformers/models/albert/modeling_tf_albert.pyhf_compute_lossz'TFAlbertPreTrainingLoss.hf_compute_lossJs,,<<Y^YeYeYoYoYtYt<u ;; % %%'LL6(CS[`1acg$h !')zz2z&QR)?TUV?W:XY*( $ "zz)9GNc  *, &)?"@NPT* &-/OOzz'BIc- )$&??zz0F)GuU\v$ %,$ FN_NeNef +m;"$--"@2==Q^C_"_zz03JJDQQN)r- tf.TensorrKr`returnr`)__name__ __module__ __qualname____doc__r^r_r]r)r)Ds  -Rr_r)cXeZdZdZdfd ZddZ d ddZxZS) TFAlbertEmbeddingszGConstruct the embeddings from word, position and token_type embeddings.c Vt|di|||_|j|_|j|_|j |_t jj|jd|_ t jj|j|_ y)N LayerNormepsilonnameraterf)super__init__r?embedding_sizemax_position_embeddingsinitializer_rangerlayersLayerNormalizationlayer_norm_epsrjDropouthidden_dropout_probdropoutrJr?kwargs __class__s r]rqzTFAlbertEmbeddings.__init__}s "6" $33'-'E'E$!'!9!988AVAV]h8i||++1K1K+L r_ctjd5|jd|jj|j gt |j|_dddtjd5|jd|jj|j gt |j|_ dddtjd5|jd|j|j gt |j|_ ddd|jryd|_ t|ddetj|jj 5|jj#dd|jj gdddyy#1swY[xYw#1swYxYw#1swYxYw#1swYyxYw) Nword_embeddingsweight)rmr1 initializertoken_type_embeddings embeddingsposition_embeddingsTrj)rA name_scope add_weightr? vocab_sizerrrrtrtype_vocab_sizerrsrbuiltgetattrrjrmbuildrJ input_shapes r]rzTFAlbertEmbeddings.builds ]], - //{{--t/B/BC+D,B,BC*DK ]]2 3 )-!{{22D4G4GH+D,B,BC*9*D & ]]0 1 '+!33T5H5HI+D,B,BC(7(D $  ::  4d + 7t~~223 O$$dD$++2L2L%MN O O 81       O Os2A F<A G *AG?3G!<G GG!G*c>| | td|At||jjt j |j |}t|dd}|t j|d}|2t jt j||d|zd }t j |j|}t j |j|} ||z| z} |j| } |j| | } | S) z Applies embedding based on inputs tensor. Returns: final_embeddings (`tf.Tensor`): output embedding tensor. Nz5Need to provide either `input_ids` or `input_embeds`.)paramsindicesr.rdimsvaluer&)startlimit)r8inputsrtraining) ValueErrorrr?rrAgatherrrfill expand_dimsrangerrrjrz) rJ input_ids position_idstoken_type_ids inputs_embedspast_key_values_lengthrrposition_embedstoken_type_embedsfinal_embeddingss r]callzTFAlbertEmbeddings.calls   !6TU U  *9dkk6L6L MIIT[[)LM /4  !WW+Q?N  >>5[^Nd=delmL))4+C+C\ZIIT-G-GQ_`(?:=NN>>1A>B<>1M>N$% QR(88r_c|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|dddtj|jj 5|jj dd|jjgdddt|dddtj|jj 5|jj dd|jjgdddt|dddtj|jj 5|jj dd|jjgdddt|ddetj|jj 5|jj dd|jjgdddyy#1swYxYw#1swYlxYw#1swYxYw#1swYxYw#1swYyxYw)NTrrrrrj) rrrArrrmrr?rrrrrjrs r]rzTFAlbertAttention.build.s ::  4$ ' 3tzz/ H   $dkk.E.E!FG H 4 % 1txx}}- FdDKK,C,CDE F 4$ ' 3tzz/ H   $dkk.E.E!FG H 4$ ' 3tzz/ H   $dkk.E.E!FG H 4d + 7t~~223 L$$dD$++2I2I%JK L L 8 H H F F H H H H L Ls< 3I <3I-3I&3I33I? II#&I03I<?Jr)r0r`rrrar`F) r7r`rr`rr`rrrrraTuple[tf.Tensor]r) rbrcrdrerqrrrrrs@r]rrs\[B7 33"3 3  3  3 3jLr_rcJeZdZdfd Z d ddZddZxZS) TFAlbertLayerc t|d i|t|d|_tj j |jt|jd|_ t|jtrt|j|_n|j|_tj j |j t|jd|_tj j%|j&d|_tj j+|j,|_||_y) N attentionrmffnr ffn_outputfull_layer_layer_normrkrnrf)rprqrrrrurintermediate_sizerrtr isinstance hidden_actstrr activationrrrvrwrrxryrzr?r{s r]rqzTFAlbertLayer.