i0ddlZddlZddlZddlmZmZddlmZddlm Z m Z m Z m Z ddl mZddlmZmZddlmZmZmZmZmZmZmZmZmZmZmZmZddlm Z dd l!m"Z"m#Z#e$ejJZ&e&Dcgc] }| c}Z'e'ed ed ed ededededededededg z Z'dZ( ddZ)dZ*ejVjYde'e(dZ-egdej\ ej\gfegdej\ ej\gfegdej\ ej\gfegdej\ ej\gfeddgej\ ddej\gfeddgej\ dddej\gfedddgej\ ej\gfedddgej\ ej\gfedddgej\ ddej\gfedddgej\ dddej\gfedddgej\ dddej\gfedddgej\ ej\gfedgdej\ ej\gfedddgej\ ddej\gfedddgej\ dddej\gfedddgej\ dddej\gfegd ej\ d!dej\gfegej\ d!d"ej\gfgZ/edggfedggfedgd#gfedd$dggfeddggfedgd#gfedd$dggfeddggfeddggfegd%gfg Z0egdgfedggd&fedggd&fedgdgfedggd&fedgd&gfedgdgfegddgfedd'gddgfg Z1ejVjYd(e/e0zd)Z2ejVjYd*e/e1zd+Z3ejVjYd,ged-d.d/d0dfed-d.d1d-dfed' d-d.dd'dfed d-d.dd dfed d.d-d-d2dfed'd3d-d.d4ddfed'd3d-dd5ddfed5ejhd0d0dejhd0zz dd0fed5ejhd0ejhd0d'zd'd'fedd5ejhd0d6ddfedd5d1ddejhdzz ddfedd5d1ejhdd'z ddfedd5d1d7ddfed ejhd0ejjd0d ejhd0zz ddfed8d9dddfed-d9d:d!dfed8d;dddfed-d;dd!dfed-d?d@dAdBfed8d?gdBgdzed-dCdDdAdEfed8dCgdEgdzed8dFd@dGdBfed-dFgdHgdzed8d>d>ddIfed8dJdJddKfed-d:dddfed8d:d:ddfed-dJddLdKfed-dMdddfed-dejVjYde'e(ejVjYdZdd_gdgZ?ejVjYdhgdie#djZ@ejVjYde'e(ejVjYdZdd_gdkZAejVjYdleejdmfeejdmfed dndmfeejdofeejdpfeejdpfeejdqfgdrZDdsZEdtZFduZGejVjYdvejgdwejgdxfejVjYdyejgdzejgd{fd|ZIejVjYde'e(d}ZJejVjYde&ejVjYdZdd_gejVjYd~ejnejpfejVjYdddZKejVjYde&ejVjYdd~ejidejdfgdZMejVjYdeddieNdfeddieOdfedd"ieOdfeddieNdfeddieOdfedd"ieOdfgdZPejVjYde'e(dZQejVjYdgddZRycc}w)N)assert_allcloseassert_array_equal)approx)LinearConstraintminimizeminimize_scalarnewton) logsumexp) IdentityLink_inclusive_low_high) _LOSSES AbsoluteErrorBaseLossHalfBinomialLoss HalfGammaLossHalfMultinomialLossHalfPoissonLossHalfSquaredErrorHalfTweedieLossHalfTweedieLossIdentity HuberLoss PinballLoss)assert_all_finite)create_memmap_backed_data skip_if_32bitg?)quantileg?power@ct|tr|}|jj}t|tr|d|j j dz }|St|tr|d|j z }|St|dr3t|j dr|d|j jdz }|St|S)Nz (quantile=)clossrz(power=) isinstancer __class____name__rr%rrhasattrrstr)paramlossnames /var/www/ITT-ServiceAPI-Services/ITT-Document-Scan/mrzscanner/mrzenv/lib/python3.12/site-packages/sklearn/_loss/tests/test_loss.pyloss_instance_namer/4s%"~~&& dK ( j!4!4 5Q7 7D  i ( j0 0D T7 # G(D gdjj../q1 1D 5zctjj|}|jrtj||j f}|j |d|d||j z|jddtj|jt|j z}||fSt|jtrPt|j\}} tj ||dg}tj"| |dg} || f}|j |d|d|}t|j$\}} t'||d}t)| |d} |j || |}|j$j*dk(r!|j$j,r d|dd|dz<|j$j.dk(r!|j$j0r d|dd|dz<||fS)z9Random generate y_true and raw_prediction in valid range.rr )lowhighsizeNr4)nprandom RandomState is_multiclassempty n_classesuniformflatarangeastypefloatr&linkr r interval_y_predamaxamininterval_y_truemaxminr2 low_inclusiver3high_inclusive) r, n_samplesy_bound raw_boundseedrngraw_predictiony_truer2r3s r.random_y_true_raw_predictionrRCs ))   %C 9dnn"=>!