
    h"c                       S SK Jr  S SKrS SKrS SKJr  S SKJrJrJ	r	  S SK
r
S SKrS SKJr  S SK
JrJr  S SKJr  S SKJr  S S	KJrJr  S S
KJrJrJrJr  S SKJr  S SKJr  S SKJ r   S SK!J"r"  S SK#J$r$  S SK%J&r&  S SK'J(r(  S SK)J*r*  S SK+J,r,  \," 5       (       a  S SK-J.r.  \(       a  S SK/J0r0  S SK1J&r&  \Rd                  " \35      r4 " S S\5      r5 " S S\5      r6 " S S\5      r7 " S S\Rp                  5      r9 " S S5      r:g)     )annotationsN)Path)TYPE_CHECKINGAnyCallable)version)Tensornn)	Optimizer)
DataLoader)tqdmtrange)TrainerCallbackTrainerControlTrainerStateis_torch_npu_available)BatchEncoding)CrossEncoderTrainingArguments)NoDuplicatesDataLoader)SentenceLabelDataset)SentenceEvaluator)InputExample)SentenceTransformer)BatchSamplers)is_datasets_available)Dataset)CrossEncoderc                  \   ^  \ rS rSrSrSU 4S jjrSS jr            S	S jrSrU =r	$ )
SaveModelCallback&   a[  A Callback to save the model to the `output_dir`.

If save_best_model is True and evaluator is defined, then we save on evaluate, but only if the new model is
better than the currently saved one according to the evaluator.

This differs from the `SaveModelCallback` used in SentenceTransformer.fit where the model is saved after
training as well.
c                T   > [         TU ]  5         Xl        X l        X0l        S U l        g N)super__init__
output_dir	evaluatorsave_best_modelbest_metric)selfr%   r&   r'   	__class__s       g/home/james-whalen/.local/lib/python3.13/site-packages/sentence_transformers/cross_encoder/fit_mixin.pyr$   SaveModelCallback.__init__0   s&    $".    c                r    [        U R                  SS5      (       a  XR                  :  $ XR                  :  $ )Ngreater_is_betterT)getattrr&   r(   )r)   
new_metrics     r+   	is_betterSaveModelCallback.is_better7   s4    4>>#6== 0 000,,,,r-   c                `   U R                   b  U R                  (       a  [        U R                   SS5      nUR                  5        Hc  u  pUR	                  U5      (       d  M  U R
                  b  U R                  U	5      (       d  MB  Xl        UR                  U R                  5        Me     g g g Nprimary_metricr&   )	r&   r'   r0   itemsendswithr(   r2   saver%   )
r)   argsstatecontrolmetricsmodelkwargs
metric_keykeyvalues
             r+   on_evaluateSaveModelCallback.on_evaluate<   s     >>%$*>*> 1A;OJ%mmo
<<
++''/4>>%3H3H+0(

4??3	 . +?%r-   )r(   r&   r%   r'   )r%   strr&   SentenceEvaluator | Noner'   boolreturnNone)r1   floatrH   rG   )r:   r   r;   r   r<   r   r=   dict[str, Any]r>   r   rH   rI   )
__name__
__module____qualname____firstlineno____doc__r$   r2   rC   __static_attributes____classcell__r*   s   @r+   r   r   &   sV     -
4+4 4  	4
  4 4 
4 4r-   r   c                  R   ^  \ rS rSrSrSSU 4S jjjr          SS jrSrU =r$ )	EvaluatorCallbackN   zThe CrossEncoder.fit method always ran the evaluator on every epoch,
in addition to every "evaluation_steps". This callback is responsible for that.