__init__Ds  "6"*6 D<<%%**vOgOg?hot&  f'' -/0A0ABDO$//DO,,,,$$IaIa9biu- &+\\%D%D))0G&E& "||++1K1K+L  r_c |j|||||}|j|d}|j|}|j|}|j ||}|j ||dz}|f|ddz}|S)N)r7rrrrrrrr&)rrrrrzr) rJrrrrrattention_outputsrrs r]rzTFAlbertLayer.callZs!NN&)/ + XX%6q%9X: __Z0 __J_7 \\h\G 22*GXYZG[:[2\ !"%6qr%::r_c|jryd|_t|ddMtj|jj 5|jj ddddt|dddtj|jj 5|jj dd|jjgdddt|dddtj|jj 5|jj dd|jjgdddt|ddetj|jj 5|jj dd|jjgdddyy#1swY_xYw#1swYxYw#1swYxYw#1swYyxYw)NTrrrr) rrrArrrmrrr?rrrrrs r]rzTFAlbertLayer.buildts ::  4d + 7t~~223 +$$T* + 4 % 1txx}}- FdDKK,C,CDE F 4t , 8t334 S%%tT4;;3P3P&QR S 40$ 7 Ct99>>? X**00$dkk>U>U1VW X X D + + F F S S X Xs0 G%3G3G3G)GGG&)G2rr) rr`rr`rr`rrrrrarrrbrcrdrqrrrrs@r]rrCsR8  "      4Xr_rcNeZdZdfd Z d ddZddZxZS)TFAlbertLayerGroupc t|di|t|jDcgc]}t |d|c}|_ycc}w)Nzalbert_layers_._rrf)rprqrinner_group_numr albert_layersrJr?r|ir}s r]rqzTFAlbertLayerGroup.__init__sN "6"INfNdNdHe CDM&)9!'= >  sAc|rdnd}|rdnd}t|jD].\} } |r||fz}| |||| ||} | d}|s&|| dfz}0|r||fz}td|||fDS)Nrf)rrrrrrr&c3&K|] }|| ywrrf.0vs r] z*TFAlbertLayerGroup.call..sh1Z[ZgQh) enumeratertuple) rJrrrroutput_hidden_statesrlayer_hidden_stateslayer_attentions layer_index albert_layer layer_outputs r]rzTFAlbertLayerGroup.calls%9bd!22)243E3E)F I %K#&9], ( TYY_YqYqSr$ NO v.DQC,H I$   $ s5Cc |j|}|rdnd}|r|fnd} t|jD]z} t| |j|jz z } |j | |||| |j z| dz|j z|||} | d}|r|| dz}|su| |fz} ||std|| |fDSt|| |S) Nrrfr&)rrrrrrrr.c3&K|] }|| ywrrfrs r]rz+TFAlbertTransformer.call..shqZ[Zghr)last_hidden_stater attentions) rrrrrrrrr ) rJrrrrr return_dictrall_attentionsall_hidden_statesr group_idxlayer_group_outputs r]rzTFAlbertTransformer.calls88 8N 0d0D],$t--. IAA!7!7$:P:P!PQRI!D!9!9)!D+-#I0E0E$EUVZ^ZoZoHop"3%9! " /q1M !/2DR2H!H#$58H$H!# I&h]4E~$Vhh h +;LYg  r_c|jryd|_t|dddtj|jj 5|jj dd|jjgdddt|ddK|jD];}tj|j 5|j dddd=yy#1swYbxYw#1swYUxYw)NTrr) rrrArrrmrr?rrrr s r]rzTFAlbertTransformer.builds ::  46 = It??DDE a0066dDKKD^D^7_` a 4. 5 A11 &]]5::.