$! 1T^^+"-" A 9%,,U3dnnD, > !!' dii .+D,@,@AIC''3 ! -.C77D)A,/0Dd I! 9Q !!r0ctj||}||d|zz }|||z }|||z}||d|zz}| d|zzd|zz |zd|zz S)z2Helper function for numerical (first) derivatives.) fill_valuer!(@)r7 full_like)funcxepsh f_minus_2h f_minus_1h f_plus_1h f_plus_2hs r.numerical_derivativer`hsy Q3'Aa!a%iJa!eJQU IQQYI JY &Z 7* DPS TTr0r,)idsc|jr2d}tjtjd|dz |d}n0t |j \}}tj||d}|j j r)tj||j jf}|j jr)tj||j jf}|j|sJ|jd}t |j\}}|jrctj|f}tj||||dddf<dd|dddfz z|dddf<dd|dddfz z|dddf<ntj|||}|j|sJ|j j!|}|j#|| y) z4Test interval ranges of y_true and y_pred in losses.r6rr num N?r!rQrP)r:r7tilelinspacer rFrIr_r2rJr3in_y_true_rangeshaperCr;in_y_pred_rangerBr,)r,r<rQr2r3ny_predrPs r.test_loss_boundaryrpvs  Q A 9EqI'(<(<= TS$B/ ))vt337778 **vt338889    '' ' QA#D$8$89IC 1i.){{3!4q!t a&A,./q!t a&A,./q!t S$A.    '' 'YY^^F+NIIVNI;r0)r皙?drrrs皙?r6)rtrurrrrs)rrrf?g?)rqrurrq)?r!)rtrurrfz!loss, y_true_success, y_true_failc|D])}|jtj|gr)J|D])}|jtj|gs)Jy)z-Test boundaries of y_true for loss functions.N)rkr7array)r,y_true_success y_true_failys r.test_loss_boundary_y_truer^ 3##BHHaSM2223 7''! 6667r0z!loss, y_pred_success, y_pred_failc|D])}|jtj|gr)J|D])}|jtj|gs)Jy)z-Test boundaries of y_pred for loss functions.N)rmr7r|)r,y_pred_success y_pred_failrs r.test_loss_boundary_y_predrrr0zDloss, y_true, raw_prediction, loss_true, gradient_true, hessian_truerzg@rUg@gп)rdeltag@@gȿg?ryg@xg@xDg@g@@gBgB@gB%ggT`Grc|tj|gtj|g}|jtj|gtj|g}|jtj|gtj|g\}} |j tj|gtj|g\} } |t |ddk(sJ|t |ddk(sJ|9|t |ddk(sJ| t |ddk(sJ| t |ddk(sJ|| t |ddk(sJyy)z7Test losses, gradients and hessians at specific values.rgV瞯<)relabsN)r7r|gradient loss_gradientgradient_hessianr) r,rQrP loss_true gradient_true hessian_trueloss1grad1loss2grad2grad3hesss r.test_loss_on_specific_valuesrsbx &*288^DT;U VE MMxx!"((N;K2L  E%%xx!"((N;K2L&LE5''xx!"((N;K2L(KE4 F9%U; ;; ; F9%U; ;; ; }%UCCCC}%UCCCC}%UCCCCvl5AAAA r0readonly_memmapFTdtype_in dtype_out sample_weightout1out2 n_threadsc|}d}t||ddd\} } | j|} | j|} |tjdg|z|}|tj| |}|tj| |}|r#t | } t | } | t |}|j | | ||| } || |usJ |j| | ||| } || |usJ |j| | |||| \} } || |usJ || |usJ |#|jrtj| |}|j| | |||| \} } || |usJ || |usJ || | | |j| ||j| |t|dr|j| t|dr-|j| | ||||\} }|| |usJ |||usJyyy)a0Test acceptance of dtypes, readonly and writeable arrays in loss functions. Check that loss accepts if all input arrays are either all float32 or all float64, and all output arrays are either all float32 or all float64. Also check that input arrays can be readonly, e.g. memory mapped. rqrsre*r,rKrLrMrNNrdtype)rQrPrloss_outr)rQrPr gradient_outr)rQrPrrrr)rQrPrr hessian_outrrQrPrrQr predict_proba)rPgradient_proba)rQrPrr proba_outr)rRr@r7r| empty_likerr,rrr:rfit_intercept_onlyconstant_to_optimal_zeror)rr)r,rrrrrrrrKrQrPlgr[ps r.test_loss_dtyperxs$ 6DI9   FN]]8 $F#**84N #!2(C  }}V95 }}^9=*622>B  $5mDM %#  A(192 2d %#  A(192 2d   %#  DAq(192 2d(192 2d D..