The `.trainer` must be provided after the trainer has been created.
c                  > [         TU ]  5         Xl        X l        U R                  bN  [        R
                  R                  U R                  S5      U l        [        R                  " U R                  SS9  SU l        S U l	        g )NevalTexist_ok)
r#   r$   r&   output_pathospathjoinmakedirsmetric_key_prefixtrainer)r)   r&   r[   r*   s      r+   r$   EvaluatorCallback.__init__U   se    "&'!ww||D,<,<fEDKK((48!'r-   c                   U R                  X@R                  UR                  UR                  S9n[	        U[
        5      (       d  SU0n[        UR                  5       5       HJ  nUR                  U R                   S35      (       a  M(  UR                  U5      X`R                   SU 3'   ML     U R                  b%  U R                  R                  R                  XX6S9  g g )Nr[   epochstepsr&   _)r=   )r&   r[   re   global_step
isinstancedictlistkeys
startswithr`   popra   callback_handlerrC   )r)   r:   r;   r<   r>   r?   evaluator_metricsrA   s           r+   on_epoch_endEvaluatorCallback.on_epoch_end`   s     !NN//u{{%J[J[ + 
 +T22!,.? @ )..01C>>T%;%;$<A">??GXG\G\]`Ga!%;%;$<AcU"CD 2 <<#LL))55d75f $r-   )r&   r`   r[   ra   r"   )r&   r   r[   
str | NonerH   rI   )
r:   r   r;   r   r<   r   r>   r   rH   rI   )	rL   rM   rN   rO   rP   r$   rq   rQ   rR   rS   s   @r+   rU   rU   N   sS    	 	g+g g  	g
 g 
g gr-   rU   c                  N   ^  \ rS rSrSrSU 4S jjr          SS jrSrU =r$ )OriginalCallbackw   zA Callback to invoke the original callback function that was provided to CrossEncoder.fit()

This callback has the following signature: `(score: float, epoch: int, steps: int) -> None`
c                :   > [         TU ]  5         Xl        X l        g r"   )r#   r$   callbackr&   )r)   rx   r&   r*   s      r+   r$   OriginalCallback.__init__}   s     "r-   c                    [        U R                  SS5      nUR                  5        HC  u  pxUR                  U5      (       d  M  U R	                  XR
                  UR                  5      s  $    g r5   )r0   r&   r7   r8   rx   re   rh   )	r)   r:   r;   r<   r=   r?   r@   rA   rB   s	            r+   rC   OriginalCallback.on_evaluate   sS     T^^-={K
!--/JC||J''}}UKK9J9JKK *r-   )rx   r&   )rx   !Callable[[float, int, int], None]r&   r   rH   rI   )
r:   r   r;   r   r<   r   r=   rK   rH   rI   )	rL   rM   rN   rO   rP   r$   rC   rQ   rR   rS   s   @r+   ru   ru   w   sN    
#
L+L L  	L
  L 
L Lr-   ru   c                  ^   ^  \ rS rSrSr\R                  " 5       4SU 4S jjjrSS jrSr	U =r
$ )FitMixinLoss   znA wrapper around the torch loss function that just accepts logits and labels, to be used in CrossEncoder.fit()c                F   > [         TU ]  5         Xl        X l        X0l        g r"   )r#   r$   r>   loss_fctactivation_fn)r)   r>   r   r   r*   s       r+   r$   FitMixinLoss.__init__   s    
 *r-   c                &   [        U5      S:w  a  [        S[        U5       S35      e[        [        US   US   5      5      nU R                  R                  USSSS9nUR                  U R                  R                  5        U R                  " S