&KK%&& & B a a&&s 3CC+C(+C4 rr)rr`rr`rr`rrrrrrrrrarrrrs@r]rrsf2&  & "&  &  & # & & &  4& P &r_rceZdZdZeZdZy)TFAlbertPreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. albertN)rbrcrdrer' config_classbase_model_prefixrfr_r]r#r#s  L r_r#cPeZdZdfd Zd dZd dZd dZd dZd dZd dZ xZ S)TFAlbertMLMHeadc t|di|||_|j|_tj j |jt|jd|_ t|jtrt|j|_n|j|_tj j|j d|_||_y)Nr)rrmrjrkrf)rprqr?rrrrurrrtrrrrr rrvrwrjdecoder)rJr?input_embeddingsr|r}s r]rqzTFAlbertMLMHead.__init__ s "6" $33\\''  ! !ofF^F^6_fm(   f'' -/0A0ABDO$//DO88AVAV]h8i( r_c|j|jjfddd|_|j|jjfddd|_|j ryd|_t |dddtj|jj5|jjdd|jjgdddt |ddetj|jj5|jjdd|jjgdddyy#1swY|xYw#1swYyxYw)NzerosTbias)r1r trainablermz decoder/biasrrj)rr?rr. decoder_biasrrrArrrmrrrjrrrs r]rzTFAlbertMLMHead.builds5OO4;;+A+A*CQXdhouOv  OO;;))+DWe,  ::  4$ ' 3tzz/ H   $dkk.E.E!FG H 4d + 7t~~223 O$$dD$++2L2L%MN O O 8 H H O Os)3E3E#E #E,c|jSr)r*rJs r]get_output_embeddingsz%TFAlbertMLMHead.get_output_embeddings/s ||r_c`||j_t|d|j_yNr)r*rrrrJrs r]set_output_embeddingsz%TFAlbertMLMHead.set_output_embeddings2s$# ",U"3A"6 r_c4|j|jdS)Nr.r0r9r2s r]get_biaszTFAlbertMLMHead.get_bias6s 43D3DEEr_cl|d|_|d|_t|dd|j_y)Nr.r0r)r.r0rr?rr6s r]set_biaszTFAlbertMLMHead.set_bias9s4&M !.1!+E&M!:1!= r_c|j|}|j|}|j|}t|d}t j |d|j g}t j||jjd}t j |d||jjg}tjj||j}|S) Nr)r0r&r.r/T)abr)rr.)rrrjrrArCrrrr*rr?rrFbias_addr0)rJr seq_lengths r]rzTFAlbertMLMHead.call>s - 8  6 m< }5a8  -DDWDW?XY  MT\\5H5HVZ[  -JPTP[P[PfPf?gh ]ARARS r_)r?r'r+keras.layers.LayerrrarBrz tf.Variable)razDict[str, tf.Variable])rr`rar`) rbrcrdrqrr3r7r:r<rrrs@r]r(r( s((&O 7F> r_r(ceZdZeZddfd Zd dZd dZdZe d d dZ d dZ xZ S)TFAlbertMainLayerc t|di|||_t|d|_t |d|_|rFtjj|jt|jdd|_ yd|_ y)Nrrencodertanhpooler)rrrrmrf)rprqr?rhrrrHrrurrrrtrJ)rJr?add_pooling_layerr|r}s r]rqzTFAlbertMainLayer.__init__Os "6" ,V,G*6 B ! LL  ((#263K3K#L!    r_c|jSr)rr2s r]get_input_embeddingsz&TFAlbertMainLayer.get_input_embeddingsas r_c`||j_t|d|j_yr5)rrrrr6s r]set_input_embeddingsz&TFAlbertMainLayer.set_input_embeddingsds$!&%/%6q%9"r_ct)z Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base class PreTrainedModel )NotImplementedError)rJheads_to_prunes r] _prune_headszTFAlbertMainLayer._prune_headshs "!r_c | | td| t|} n|t|dd} n td|tj| d}|tj| d}|j ||||| } tj || ddd| df} tj | | j} tjd | j}tjd | j}tjtj|| |} |tdg|jjz}|j| | |||| | }|d}|j|j|dddf nd}| s ||f|ddzSt!|||j"|j$ S)NzDYou cannot specify both input_ids and inputs_embeds at the same timer.z5You have to specify either input_ids or inputs_embedsr&rr)rrrrrr9g?g)rrrrrrrr)r pooler_outputrr)rrrArrrCrHr:constantrsubtractrQr?rrHrJr rr)rJrrrrrrrrrrrembedding_outputextended_attention_maskone_cstten_thousand_cstencoder_outputssequence_output pooled_outputs r]rzTFAlbertMainLayer.callos  ]%>cd d  "$Y/K  &$]3CR8KTU U  !WW+Q?N  !WW+Q?N??