}}^9=  %# ! DAq(192 2d(192 2d~]S6G!!}!Mt_% .9t%&"")' # 1!,qDy66$ ,qDy66$y'r0rangect|dddd\}}|dk(r2tjd|jd|jd }tj|}tj|}tj|}tj|}tj|}tj|} |j |||| |j j |||| t|||j|||| |j j|||| t|||j j||||| |j j||||| t||t|||j||||| |j j||||| t||t|| y)z:Test that Python and Cython functions return same results.rrrrrr rrcrQrPrrrQrPrrrQrPrrrrQrPrrrN) rRr7rirlrr,r%rrrr) r,rrQrPout_l1out_l2out_g1out_g2out_h1out_h2s r.test_loss_same_as_C_functionsrs:   FN Av||AFLLOL ]]6 "F ]]6 "F ]]> *F ]]> *F ]]> *F ]]> *FII%#   JJOO%#  FF#MM%#   JJ%#  FF#JJ%#  JJ%# FF#FF#%#  JJ%#  FF#FF#r0cFt|ddd|\}}|dk(r2tjd|jd|jd}tj|}tj|}tj|}tj|}tj|} tj|} |j |||| } |j |||| } |j||||| \} }|j|||| | \}}t| | t| |tj| |sJt| |tj| |sJt| |t| |t| |tj| |sJt||tj||sJt|| tj|| sJt|d r|jsJtj|}tj|}|j|||||\}}t| |t| |t||ttj|dddyy)zTest that loss and gradient are the same across different functions. Also test that output arguments contain correct results. rrrrrr rrcrrrrrrQrPrrraxisdy=)rtolN)rRr7rirlrr,rrrrr shares_memoryr)r:rsum)r,rglobal_random_seedrQrPrrrrout_g3out_h3l1g1l2g2g3h3out_g4 out_probag4probas r. test_loss_gradients_are_the_samer,s:   FN Av||AFLLOL ]]6 "F ]]6 "F ]]> *F ]]> *F ]]> *F ]]> *F %#  B %#  B   %#  FB " "%# #FBBr6"  B '' 'r6"  B '' 'BBr6"  B '' 'r6"  B '' 'r6"  B '' 't%&!!!!~.MM.1 '')' ( E F#By)u1-qu='r0onesr8c nd}t||dd|\}}|dk(r&tj|tj}nNtjj |}|j |jtj}t|j|||||j||d z|j||d \}}|j|||\} } t||z| |jst||z| nt||d d d fz| |j||d \}} |j|||\} } |jst||z| t| |z| y t||d d d fz| t| |d d d fz| y ) zTest sample weights in loss, gradients and hessians. Make sure that passing sample weights to loss, gradient and hessian computation methods is equivalent to multiplying by the weights. rsrrrr)rlrr5rN) rRr7rfloat64r8r9normalr@rr,rr:r) r,rrrKrQrPrOlossesr losses_sw gradient_swhessian hessian_sws r.test_sample_weight_multipliesr}sI9   FNirzzB ii##$67 299"**E  )'   )))   ))%*FH "//%#0I{ F]*I6   =0+>=D#99;G--%.Hg #33%#4K   =0+>-/<=D#99;G-4"88*Er0ct|dddd\}}|jdk(r|dddf}t|j|||j||t|j |||j ||t|j |||j ||t|j |||j ||yy) z5Test that reshaped raw_prediction gives same results.rrrrrr Nrg)rRndimrr,rrr)r,rQrPraw_prediction_2ds r.test_graceful_squeezingrs:   FNa*1d73 IIV4EI F IIVNI C     f=N  O   f^  L   MM8IM J MMM G    ! !