0 UD6S   nU R                  R                  R                  S:X  a"  UR                  S	5      nUR                  5       nOUR                  5       nU R                  U5      nU R                  XR5      nU$ )N   z\BinaryCrossEntropyLoss expects a dataset with two non-label columns, but got a dataset with z	 columns.r      Tptpadding
truncationreturn_tensors )len
ValueErrorrk   zipr>   	tokenizertodeviceconfig
num_labelsviewrJ   longr   r   )r)   inputslabelspairstokenslogitslosss          r+   forwardFitMixinLoss.forward   s   v;!norsyozn{  |E  F  SF1I./%%	 & 
 			$**##$%f%a(::''1,[[_F\\^F[[]F##F+}}V,r-   )r   r   r>   )r>   r   r   	nn.Moduler   r   rH   rI   )r   zlist[list[str]]r   r	   rH   r	   )rL   rM   rN   rO   rP   r
   Identityr$   r   rQ   rR   rS   s   @r+   r~   r~      s&    x\^\g\g\i + + r-   r~   c                     \ rS rSrSrSSS\R                  " 5       SS\R                  R                  SS0S	S
SSSSSS4                                 SS jjr
SS jrSS jrSSS\R                  " 5       SS\R                  R                  SS0S	S
SSSSSS4                               SS jjrSS jrSrg)FitMixin   zUMixin class for injecting the `fit` and `old_fit` methods into the CrossEncoder classNr   WarmupLineari'  lrgh㈵>g{Gz?r   TFc                8  ^' [        5       (       d  [        S5      eSSKJn  SSKJn  SSKJn  S nUUl        [        USS5      n[        U[        5      (       a  [        R                  nOQ[        US	5      (       a0  [        UR                  [         5      (       a  [        R"                  nO[        R$                  n/ n/ nU H?  n['        U Vs/ s H  nUR(                  UR*                  4PM     sn6 u  nnUU-  nUU-  nMA     [,        R.                  " [1        ['        U6 5       VVs0 s H  u  nnS
U 3U_M     snn5      n Sn! [3        U5      S1:X  a  Sn!U!(       a  U R7                  SU5      n S!S jn"[8        R:                  " [<        R>                  5      [8        R:                  " S5      :  a  SOSn#[A        S"U"" 5       UUUUS.U#Ub  US:  a  SOS0DUUUU(       + SS.D6n$[C        U RE                  5       5      n%/ SQn&U% V'^'V(s/ s H$  u  m'n([G        U'4S jU& 5       5      (       a  M"  U(PM&     sn(n'U
S.U% V'^'V(s/ s H$  u  m'n([G        U'4S jU& 5       5      (       d  M"  U(PM&     sn(n'SS./n)U" U)40 U	D6n*[        U[H        5      (       a2  [K        U 5      U-  n+[M        U+U-  5      n,[N        RP                  " U*XgU,S9n-Uc(  U RR                  RT                  S:X  a  U" XS9nOU" XS9nO
[W        XUS9n/ n.Ub8  U.RY                  [[        X,5      5        Ub  U.RY                  []        UU5      5        U" U U$U SUUU*W-4U.S 9n/Ub  U/R_                  [a        XU5      5        U/Rc                  5         gs  snf s  snnf ! [4         a     GN3f = fs  sn(n'f s  sn(n'f )#a	  
Deprecated training method from before Sentence Transformers v4.0, it is recommended to use
:class:`~sentence_transformers.trainer.CrossEncoderTrainer` instead. This method uses
:class:`~sentence_transformers.trainer.CrossEncoderTrainer` behind the scenes, but does
not provide as much flexibility as the Trainer itself.

This training approach uses a DataLoader and Loss function to train the model.

This method should produce equivalent results in v4.0 as before v4.0, but if you encounter any issues
with your existing training scripts, then you may wish to use
:meth:`CrossEncoder.old_fit <sentence_transformers.cross_encoder.CrossEncoder.old_fit>` instead.
That uses the old training method from before v4.0.