%)' + #%**^k!naQRT_`aTb=c"d#%''*AIYI_I_"`++c)9)?)?@;;x7G7M7MN"$++bkk'CZ.[]m"n  % %!>!>>I,,*2/!5#' *!,EI[[E\ ?1a4+@ Abf  #$ $ ,-')77&11   r_c|jryd|_t|ddMtj|jj 5|jj ddddt|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swYxYw#1swYxYw#1swYyxYw)NTrrHrJ) rrrArrrmrrHrJr?rrs r]rzTFAlbertMainLayer.builds ::  4t , 8t334 ,%%d+ , 4D ) 5t||001 ) ""4( ) 44 ( 4t{{//0 I !!4t{{/F/F"GH I I 5  , , ) ) I Is$ D<%E?3E<EEE)T)r?r'rKrrCrD NNNNNNNNNF)rTFModelInputType | Nonernp.ndarray | tf.Tensor | NonerrbrrbrrbrrbrOptional[bool]rrcrrcrrra5Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]r) rbrcrdr'r%rqrMrOrSrrrrrs@r]rFrFKsL $:".28<8<6:377;,0/3&*W *W 6W 6 W 4 W 1 W 5W *W -W $W W  ?W W r Ir_rFcXeZdZUdZdZded<dZded<dZded<dZded<dZ ded <y) TFAlbertForPreTrainingOutputa4 Output type of [`TFAlbertForPreTraining`]. Args: prediction_logits (`tf.Tensor` of shape `(batch_size, sequence_length, config.vocab_size)`): Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). sop_logits (`tf.Tensor` of shape `(batch_size, 2)`): Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation 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. Nrlossprediction_logitsrXzTuple[tf.Tensor] | Nonerr) rbrcrdrerg__annotations__rhrXrrrfr_r]rfrfsA,!%D $-1*1&*J#*-1M*1*.J'.r_rfav 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 ([`AlbertConfig`]): 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). z`The bare Albert Model transformer outputting raw hidden-states without any specific head on top.ceZdZdfd Zeeejdee e e  d ddZ d dZ xZS) TFAlbertModelcPt||g|i|t|d|_y)Nr$r)rprqrFr$rJr?rr|r}s r]rqzTFAlbertModel.__init__^s( 3&3F3'X> r_batch_size, sequence_length checkpoint output_typer%c <|j||||||||| |  } | S)N rrrrrrrrrr)r$) rJrrrrrrrrrrrs r]rzTFAlbertModel.callcs<(++))%'/!5#  r_c|jryd|_t|ddNtj|jj 5|jj ddddyy#1swYyxYw)NTr$)rrrArr$rmrrs r]rzTFAlbertModel.buildsi ::  44 ( 4t{{//0 ( !!$' ( ( 5 ( (s A11A:rr`)rrarrbrrbrrbrrbrrbrrcrrcrrcrrcrardr)rbrcrdrqrr#ALBERT_INPUTS_DOCSTRINGformatr!_CHECKPOINT_FOR_DOCr _CONFIG_FOR_DOCrrrrs@r]rkrkYs ? *+B+I+IJg+hi&0$.28<8<6:377;,0/3&*#(*66  4  1 5*-$! ? j8(r_rkz Albert Model with two heads on top for pretraining: a `masked language modeling` head and a `sentence order prediction` (classification) head. ceZdZdgZdfd Zd dZeeejde e e  d d dZ d dZxZS) TFAlbertForPreTrainingpredictions.decoder.weightct||g|i||j|_t|d|_t ||jj d|_t|d|_ y)Nr$r predictionsr+rmsop_classifier) rprq num_labelsrFr$r(rr}TFAlbertSOPHeadrrms r]rqzTFAlbertForPreTraining.