@Q ! R  ! ! ! O  r0c|jstjgd}t|jt r{d}|j j}|j js||z}|j j}|j js||z }tj|||}|jj|}ntj|jjt }tj"|j|jftj$d t }tj$d|j&dd|jdz<|dk(r2tj(d|j*d|j*d }|j-||| }|j/|| }t1|| d d y)z~Test value of perfect predictions. Loss of y_pred = y_true plus constant_to_optimal_zero should sums up to zero. )rrurrrr6re绽|=re)rlrTrNr rrrcrr+=r)atolr)r:r7r|r&rBr rCr2rIr3rJclipinverser?r<r@rAfullexpr>rirlr,rr) r,rrPrZr2r3rQ loss_value constant_terms r.test_loss_of_perfect_predictionrs   "<= dii .C&&**C''55Ci'',,D''66czWW^S$?N"">24>>*11%8>>4>>2r {  68VVBZ1t~~112 Av||AFLLOL %#J 11]2M JUGr0c" d}t|dd|\ dk(r2tjd jd jdj  \}}|j jk(sJ|j jk(sJj sV fd }t | d }t||d d fd}t | d } jryt|| d dytjD]w  fd}t | dd fd }t|dd f|d d  fd}t | dd fd } jrbt|dd f| d dyy)zTest gradients and hessians with numerical derivatives. Gradient should equal the numerical derivatives of the loss function. Hessians should equal the numerical derivatives of gradients. rrrrrr rrcrc,j|SNrr,rYr,rrQs r. loss_funcz6test_gradients_hessians_numerically..loss_func4s"99 + r0gư>)rZh㈵>rrrc,j|Srrrs r. grad_funcz6test_gradients_hessians_numerically..grad_func>s"== +! r0c^j}||ddf<j|Sr)copyr,rYrawkr,rPrrQs r.rz6test_gradients_hessians_numerically..loss_funcSs=$))+AqD yy!#&"/!r0Ngh㈵>clj}||ddf<j|ddfSr)r rr s r.rz6test_gradients_hessians_numerically..grad_func_sK$))+AqD }}!#&"/%Q$ r0) rRr7rirlrr:r`rapprox_hessianrr<) r,rrrKrr[r g_numericr h_numericr rPrQs `` @@@r.#test_gradients_hessians_numericallyrsI9   FN Av||AFLLOL  %# ! DAq 77n** ** * 77n** ** *    )NM 94e< )NM     Ayt% @t~~& KA  -Yq!t8LRVWI AadGYT F  -Yq!t8LRVWI""!Q$EJ9 Kr0zloss, x0, y_true) ) squared_errorgr)rg@]@g?)rryry) binomial_lossrrr)rir)rrw) poisson_lossrVrz)rryr)rg6g$@ctdtjgtjtj|gtj}dtjdtjffd }dtjdtjffd }dtjdtjffd }t ||||d d }j |j }tjj|t||d d tj|d dy)acTest that gradients are zero at the minimum of the loss. We check this on a single value/sample using Halley's method with the first and second order derivatives computed by the Loss instance. Note that methods of Loss instances operate on arrays while the newton root finder expects a scalar or a one-element array for this purpose. NrrrYreturncPj|jzS)zCompute loss plus constant term. The constant term is such that the minimum function value is zero, which is required by the Newton method. rgrQ)r,rrYr,rQs r.rXztest_derivatives..funcs6 yy!  ) ) ) 89 9r0c*j|S)Nrgrrs r.fprimez test_derivatives..fprimes}}F1}==r0c0j|dS)Nrgr )rrs r.fprime2z!test_derivatives..fprime2s$$F1$EaHHr0rsgHj>)x0rrmaxitertolrrrrggƠ>) r r7r|rndarrayr ravelrrBrr)r,r rQrXrroptimums` ` r.test_derivativesr'ps 4 4=t ,D XXvhbjj 1F 2$bjj )B9 9rzz9>"**>>I2::I"**I   G\\^FmmoGDII%%g.7DM151DMMMH!RVWr0c djs2jjtjddnGtj j tjjzdddd<dk(rtjd d j }fd }jst|d ddi}jtj|}|jtk(sJ|jjk(sJt!||t#|j$d k(|j't#ddk(yt)|tj*jdddidt-tj.djfdd}jtj0|df}|jjk(sJt!|t3||j$ddt3|j'dddy)zzTest that fit_intercept_only returns the argmin of the loss. Also test that the gradient is zero at the minimum. 