Args:
    train_dataloader: The DataLoader with InputExample instances
    evaluator: An evaluator (sentence_transformers.cross_encoder.evaluation)
        evaluates the model performance during training on held-
        out dev data. It is used to determine the best model
        that is saved to disk.
    epochs: Number of epochs for training
    loss_fct: Which loss function to use for training. If None,
        will use BinaryCrossEntropy() if self.config.num_labels == 1
        else CrossEntropyLoss(). Defaults to None.
    activation_fct: Activation function applied on top of logits
        output of model.
    scheduler: Learning rate scheduler. Available schedulers:
        constantlr, warmupconstant, warmuplinear, warmupcosine,
        warmupcosinewithhardrestarts
    warmup_steps: Behavior depends on the scheduler. For
        WarmupLinear (default), the learning rate is increased
        from o up to the maximal learning rate. After these many
        training steps, the learning rate is decreased linearly
        back to zero.
    optimizer_class: Optimizer
    optimizer_params: Optimizer parameters
    weight_decay: Weight decay for model parameters
    evaluation_steps: If > 0, evaluate the model using evaluator
        after each number of training steps
    output_path: Storage path for the model and evaluation files
    save_best_model: If true, the best model (according to
        evaluator) is stored at output_path
    max_grad_norm: Used for gradient normalization.
    use_amp: Use Automatic Mixed Precision (AMP). Only for
        Pytorch >= 1.6.0
    callback: Callback function that is invoked after each
        evaluation. It must accept the following three
        parameters in this order: `score`, `epoch`, `steps`
    show_progress_bar: If True, output a tqdm progress bar
zGPlease install `datasets` to use this function: `pip install datasets`.r   )BinaryCrossEntropyLoss)CrossEntropyLoss)CrossEncoderTrainerc                    U $ r"   r   )batchs    r+   identityFitMixin.fit.<locals>.identity  s    Lr-   
batch_size   dataset	sentence_TFlabelc                 Z   Sn Sn[        U 5      R                  5       (       a  [        [        [        U 5      R	                  5       5      5      S:w  aY  SU 3n US-  n[        U 5      R                  5       (       a1  [        [        [        U 5      R	                  5       5      5      S:w  a  MY  U $ )Nzcheckpoints/modelr   r   zcheckpoints/model_)r   existsr   rk   iterdir)dir_nameidxs     r+   _default_checkpoint_dir-FitMixin.fit.<locals>._default_checkpoint_dir!  s    *HCx.''))c$tH~7M7M7O2P.QUV.V/u5q x.''))c$tH~7M7M7O2P.QUV.V Or-   z4.41.0eval_strategyevaluation_strategy)r%   batch_samplerper_device_train_batch_sizeper_device_eval_batch_sizenum_train_epochsNrf   no)
eval_stepsmax_grad_normfp16disable_tqdmsave_strategybiaszLayerNorm.biaszLayerNorm.weightc              3  ,   >#    U  H	  oT;   v   M     g 7fr"   r   .0ndns     r+   	<genexpr>FitMixin.fit.<locals>.<genexpr>E       C_V^PR!GV^   paramsweight_decayc              3  ,   >#    U  H	  oT;   v   M     g 7fr"   r   r   s     r+   r   r   H       <Xx1Wxr           	schedulerwarmup_stepst_totalr   )r   )r   r   )r>   r:   train_dataseteval_datasetr   r&   
optimizers	callbacks)rH   rE   r   )2r   ImportErrorAsentence_transformers.cross_encoder.losses.BinaryCrossEntropyLossr   ;sentence_transformers.cross_encoder.losses.CrossEntropyLossr   +sentence_transformers.cross_encoder.trainerr   
collate_fnr0   ri   r   r   NO_DUPLICATEShasattrr   r   GROUP_BY_LABELBATCH_SAMPLERr   textsr   r   	from_dict	enumerateset	TypeError
add_columnr   parsetransformers__version__r   rk   named_parametersanyrE   r   intr   _get_schedulerr   r   r~   appendrU   ru   add_callbackr   train)0r)   train_dataloaderr&   epochsr   activation_fctr   r   optimizer_classoptimizer_paramsr   evaluation_stepsr[   r'   r   use_amprx   show_progress_barr   r   r   r   r   r   r   r   r   examplebatch_textsbatch_labelsr   textr   add_label_columnr   eval_strategy_keyr:   param_optimizerno_decayr   poptimizer_grouped_parameters	optimizersteps_per_epochnum_train_stepsscheduler_objr   ra   s0                                          `        r+   fitFitMixin.fit   s   J %&&ghh 	m`S	 '/#-|Q?