__init__sa 3&3F3 ++'X> *6DKKDZDZano-f;KLr_c|jSrr}r2s r] get_lm_headz"TFAlbertForPreTraining.get_lm_headr_rnrqr%c N|j||||||||| |  } | dd\}}|j|}|j|| }d}| | d| i}| |d<|j|||f}| s||f| ddz}||f|zS|St |||| j | j S) ao Return: Example: ```python >>> import tensorflow as tf >>> from transformers import AutoTokenizer, TFAlbertForPreTraining >>> tokenizer = AutoTokenizer.from_pretrained("albert/albert-base-v2") >>> model = TFAlbertForPreTraining.from_pretrained("albert/albert-base-v2") >>> input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :] >>> # Batch size 1 >>> outputs = model(input_ids) >>> prediction_logits = outputs.prediction_logits >>> sop_logits = outputs.sop_logits ```rsNr2)r)r^rr-r4r-rK)rgrhrXrr)r$r}rr^rfrr)rJrrrrrrrrrr-r4rrr]r^prediction_scores sop_scores total_lossd_labelsoutputs r]rzTFAlbertForPreTraining.calls L++))%'/!5#  *1!& ,,?,K((}x(X   "6"B &)H/CH+ ,--XGXZdFe-fJ'4wqr{BF/9/EZMF* Q6 Q+/!!//))   r_c|jryd|_t|ddMtj|jj 5|jj ddddt|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYxYw#1swYqxYw#1swYyxYw)NTr$r}r) rrrArr$rmrr}rrs r]rzTFAlbertForPreTraining.builds  ::  44 ( 4t{{//0 ( !!$' ( 4 - 9t//445 -  &&t, - 4)4 0 <t22778 0##))$/ 0 0 =  ( ( - - 0 0s$ D%%D1?D=%D.1D:=ErrC NNNNNNNNNNNF)rrarrbrrbrrbrrbrrbrrcrrcrrcr-rbr4rbrrcraz5Union[TFAlbertForPreTrainingOutput, Tuple[tf.Tensor]]r)rbrcrd"_keys_to_ignore_on_load_unexpectedrqrrr#rurvr%rfrxrrrrs@r]rzrzs+H)H&M *+B+I+IJg+hi+GVef.28<8<6:377;,0/3&*04>B#(C *C 6C 6 C 4 C 1 C 5C *C -C $C .C <C !C  ?C gjC J 0r_rzc0eZdZdfd ZddZddZxZS)rc t|di|tjj |j |_tjj|jt|jd|_ ||_ y)Nrn classifierrrf) rprqrrurxclassifier_dropout_probrzrrrrtrr?r{s r]rqzTFAlbertSOPHead.__init__sp "6"||++1O1O+P ,,,,##.v/G/GH-   r_cP|j||}|j|}|S)Nrr)rzr)rJr^rdropout_pooled_outputrKs r]rzTFAlbertSOPHead.call s+ $ MH U(=> r_c(|jryd|_t|ddetj|jj 5|jj dd|jjgdddyy#1swYyxYw)NTr) rrrArrrmrr?rrs r]rzTFAlbertSOPHead.builds ::  4t , 8t334 M%%tT4;;3J3J&KL M M 9 M Ms 3BBr)r^r`rrrar`rrrs@r]rrs  Mr_rz4Albert Model with a `language modeling` head on top.ceZdZddgZd fd Zd dZeeejde e e  d d dZ d dZxZS)TFAlbertForMaskedLMrJr{ct||g|i|t|dd|_t ||jj d|_y)NFr$rKrmr}r~)rprqrFr$r(rr}rms r]rqzTFAlbertForMaskedLM.