2rrcrNrrrrr!rcjstj|}n6tjtj|j f}|S)N)rlrTrlr)r:r7rascontiguousarray broadcast_tor<)rYrPr,rKrrQs r.funz%test_loss_intercept_only..funs[!!WWI1EN11)T^^)DEN)'  r0gHz>r!rs)r"optionsrrg-q=)rgvIh%<=SLSQPr )r"r0method constraintsrrrr#)r:rBrr7rir?r@rr<rrrrWrltuplerrrrYrrzerosrrrhr)r,rar/optgradrKrQs`` @@r.test_loss_intercept_onlyr:sI   ""2;;r1)#DE9%,,RZZ84>>Iss  C :  v]KA     cti5EF}}<<2'  ww%'!!!ww&,,&&&! VCEEt $$  fQE**  HHdnn &$(!T^^1D)Eq!L  }}771y!n5'  ww&,,&&&!355t%8a(!%8r0zloss, func, random_distrc0tj|dS)N)q)r7 percentile)rYs r.r?sr}}Q"/Er0poisson exponentialbinomialc^tjj|}|dk(r|jddd}nt ||d}|j |}t ||t|jj||k(sJ|jj|t||k(sJt|tr%t|jj|||jjrB|j|jj |j |}t ||jj"rC|j|jj$|j |}t |yy)zTest that fit_intercept_only returns the correct functional. We test the functional for specific, meaningful distributions, e.g. squared error estimates the expectation of a probability distribution. rBr rfrsr5rN)r7r8r9rBgetattrrrrrBrr&r rrFrIfillr2rJr3)r,rX random_distrrOy_trainbaseline_predictions r. test_specific_fit_intercept_onlyrIsf$ ))   2 3Cj ,,q#C,0+'#{+5111A)* &W )F"G GG G 99  0 1VDM5J JJ J$ % ))*=>@ST )) T))--."55W5E-. ** T))../"55W5E-.+r0c tjjd}d}t|}|j d|dzdj tj }|j|}|j|fk(sJtj||j}t|D]}||k(j||<t|tj|tjtj|z t|d d d f|jj|d d d ftjd tj d fD]Y}|j tj }|j|}|j|jk(sJt#|[y ) zATest that fit_intercept_only returns the mean functional for CCE.rrrr rsr5rrNrer,)r7r8r9rrandintr@rrrlr6rrmeanrlogrBrr)rOr<r,rGrHrr s r.(test_multinomial_loss_fit_intercept_onlyrNsu ))   "CI  3Dkk!Y]k5<.alt_lossus; EAIwwrvva"&&!h"77899r0cPd|zdz }| dtj||zzz SrZ)r7rr[s r. alt_gradientz>test_binomial_vs_alternative_formulation..alt_gradientys. EAIrQH --..r0N)rr@rrr)rQro global_dtyper^r`bin_lossdatums r.(test_binomial_vs_alternative_formulationrddsv":/  !H ]]< (F ]]< (F V E U vh&67 77 7%H%%u-|U/CDr0c d}t||dd|\}}t|drU|j|}|j||jfk(sJt j |dtdd k(sJt|d rd d t j|ft j|d ft j|t j|ffD]\}}|j||d || \}}|j||jfk(sJt j |dtdd k(sJt||j||d d y y )z2{{y$..9999vve!$qe(<<<<t%&  2==0 1 ]]> *D 1 ]]> *BMM.,I J   KD% ---"! .KD%;;9dnn"== ==66%a(F1%,@@ @@  !#1"&!%   'r0rorder)CFcd}|dk(rtj|}||}|j|||\}}|jr#|j|fk(sJ|jdk(slJ|j r:|j||j fk(sJ|j||j fk(s&J|j|fk(sJ|j|fk(sJ|j|k(sJ|j|k(sJ|dk(r1|jjsJ|jjsJy|jjsJ|jjsJy)zTest that init_gradient_and_hessian works as expected. passing sample_weight to a loss correctly influences the constant_hessian attribute, and consequently the shape of the hessian array. rrr)rKrrg)r rhN) r7rinit_gradient_and_hessianconstant_hessianrlr:r<rflags c_contiguous f_contiguous)r,rrrgrKrrs r.test_init_gradient_and_hessiansrps[I * m ,D667Hg  ~~)---}}$$$  ~~)T^^!<<<<}}DNN ;;;;}} ,,,}} ,,, >>U "" " ==E !! ! |~~****}}))))~~****}}))))r0zparams, err_msgz+Valid options for 'dtype' are .