&(>??)77M%y11jAQAYAY[o6p6p)88M)77M%E(+\a-b\aQXw}}gmm.L\a-b(c%K[ El"F &  ))T]^ach^iTj*kTjysDYse+<d+BTj*kl	6{qc!#(  )44WfEM	 }}\556'--:QQ & 	
 - 
.0'(2'1#
 ".>.JO_bcOc7im
 ('..
" t4467A *9`AC_V^C_@_1` , '6Yoda<Xx<X9XoYkno(
$ $$@UDTU	i%%!-0J>O!/F":;O/>>YSbM {{%%*1$U+DO#D>ZH 	 .yFG#  !1(I!FG%'!=1	
 "  !2;?![\{ .c +l  		R a Zs6   "M4
)M9
M? 
!N/N!N&N?
NNc                   [        [        US   R                  5      5       Vs/ s H  n/ PM     nn/ nU H]  n[        UR                  5       H&  u  pgX6   R	                  UR                  5       5        M(     UR	                  UR                  5        M_     U R                  " USSSS.6nU R                  b"  US   R                  S   U R                  ::  d   e[        R                  " X@R                  R                  S:X  a  [        R                  O[        R                  S9R!                  U R"                  R$                  5      nU H,  n	X   R!                  U R"                  R$                  5      X'   M.     X4$ s  snf )	Nr   Tlongest_firstr   r   	input_idsr   )dtype)ranger   r   r   r   stripr   r   
max_lengthshapetorchtensorr   r   rJ   r   r   r>   r   )
r)   r   rg   r   r   r  r   r  	tokenizednames
             r+   smart_batching_collateFitMixin.smart_batching_collater  s>   "3uQx~~#678778G&w}}5	
!!$**,/ 6 MM'--(	  NN&	
	 &)K*@*F*Fq*IT__*\\\f;;;Q;QUV;VEKK\a\f\fgjjJJ
 D'o001B1BCIO    / 9s   E>c                   [        [        US   5      5       Vs/ s H  n/ PM     nnU H8  n[        U5       H&  u  pVX5   R                  UR	                  5       5        M(     M:     U R
                  " USSSS.6nU R                  b"  US   R                  S   U R                  ::  d   eU H,  nXx   R                  U R                  R                  5      Xx'   M.     U$ s  snf )Nr   Tr   r   r  )r  r   r   r   r  r   r  r  r   r>   r   )	r)   r   rg   r   r  r   r  r  r  s	            r+    smart_batching_collate_text_only)FitMixin.smart_batching_collate_text_only  s    "3uQx=12112G&w/	
!!$**,/ 0  NN	
	 &)K*@*F*Fq*IT__*\\\D'o001B1BCIO  # 3s   Cc           
     
  ^ U R                   Ul        U(       a`  [        5       (       a)  [        R                  R
                  R                  5       nO([        R                  R
                  R                  5       nUb  [        R                  " USS9  SU l
        [        [        U5      U-  5      n[        U R                  R                  5       5      n/ SQnU V^Vs/ s H$  u  mn[!        U4S jU 5       5      (       a  M"  UPM&     snnU
S.U V^Vs/ s H$  u  mn[!        U4S jU 5       5      (       d  M"  UPM&     snnS	S./nU" U40 U	D6n[#        U[$        5      (       a  [&        R(                  " UXgUS
9nUcD  U R*                  R,                  S:X  a  [.        R0                  " 5       O[.        R2                  " 5       nSn[5        USU(       + S9 GH  nSnU R                  R7                  5         U R                  R9                  5         [;        USSU(       + S9 GH  u  nnU(       GaN  [        R<                  " U R                  R>                  R@                  S9   U R                  " S0 UDSS0D6nU" URB                  5      n U R*                  R,                  S:X  a  U RE                  S5      n U" U U5      n!SSS5        WRG                  5       n"URI                  W!5      RK                  5         URM                  U5        [        R.                  