__init__sF 3&3F3'%hW *6DKKDZDZanor_c|jSrrr2s r]rzTFAlbertForMaskedLM.get_lm_head#rr_rnrc |j||||||||| |  } | d} |j| | }| dn|j| |}| s|f| ddz}||f|zS|St||| j| j S)a labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]` Returns: Example: ```python >>> import tensorflow as tf >>> from transformers import AutoTokenizer, TFAlbertForMaskedLM >>> tokenizer = AutoTokenizer.from_pretrained("albert/albert-base-v2") >>> model = TFAlbertForMaskedLM.from_pretrained("albert/albert-base-v2") >>> # add mask_token >>> inputs = tokenizer(f"The capital of [MASK] is Paris.", return_tensors="tf") >>> logits = model(**inputs).logits >>> # retrieve index of [MASK] >>> mask_token_index = tf.where(inputs.input_ids == tokenizer.mask_token_id)[0][1] >>> predicted_token_id = tf.math.argmax(logits[0, mask_token_index], axis=-1) >>> tokenizer.decode(predicted_token_id) 'france' ``` ```python >>> labels = tokenizer("The capital of France is Paris.", return_tensors="tf")["input_ids"] >>> labels = tf.where(inputs.input_ids == tokenizer.mask_token_id, labels, -100) >>> outputs = model(**inputs, labels=labels) >>> round(float(outputs.loss), 2) 0.81 ``` rsr)rrNrr2rgrKrr)r$r}r^r rr)rJrrrrrrrrrr-rrr]rrgrs r]rzTFAlbertForMaskedLM.call&sj++))%'/!5#  "!* ,,?U],^~t4+?+?vVg+?+h')GABK7F)-)9TGf$ Ev E$!//))   r_c|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYw)NTr$r})rrrArr$rmrr}rs r]rzTFAlbertForMaskedLM.buildws ::  44 ( 4t{{//0 ( !!$' ( 4 - 9t//445 -  &&t, - - : ( ( - -s C %C CC rrC NNNNNNNNNNF)rrarrbrrbrrbrrbrrbrrcrrcrrcr-rbrrcraz)Union[TFMaskedLMOutput, Tuple[tf.Tensor]]r)rbrcrdrrqrrr#rurvr%r rxrrrrs@r]rrs +45R)S&p  *+B+I+IJg+hi+;/Z.28<8<6:377;,0/3&*04#(L *L 6L 6 L 4 L 1 L 5L *L -L $L .L !L  3L [jL \ -r_rz Albert Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled output) e.g. for GLUE tasks. c eZdZdgZdgZd fd Zeeejde de e dd d d d Z dd ZxZS)!TFAlbertForSequenceClassificationr}rzcft||g|i||j|_t|d|_t j j|j|_ t j j|jt|jd|_ ||_y)Nr$rrnrr)rprqrrFr$rrurxrrzrrrtrr?rms r]rqz*TFAlbertForSequenceClassification.__init__s 3&3F3 ++'X> ||++1O1O+P ,,,,##H`H`8aht-  r_rnzvumichien/albert-base-v2-imdbz 'LABEL_1'gQ?)rprqr%expected_output expected_lossc *|j||||||||| |  } | d} |j| | } |j| }| dn|j| |}| s|f| ddz}||f|zS|St ||| j | j S) a labels (`tf.Tensor` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). rsr&rrNrr2r)r$rzrr^rrr)rJrrrrrrrrrr-rrr^rKrgrs r]rz&TFAlbertForSequenceClassification.calls:++))%'/!5#     MH M  6~t4+?+?vV\+?+]Y,F)-)9TGf$ Ev E)!//))   r_c|jryd|_t|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYwNTr$r rrrArr$rmrrr?rrs r]rz'TFAlbertForSequenceClassification.build ::  44 ( 4t{{//0 ( !!