* Got dtype=z instead.c|}tjttf|5|jdddi|\}}dddy#1swYyxYw)zDTest that init_gradient_and_hessian raises errors for invalid input.matchrKrN)pytestraises ValueError TypeErrorrk)r,paramserr_msgrrs r.%test_init_gradient_and_hessian_raisesr{sV 6D  I.g >R:D::QQQ&Q'RRRs A  Azloss, params, err_type, err_msgrz4quantile must be an instance of float, not NoneType.zquantile == 0, must be > 0.zquantile == 1.1, must be < 1.cltj||5|di|dddy#1swYyxYw)z/Test that loss raises errors for invalid input.rrNrt)rurv)r,ryerr_typerzs r.#test_loss_init_parameter_validationr~s1B xw / vs *3cd}t||ddd\}}tj|}tj|}|||t |||k(sJy)z Test that losses can be pickled.rrrrrrgN)rRpickledumpsloadsr)r,rKrQrP pickled_lossunpickled_losss r.test_loss_picklerspI9   FN<<%L\\,/N vn =f^DB  r0r)rrr rvr!r6ct|}t|}d}t||d\}}|jj |}|j |||j |z}|j |||j |z}t|||j||\} } |j||\} } t| || zt| || z|dz| zzy)zCTest for identical losses when only the link function is different.rrerrPrgr!N) rrrRrBrr,rrr) rhalf_tweedie_loghalf_tweedie_identityrKrQrProloss_log loss_identity gradient_log hessian_loggradient_identityhessian_identitys r.%test_tweedie_log_identity_consistencyr$s2'Q/3!<I9 FN " " * *> :F $$n%11&9:H*..f/66v>?MHm,!1 A An!B!L++@*P*Pf+Q+''L&+<"<=V//&!)>N2NNr0)rrr)Srnumpyr7ru numpy.testingrrrscipy.optimizerrrr scipy.specialr sklearn._loss.linkr r sklearn._loss.lossr rrrrrrrrrrr sklearn.utilsrsklearn.utils._testingrrlistvalues ALL_LOSSESLOSS_INSTANCESr/rRr`mark parametrizerpinfY_COMMON_PARAMS Y_TRUE_PARAMS Y_PRED_PARAMSrrrMlog1prfloat32rrrrrrrrr'r:rLmedianrIrNrRrWr|rdrfrpint64r{rxrwr~rrrs0r.rs  = $@    ,K .'.." # %/0T$&0 t$!!!#!$!$!$#&    CE""J U5GH <I 1SzRVVGRVV+<=3#srvvgr4-HI1SzRVVGRq"&&+IJ1SzRVVGRq"&&+IJ2&c bffWbff4EF1%'>"&&"&&@QR3'#srvvgr45PQ1%SzRVVGRq"&&3QR1%SzRVVGRq"&&3QRBFF7B266*BCB"&&"c266 :;'2R [1#r241ay"-3!b)2&<1%ay"53'!b1!Q$M2.  QC 2M21r=13aS)2&M:1%}b93'aS1aV$S#JA/  'm#7 7'm#7 7Js  S#q!Q/s #sCd3s c "Cad;s d #S#~xN s d #S#xE s Cq )3_dDQ s Cq )3ZtLs  CAFBFF1I ,=uaHs #vrvvay&"&&)e*;YNs q !3q 3DdDQs !q )3QVRVVAY5FdSs !q )3VRVVAY5FdSs $! ,   )    s,    BFF1I BHHQK$* *    +sr  S%Aq1sst  S%r15usv  S$1a0wsx  S$R3ysz  S%r15{s|  S%Y8MN}s~  S%F+@*AA*EFs@  S%y:OPAsB  S%F+@*AA*EFCsD  S$I7LMEsF  S$E*?)@1)DEGsH  S$a1FGIsJ  S%3JKKsL  S$1a0MsN  S$a3OsR     $ #  Qs`  S%Aq1asb  S#q!Q/csd  S#sAq1esh ! ,   o & ,    gsx ! ,   o & ,    wsH ! ,   o & ,    GsX ! ,   n % .    Wsh mwpBqwpB6,*UDM:bjj"**%=>rzz2::&>?4)4$+$+q!f-Y7.,,5@?;-Y7x5GH4/:K$;IK$\5GH4/:L>;IL>^5GH68*<=BF>IBFJ5GH I <5GH4/:-H;I-H`5GH4/:VK;IVKr "+X#$+X\5GH4/:>9;I>9B  RWWh/ "))X. d #%ExP  BGGY/ "''=1  BGG]3  RWWj1 / />/."<"HBHH[$98288K;P#QRHBHH^$S#TUEVSED5GH(I(V,4/:2::rzz":;*-!*.<;-!*H,bhh 9"((9 M R-R%     B    O  )  z3'5TU     B    O  )  Z%z3RS5>?> 5GHI"56)7)E(1s:w3