RN                  RQ                  U R                  RS                  5       U5        URU                  U5        URW                  5         URG                  5       U":g  nOU R                  " S0 UDSS0D6nU" URB                  5      n U R*                  R,                  S:X  a  U RE                  S5      n U" U U5      n!U!RK                  5         [        R.                  RN                  RQ                  U R                  RS                  5       U5        URU                  5         UR7                  5         U(       d  URU                  5         US-  nUc  GMM  US:  d  GMV  UU-  S:X  d  GMb  U RY                  X,UUUU5        U R                  R7                  5         U R                  R9                  5         GM     Uc  GM  U RY                  X,UUSU5        GM     gs  snnf s  snnf ! , (       d  f       GNF= f)a  
Deprecated training method from before Sentence Transformers v4.0, it is recommended to use
:class:`~sentence_transformers.trainer.CrossEncoderTrainer` instead. This method should
only be used if you encounter issues with your existing training scripts after upgrading to v4.0.

This training approach uses a DataLoader and Loss function to train the model.

Args:
    train_dataloader: The DataLoader with InputExample instances
    evaluator: An evaluator (sentence_transformers.cross_encoder.evaluation)
        evaluates the model performance during training on held-
        out dev data. It is used to determine the best model
        that is saved to disk.
    epochs: Number of epochs for training
    loss_fct: Which loss function to use for training. If None,
        will use BinaryCrossEntropy() if self.config.num_labels == 1
        else CrossEntropyLoss(). Defaults to None.
    activation_fct: Activation function applied on top of logits
        output of model.
    scheduler: Learning rate scheduler. Available schedulers:
        constantlr, warmupconstant, warmuplinear, warmupcosine,
        warmupcosinewithhardrestarts
    warmup_steps: Behavior depends on the scheduler. For
        WarmupLinear (default), the learning rate is increased
        from o up to the maximal learning rate. After these many
        training steps, the learning rate is decreased linearly
        back to zero.
    optimizer_class: Optimizer
    optimizer_params: Optimizer parameters
    weight_decay: Weight decay for model parameters
    evaluation_steps: If > 0, evaluate the model using evaluator
        after each number of training steps
    output_path: Storage path for the model and evaluation files
    save_best_model: If true, the best model (according to
        evaluator) is stored at output_path
    max_grad_norm: Used for gradient normalization.
    use_amp: Use Automatic Mixed Precision (AMP). Only for
        Pytorch >= 1.6.0
    callback: Callback function that is invoked after each
        evaluation. It must accept the following three
        parameters in this order: `score`, `epoch`, `steps`
    show_progress_bar: If True, output a tqdm progress bar
NTrY   iigr   c              3  ,   >#    U  H	  oT;   v   M     g 7fr"   r   r   s     r+   r   #FitMixin.old_fit.<locals>.<genexpr>  r   r   r   c              3  ,   >#    U  H	  oT;   v   M     g 7fr"   r   r   s     r+   r   r$    r   r   r   r   r   FEpoch)descdisabler   	Iterationg?)r'  	smoothingr(  )device_typereturn_dictr   r   )-r  r   r   r  npuamp
GradScalercudar\   r_   
best_scorer   r   rk   r>   r   r   ri   rE   r   r   r   r   r
   BCEWithLogitsLossr   r   	zero_gradr   r   autocastr   typer   r   	get_scalescalebackwardunscale_utilsclip_grad_norm_
parametersstepupdate_eval_during_training)#r)   r   r&   r   r   r   r   r   r   r   r   r   r[   r'   r   r   rx   r   scalerr  r  r  r   r	  r
  r  skip_schedulerre   training_stepsfeaturesr   model_predictionsr   
loss_valuescale_before_steps#                         `            r+   old_fitFitMixin.old_fit  s   ~ '+&A&A#%''113224"KKd3"c"23f<= tzz::<=A *9`AC_V^C_@_1` , '6Yoda<Xx<X9XoYkno(
$ $$@UDTU	i%%+::YSbI 151G1G11Lr++-RTReReRgHF>O:OPENJJ  "JJ$( {dPaLa% & DJJ4E4E4J4JK,0JJ,T,Tt,T)!/0A0H0H!I;;11Q6%+[[_F%-ff%=
 L )/(8(8(:%LL,557OOI.HHNN224::3H3H3JMZKK	*MMO%+%5%5%7;L%LN(,

(PX(P4(P%+,=,D,DEF{{--2!'R!)&&!9J'')HHNN224::3H3H3JMZNN$##%%NN$!#(-=-AnWgFgklFl..!X` JJ((*JJ$$&U%X $**9?TY[]_ghe Q# a Z. LKs%   !S0S0!S69S6 AS<<
Tc                   Ub  U" XXES9nUb	  U" XtU5        [        U[        5      (       a/  [        US5      (       a  UR                  U;   a  XqR                     nXpR                  :  a   Xpl        U(       a  U R                  U5        gggg)z#Runs evaluation during the trainingNrd   r6   )ri   rj   r   r6   r1  r9   )r)   r&   r[   r'   re   rf   rx   scores           r+   r?  FitMixin._eval_during_training8  s     d5VE#u-%&&79>N+O+OT]TlTlpuTu667&"'"IIk* # ' !r-   )r1  )"r)   r   r   r   r&   rF   r   r   r   rE   r   r   r   type[Optimizer]r   dict[str, object]r   rJ   r   r   r[   rs   r'   rG   r   rJ   r   rG   rx   r|   r   rG   rH   rI   )r   list[InputExample]rH   ztuple[BatchEncoding, Tensor])r   rN  rH   r   ) r   r   r&   rF   r   r   r   rE   r   r   r   rL  r   rM  r   rJ   r   r   r[   rs   r'   rG   r   rJ   r   rG   rx   r|   r   rG   rH   rI   )rH   rI   )rL   rM   rN   rO   rP   r
   r   r  optimAdamWr  r  r   rG  r?  rQ   r   r-   r+   r   r      s$   _
 /3{{}'!+0;;+<+</3Tl" !"& $ 6:"&%{{${ ,{ 	{ { { ){ ,{ { {  { { {  !{" 4#{$  %{& 
'{z!4. /3{{}'!+0;;+<+</3Tl" !"& $ 6:"&%Vi$Vi ,Vi 	Vi Vi Vi )Vi ,Vi Vi Vi  Vi Vi Vi  !Vi" 4#Vi$  %Vi& 
'Vip+r-   r   );
__future__r   loggingr\   pathlibr   typingr   r   r   r  r   	packagingr   r	   r
   torch.optimr   torch.utils.datar   tqdm.autonotebookr   r   r   r   r   r   $transformers.tokenization_utils_baser   1sentence_transformers.cross_encoder.training_argsr   5sentence_transformers.datasets.NoDuplicatesDataLoaderr   3sentence_transformers.datasets.SentenceLabelDatasetr   2sentence_transformers.evaluation.SentenceEvaluatorr   sentence_transformers.readersr   )sentence_transformers.SentenceTransformerr   #sentence_transformers.training_argsr   sentence_transformers.utilr   datasetsr   0sentence_transformers.cross_encoder.CrossEncoderr   *sentence_transformers.readers.InputExample	getLoggerrL   loggerr   rU   ru   Moduler~   r   r   r-   r+   <module>rh     s    "  	  / /     ! ' * ^ ^ > [ X T P 6 I = < MG 
		8	$%4 %4P&g &gRL L2299 DQ+ Q+r-   