$' ( 4t , 8t334 M%%tT4;;3J3J&KL M M 9 ( ( M M C"%3C."C+.C7rr)rrarrbrrbrrbrrbrrbrrcrrcrrcr-rbrrcraz3Union[TFSequenceClassifierOutput, Tuple[tf.Tensor]]r)rbrcrdr_keys_to_ignore_on_load_missingrqrr#rurvr!rrxrrrrs@r]rrs+9)9&'1l# *+B+I+IJg+hi2.$# .28<8<6:377;,0/3&*04#(/ */ 6/ 6 / 4 / 1 / 5/ */ -/ $/ ./ !/  =/ j/ b Mr_rz Albert Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. ceZdZddgZdgZd fd Zeeejde e e e  d d dZd dZxZS) TFAlbertForTokenClassificationrJr}rzct||g|i||j|_t|dd|_|j |j n |j }tjj||_ tjj|jt|jd|_||_y)NFr$rrnrr)rprqrrFr$rryrrurxrzrrrtrr?)rJr?rr|rr}s r]rqz'TFAlbertForTokenClassification.__init__s 3&3F3 ++'%hW --9  * *++ ||++1H+I ,,,,##H`H`8aht-  r_rnroc *|j||||||||| |  } | d} |j| | } |j| }| dn|j| |}| s|f| ddz}||f|zS|St ||| j | j S) z labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`. rsrrrNrr2r)r$rzrr^rrr)rJrrrrrrrrrr-rrr]rKrgrs r]rz#TFAlbertForTokenClassification.calls2++))%'/!5#  "!*,,o,Q8~t4+?+?vV\+?+]Y,F)-)9TGf$ Ev E&!//))   r_c|jryd|_t|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYwrrrs r]rz$TFAlbertForTokenClassification.build4rrrr)rrarrbrrbrrbrrbrrbrrcrrcrrcr-rbrrcraz0Union[TFTokenClassifierOutput, Tuple[tf.Tensor]]r)rbrcrdrrrqrr#rurvr!rwrrxrrrrs@r]rrs+4^)D&'1l#"*+B+I+IJg+hi&+$.28<8<6:377;,0/3&*04#(- *- 6- 6 - 4 - 1 - 5- *- -- $- .- !-  :- j- ^ Mr_rz Albert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`). c eZdZddgZd fd Zeeejde de e dddd  d dd Z dd Z xZS)TFAlbertForQuestionAnsweringrJr}c t||g|i||j|_t|dd|_t j j|jt|jd|_ ||_ y)NFr$r qa_outputsr) rprqrrFr$rrurrrtrr?rms r]rqz%TFAlbertForQuestionAnswering.__init__Kst 3&3F3 ++'%hW ,,,,##H`H`8aht-  r_rnzvumichien/albert-base-v2-squad2 z'a nice puppet'gq= ףp@)rprqr%qa_target_start_indexqa_target_end_indexrrc |j||||||||| |  } | d}|j|}tj|dd\}}tj|d}tj|d}d}| | d | i}| |d <|j |||f }| s||f| ddz}||f|zS|St |||| j| j S) a start_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*): Labels for position (index) of the start of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. end_positions (`tf.Tensor` of shape `(batch_size,)`, *optional*): Labels for position (index) of the end of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. rsrrr2r.)rnum_or_size_splitsr8)inputr8Nstart_position end_positionr)rg start_logits end_logitsrr) r$rrAsplitsqueezer^rrr)rJrrrrrrrrrstart_positions end_positionsrrr]rKrrrgr-rs r]rz!TFAlbertForQuestionAnswering.callVsH++))%'/!5#  "!*8#%88&QUW#X jzz 2> ZZjr:   &=+D&8F%2F> "''v|Z>X'YD"J/'!"+=F)-)9TGf$ Ev E-%!!//))   r_c|jryd|_t|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYw)NTr$r) rrrArr$rmrrr?rrs r]rz"TFAlbertForQuestionAnswering.buildrrrr)rrarrbrrbrrbrrbrrbrrcrrcrrcrrbrrbrrcraz7Union[TFQuestionAnsweringModelOutput, Tuple[tf.Tensor]]r)rbrcrdrrqrr#rurvr!rrxrrrrs@r]rr@s+4^)D& *+B+I+IJg+hi42$ ).28<8<6:377;,0/3&*9=7;#(< *< 6< 6 < 4 < 1 < 5< *< -< $< 7< 5< !<  A< j< | Mr_rz Albert Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a softmax) e.g. for RocStories/SWAG tasks. ceZdZddgZdgZd fd Zeeejde e e e  d d dZd dZxZS) TFAlbertForMultipleChoicerJr}rzc0t||g|i|t|d|_tj j |j|_tj jdt|jd|_ ||_ y)Nr$rrnr&rr)rprqrFr$rrurxryrzrrrtrr?rms r]rqz"TFAlbertForMultipleChoice.__init__s~ 3&3F3'X> ||++1K1K+L ,,,,8P8P(QXd-  r_z(batch_size, num_choices, sequence_lengthroc  |t|d} t|d} nt|d} t|d} |tj|d| fnd}|tj|d| fnd}|tj|d| fnd}|tj|d| fnd}|&tj|d| t|dfnd}|j||||||||| |  }|d}|j || }|j | }tj|d| f}| dn|j | | }| s|f|ddz}||f|zS|St|||j|j S) a5 labels (`tf.Tensor` of shape `(batch_size,)`, *optional*): Labels for computing the multiple choice classification loss. Indices should be in `[0, ..., num_choices]` where `num_choices` is the size of the second dimension of the input tensors. (See `input_ids` above) Nr&r2r.r/r rsrrrr) rrArCr$rzrr^rrr)rJrrrrrrrrrr-r num_choicesrAflat_input_idsflat_attention_maskflat_token_type_idsflat_position_idsflat_inputs_embedsrr^rKreshaped_logitsrgrs r]rzTFAlbertForMultipleChoice.calls6  $Y/2K#I.q1J$]3A6K#M215JDMDYIJ/?@_cIWIcBJJnR4D Eim JXIcBJJnR4D Eim HTG_BJJl2z2B Cei  ( JJmB J}D]^_D`3a b  ++$..*,/!5#     MH M  6**F2{:KL~t4+?+?vVe+?+f%''!"+5F)-)9TGf$ Ev E*"!//))   r_c|jryd|_t|ddMtj|jj 5|jj ddddt|ddetj|jj 5|jj dd|jjgdddyy#1swY|xYw#1swYyxYwrrrs r]rzTFAlbertForMultipleChoice.buildrrrr)rrarrbrrbrrbrrbrrbrrcrrcrrcr-rbrrcraz4Union[TFMultipleChoiceModelOutput, Tuple[tf.Tensor]]r)rbrcrdrrrqrr#rurvr!rwrrxrrrrs@r]rrs+4^)D&'1l#*+B+I+IJt+uv&/$.28<8<6:377;,0/3&*04#(E *E 6E 6 E 4 E 1 E 5E *E -E $E .E !E  >E wE N Mr_r) r#rkrzrrrrrrF)Nre __future__rr dataclassesrtypingrrrrnumpynp tensorflowrAactivations_tfr modeling_tf_outputsr r r rrrrmodeling_tf_utilsrrrrrrrrrrrtf_utilsrrrutilsr r!r"r#r$r%configuration_albertr' get_loggerrbloggerrwrxr)ruLayerrhrrrrr#r(rFrfALBERT_START_DOCSTRINGrurkrzrrrrrr__all__rfr_r]rs " !///    SR/   H %- 3R3RlQ ++Q hrL **rLj@XELL&&@XF1&++1&hD&%,,,,D&N!/!=ell((=@II **IIIIX /;/ /<(T5pf/(+/( /(d d046Md0d0NMell((M6PRhig-13Og-jg-T TM(?A]TMTMn UM%<>WUMUMp aM#: