_ndddlZddlZddlmZmZddlZddlmZddlm Z m Z m Z m Z ddl mZddlmZmZddlmZmZmZmZmZmZmZmZmZmZmZddlmZdd l m!Z!m"Z"e#ej$Z%d e%DZ&e&ed ed ededededededededg z Z&dZ' drdZ(dZ)ej*+de&e'dZ,egdej- ej-gfegdej- ej-gfegdej- ej-gfeddgej- ddej-gfeddgej- dddej-gfedddgej- ej-gfedddgej- ej-gfed ddgej- ddej-gfedddgej- dddej-gfed!ddgej- dddej-gfedddgej- ej-gfedgd"ej- ej-gfed ddgej- ddej-gfedddgej- dddej-gfed!ddgej- dddej-gfegd#ej- d$dej-gfegej- d$d%ej-gfgZ.edggfedgd&gfedd'dggfed dggfedgd&gfedd'dggfed dggfeddggfegd(gfg Z/egdgfedggd)fedggd)fed gdgfedggd)fedgd)gfed gdgfegddgfedd*gddgfg Z0ej*+d+e.e/zd,Z1ej*+d-e.e0zd.Z2ej*+d/ed0d1d2fed0d1d3fed* d0d1dfed d0d1dfed d1d0d0fed4ej3d3d3dej3d3zz fed4ej3d3ej3d3d*zfed!d4ej3d3d5fedd4d6ddej3dzz fedd4d6ej3dd*z fed!d4d6d7fed ej3d3ej3d8d ej3d3zz fed!9d:gd;egd;dZ4ej*+de%ej*+d?d@dAgej*+dBej5ej6gej*+dCej5ej6gej*+dDddgej*+dEddgej*+dFddgej*+dGddgdHZ7ej*+de&e'ej*+dDddIgdJZ8ej*+de&e'ej*+dDddIgdKZ9ej*+de&e'ej*+dDdLdMgdNZ:ej*+de&e'dOZ;ej*+de&e'ej*+dDddIgdPZej*+de&e'ej*+dDddIgdUZ?ej*+dVeej@dWfeejAdWfed dXdWfeej@dYfeej@dZfeej@dZfeej@d[fgd\ZBd]ZCd^ZDej*+de&e'd_ZEej*+de%ej*+dDddIgej*+d`ej5ej6fej*+dadbdcZFej*+de%ej*+ddd`ejGideejGdffgdgZHej*+dhedidieIdjfedidieJdkfedid%ieJdlfgdmZKej*+de&e'dnZLej*+dogdpdqZMdS)sN)assert_allcloseassert_array_equal)approx)minimizeminimize_scalarnewtonLinearConstraint) logsumexp)_inclusive_low_high IdentityLink) _LOSSESBaseLoss AbsoluteErrorHalfBinomialLoss HalfGammaLossHalfMultinomialLossHalfPoissonLossHalfSquaredErrorHalfTweedieLossHalfTweedieLossIdentity PinballLoss)assert_all_finite)create_memmap_backed_data skip_if_32bitc"g|] }| Sr).0losss =lib/python3.11/site-packages/sklearn/_loss/tests/test_loss.py r #s000T$$&&000g?)quantilepowerg@ct|trW|}|jj}t |dr|d|jjdz }n#t |dr|d|jjdz }|St|S)Nr"z (quantile=)r%z(power=) isinstancer __class____name__hasattrclossr"r%str)paramrnames rloss_instance_namer23s%"" ~& 4 $ $ 2 7!4777 7DD T7 # # 2 1dj.111 1D 5zzr!d*ctj|}|jrtj||jf}||d|d||jz|jdd<tj| t|jz}n:t|j trSt|j\}} tj||dg}tj| |dg} || f}||d|d|}t|j\}} t'||d}t)| |d} ||| |}|jjdkr|jjr d|dd|dz<|jjdkr|jjr d|dd|dz<||fS)z9Random generate y_true and raw_prediction in valid range.rr&)lowhighsizeNr=)nprandom RandomState is_multiclassempty n_classesuniformflatarangeastypefloatr*linkr r interval_y_predamaxamininterval_y_truemaxminr; low_inclusiver<high_inclusive) r n_samplesy_bound raw_boundseedrngraw_predictiony_truer;r<s rrandom_y_true_raw_predictionr[@s )   % %C .9dn"=>>!$! 1T^+"-" " AAA 9%%,,U33dnD di . . $+D,@AAIC'3 ! -..C7D)A,/00Dd I! 9Q !!r!ctj||}||d|zz }|||z }|||z}||d|zz}| d|zzd|zz |zd|zz S)z2Helper function for numerical (first) derivatives.) fill_valuer'(@)r@ full_like)funcxepsh f_minus_2h f_minus_1h f_plus_1h f_plus_2hs rnumerical_derivativeries Q3'''Aa!a%iJa!eJQU IQQYI JY &Z 7* DPS TTr!r)idscX|jrtjddd}n.t|j\}}tj||d}|jjrtj||jjf}|jjrtj||jj f}| |sJ|j d}t|j \}}|jrgtj |df}tj||||dddf<dd|dddfz z|dddf<dd|dddfz z|ddd f<ntj|||}||sJ|j|}||| dS) z4Test interval ranges of y_true and y_pred in losses.r numr?N?r&r'rZrY)rCr@linspacer rOrRr_r;rSr<in_y_true_rangeshaperLrDin_y_pred_rangerKr)rrZr;r<ny_predrYs rtest_loss_boundaryryss 0Qr***'(<== TS$B/// )9vt3778 *:vt3889    ' '''' QA#D$899IC /1a&!!{3!444qqq!t a&A,./qqq!t a&A,./qqq!t S$A...    ' ''''Y^^F++NIIVNI;;;;;r!)r4r皙?r5rzr5皙?r?)r{r|rrzr5)rzrp?g?)r4r|rr4)?r')r{r|rrpz!loss, y_true_success, y_true_failc|D],}|tj|gsJ-|D],}|tj|grJ-dS)z-Test boundaries of y_true for loss functions.N)rtr@array)ry_true_success y_true_failys rtest_loss_boundary_y_truerz 33##BHaSMM222222 77''! 66666677r!z!loss, y_pred_success, y_pred_failc|D],}|tj|gsJ-|D],}|tj|grJ-dS)z-Test boundaries of y_pred for loss functions.N)rvr@r)ry_pred_success y_pred_failrs rtest_loss_boundary_y_predrrr!z'loss, y_true, raw_prediction, loss_truerg@r^@gȿg@g?r8rEr)皙?rp333333?rrc|tj|gtj|gt|ddksJdS)zTest losses at specific values.rqdy=-q=)relabsN)r@rr)rrZrY loss_trues rtest_loss_on_specific_valuesrsgL 4x!!"(N;K2L2L     u% 0 0 0 1 1 1 1 1 1r!readonly_memmapFTdtype_in dtype_out sample_weightout1out2 n_threadsc|}d}t||ddd\} } | |} | |} |tjdg|z|}|tj| |}|tj| |}|r5t | d } t | d } |t |d }|| | ||| || | ||| || | |||| ||j rtj| |}| | | |||||| | || | || | |t|dr|| t|dr|| | ||||dSdS)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. r8r3rmr9rrTrUrVrWNrdtypeT)aligned)rZrYrloss_outr)rZrYr gradient_outr)rZrYrrrr)rZrYrr hessian_outrrZrYrrZr predict_proba)rYgradient_proba)rZrYrr proba_outr)r[rIr@r empty_likerrgradient loss_gradientrCgradient_hessianfit_intercept_onlyconstant_to_optimal_zeror-rr) rrrrrrrrrTrZrYs rtest_loss_dtypers$ 466DI9   FN]]8 $ $F#**844N #!2(CCC  }V9555 }^9===S*64@@@2>4PPP  $5mTRRRMII%#  MM%#  %#  D.}^9===%#  D~]SSSS6GGG!!}!MMMt_%%: .999t%&&  )'        r!rangec t|dddd\}}|dkr-tjd|jd|jd }tj|}tj|}tj|}tj|}tj|}tj|} t ||||| |j|||| t ||||| |j|||| |j ||||| |j ||||| t ||t ||| ||||| |j ||||| t ||t || dS)z:Test that Python and Cython functions return same results.r3rr9rrr&rrnrZrYrrrZrYrrrZrYrrrrZrYrrrN) r[r@rrrurrrr.rrr) rrrZrYout_l1out_l2out_g1out_g2out_h1out_h2s rtest_loss_same_as_C_functionsrdss:   FN Av|AFLOLLL ]6 " "F ]6 " "F ]> * *F ]> * *F ]> * *F ]> * *F )'   )'       )'   )'       J%#  J%# FF###FF###%#  J%#  FF###FF#####r!ct|ddd|\}}|dkr-tjd|jd|jd}tj|}tj|}tj|}tj|}tj|} tj|} ||||| } ||||| } |||||| \} }||||| | \}}t| | t| |tj | |sJt| |tj | |sJt| |t| |t| |tj | |sJt||tj ||sJt|| tj || sJt|d r|j sJtj|}tj|}||||||\}}t| |t| |t||ttj|ddddSdS)zTest that loss and gradient are the same across different functions. Also test that output arguments contain correct results. rr3rrrr&rrnrrrrrrZrYrrraxisr)rtolN)r[r@rrrurrrrrrr shares_memoryr-rCrsum)rrglobal_random_seedrZrYrrrrout_g3out_h3l1g1l2g2g3h3out_g4 out_probag4probas r test_loss_gradients_are_the_samers/:   FN Av|AFLOLLL ]6 " "F ]6 " "F ]> * *F ]> * *F ]> * *F ]> * *F %#   B %#   B   %#  FB " "%# #FBBr6""" B ' ''''r6""" B ' ''''BBr6""" B ' ''''r6""" B ' ''''r6""" B ' ''''t%&&>!!!!~..M.11 '')' (  E F###By)))u1---qu======>>r!onesrAc d}t||dd|\}}|dkr!tj|tj}nRtj|}||tj}t| ||||| ||d z| ||d \}}| |||\} } t||z| |j st||z| nt||d d d fz| | ||d \}} | |||\} } |j s(t||z| t| |z| d St||d d d fz| t| |d d d fz| d S) 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. r5r3r6rr)rurr>rN) r[r@rfloat64rArBnormalrIrrrrCr) rrrrTrZrYrXlossesr losses_sw gradient_swhessian hessian_sws rtest_sample_weight_multipliesrstI9   FNirzBBB i##$677 2299"*EE  )'    )))      ))%*FH "//%#0I{ F]*I666  H=0+>>>>=D#99;GGG--%.Hg #33%#4K  F=0+>>>-/<<<<<=D#99;GGG-4"88*EEEEEr!c6t|dddd\}}|jdkr|dddf}t||||||t||||||t||||||t||||||dSdS) z5Test that reshaped raw_prediction gives same results.rr3rr9rr&Nrq)r[ndimrrrrr)rrZrYraw_prediction_2ds rtest_graceful_squeezingrNsW:   FNa*111d73 IIV4EI F F IIVNI C C       f=N  O O   f^  L L     MM8IM J J MMM G G      ! !@Q ! R R  ! ! ! O O      r!c`|jstjgd}t|jt rRd}|jj}|jjs||z}|jj }|jj s||z }tj |||}|j |}ntj |jt }tj|j|jftjd t }tjd|jdd|jdz<|dkr-tjd|jd|jd }|||| }||| }t1|| d d dS)z~Test value of perfect predictions. Loss of y_pred = y_true plus constant_to_optimal_zero should sums up to zero. )rr|rrzr?rm绽|=rm)rur]rNr&rrrnrr+=gV瞯<)atolr)rCr@rr*rKr rLr;rRr<rSclipinverserHrErIrJfullexprGrrrurrr) rrrYrcr;r<rZ loss_value constant_terms rtest_loss_of_perfect_predictionrms  @"<"<"<== di . . @C&*C'5 Ci',D'6 "czW^S$??N"">224>**11%88>4>2r {    68VBZZ11t~112 Av|AFLOLLL %#J 11]2M JUGGGGGGr!c@ d}t|dd|\ dkr-tjd jd jd  \}}|j jksJ|j jksJjsc fd }t | d }t||d d fd}t | d } jrdSt|| d ddStj D]  fd}t | dd fd }t|dd f|d d  fd}t | dd fd } jrqt|dd f| d ddS)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. rr3r6rrr&rrnrc4|SNr)rrbrrrZs r loss_funcz6test_gradients_hessians_numerically..loss_funcs'99 + r!gư>)rch㈵>rrrc4|Srrrs r grad_funcz6test_gradients_hessians_numerically..grad_funcs'== +! r!cn}||ddf<|Sr)copyrrbrawkrrYrrZs rrz6test_gradients_hessians_numerically..loss_funcsI$))++AAAqD yy!#&"/!r!Ngh㈵>c}||ddf<|ddfSr)rrrs rrz6test_gradients_hessians_numerically..grad_funcs[$))++AAAqD }}!#&"/%!!Q$ r!) r[r@rrrurrCrirapprox_hessianrrE) rrrrTgrdr g_numericr h_numericrrYrZs `` @@@r#test_gradients_hessians_numericallyrsI9   FN Av|AFLOLLL  %# !  DAq 7n* * * * * 7n* * * * *  ;K       )NMMM 94e<<<<       )NMMM   A D Ayt% @ @ @ @ @ @t~&& K KA         -Yqqq!t8LRVWWWI AaaadGYT F F F F         -Yqqq!t8LRVWWWI" K!!!Q$EJJJJJ9 K Kr!zloss, x0, y_true) ) squared_errorgr9)rg@]@g?)rrr) binomial_lossrrz)rir)rr~) poisson_lossr_r)r rr)r g6g$@ctdtjgtjtj|gtj}dtjdtjffd }dtjdtjffd }dtjdtjffd }t ||||d d }|}tj |t||d d t |d ddS)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. Nrrrbreturnc`|zS)zCompute loss plus constant term. The constant term is such that the minimum function value is zero, which is required by the Newton method. rqrZ)rrrbrrZs rraztest_derivatives..funcs> yy!   ) ) ) 8 89 9r!c2|S)Nrqrrs rfprimez test_derivatives..fprime"s}}F1}===r!c>|dS)Nrqr&)rrs rfprime2z!test_derivatives..fprime2%s!$$F1$EEaHHr!r5gHj>)x0rrmaxitertolrrrrqgƠ>) r r@rrndarrayrravelrrKrr)rrrZrarroptimums` ` rtest_derivativesrs4 4=t , , ,D Xvhbj 1 1 1F 2$bj ) ) )B9 9rz9999999>"*>>>>>>>>I2:I"*IIIIIII   G\\^^FmmooGDI%%g..777DDMM151111DMMMHH!RVWWWWWWr!c djs0jtjddnAtjtjjzdddd<dkrtjd d   }fd }jst|d ddi} tj |}|j tksJ|jjksJt!||t#|jd k|t#ddkdSt)|tjjdddidt-tjdjfdd} tj|df}|jjksJt!|t3||jddt3|ddddS)zzTest that fit_intercept_only returns the argmin of the loss. Also test that the gradient is zero at the minimum. 2rrnrNr8rrzr'rcjstj|}n/tjtj|jf}|S)N)rur]rur)rCr@rascontiguousarray broadcast_torE)rbrYrrTrrZs rfunz%test_loss_intercept_only..funNss! WI1EEENN1)T^)DEEENt)'    r!gHz>rr5)roptionsrrr)rgvIh%<=SLSQPr&)rr%method constraintsrrrr)rCrKrr@rrrHrIrrErrrr`rutuplerrrrbrrzerosr rtiler)rrar$optgradrTrZs`` @@rtest_loss_intercept_onlyr0:sI  ""2;r1)#D#D#DEE9%%,,RZ884>Isss  C :::  v]KKA           9cti5EFFF}}<22'   w%''!!!!w&,&&&&! VCEt $ $ $$$  fQE*******  Hdn & &$(!T^1D)E)Eq!LL    }}71y!n55'   w&,&&&&!35t%8888a((!%888888r!zloss, func, random_distrc.tj|dS)N)q)r@ percentile)rbs rr5sr}Q"/E/E/Er!poisson exponentialbinomialctj|}|dkr|ddd}nt ||d}||}t ||t|j||ksJ|j |t||ksJt|tr(t|j |||j jrD||j j||}t ||j jrF||j j||}t |dSdS)zTest that fit_intercept_only returns the correct functional. We test the functional for specific, meaningful distributions, e.g. squared error estimates the expectation of a probability distribution. r8r&rpr5r>rN)r@rArBr8getattrrrrrKrr*r rrOrRfillr;rSr<)rra random_distrrXy_trainbaseline_predictions r test_specific_fit_intercept_onlyr?}s$ )   2 3 3Cj  ,,q#C,00+'#{++555111AA)*** &W )F)F"G"G G G G G 9  0 1 1VDDMM5J5J J J J J$ %%U ))*=>>@STTT )/ T)-..."55W5EE-... */ T).///"55W5EE-.....//r!c tjd}d}t|}|d|dzdtj}||}|j|fksJtj ||j }t|D]}||k ||<t|tj|tj tj|z t|d d d f|j|d d d ftj d tjd fD]X}|tj}||}|j |j ksJt#|Yd S) zATest that fit_intercept_only returns the mean functional for CCE.rrrr&r5r>rrNrmr!)r@rArBrrandintrIrrrur+rrmeanrlogrKrr)rXrErr=r>prs r(test_multinomial_loss_fit_intercept_onlyrEs )   " "CI  3 3 3Dkk!Y]k55<gNrpr&rq) r@rArBrrrArIrrrDrr)rrXrTbinommultinomr=rY raw_multinoms r"test_binomial_and_multinomial_lossrJs )   2 3 3CI   E"Q///Hkk!QYk//66rzBBGZZYZ//N8YN++L.LA~-LA '. AA W\ BBr!c d}t||dd|\}}t|drT||}|j||jfksJt j|dtdd ksJt|d rd d t j|ft j|d ft j|t j|ffD]\}}| ||d || \}}|j||jfksJt j|dtdd ksJt|| ||d d d Sd S)z22{y$.99999ve!$$$qe(<(<(<<<<<t%&&  2=00 1 ]> * *D 1 ]> * *BM.,I,I J    KD% ---"! .KD%;9dn"=====6%a(((F1%,@,@,@@@@@  !#1"&!%     !  r!rorder)CFcd}|dkrtj|}||}||||\}}|jr|j|fksJ|jdksJnL|jr)|j||jfksJ|j||jfksJn|j|fksJ|j|fksJ|j|ksJ|j|ksJ|dkr|jj sJ|jj sJdS|jj sJ|jj sJdS)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. r8rr )rTrrM)r&rNN) r@rinit_gradient_and_hessianconstant_hessianrurCrErflags c_contiguous f_contiguous)rrrrMrTrrs rtest_init_gradient_and_hessiansrVsI ** 4m , , ,D667Hg  -~)----}$$$$$ -~)T^!<<<<<}DN ;;;;;;} ,,,,} ,,,, >U " " " " =E ! ! ! ! ||~****}))))))~****}))))))r!zparams, err_msgz+Valid options for 'dtype' are .* Got dtype=z instead.c|}tjttf|5|jdddi|\}}ddddS#1swxYwYdS)zDTest that init_gradient_and_hessian raises errors for invalid input.matchrTr8Nr)pytestraises ValueError TypeErrorrQ)rparamserr_msgrrs r%test_init_gradient_and_hessian_raisesr`$s 466D  I.g > > >RR:D:QQQQ&QQ'RRRRRRRRRRRRRRRRRRsA  AAzloss, params, err_type, err_msgr"z4quantile must be an instance of float, not NoneType.zquantile == 0, must be > 0.zquantile == 1.1, must be < 1.cttj||5|di|ddddS#1swxYwYdS)z/Test that loss raises errors for invalid input.rXNr)rZr[)rr^err_typer_s r#test_loss_init_parameter_validationrc5s( xw / / / vs -11cd}t||ddd\}}tj|}tj|}|||t |||ksJdS)z Test that losses can be pickled.rr3r6r9rrqN)r[pickledumpsloadsr)rrTrZrY pickled_lossunpickled_losss rtest_loss_picklerjMsI9   FN<%%L\,//N 4vn = = =f^DDDBB      r!rD)r#rr&r}r'r?cBt|}t|}d}t||d\}}|j|}|||||z}|||||z}t|||||\} } |||\} } t| || zt| || z|dz| zzdS)zCTest for identical losses when only the link function is different.r$rmr9)rrTrWrqr'N) rrr[rKrrrrr) rDhalf_tweedie_loghalf_tweedie_identityrTrZrYrxloss_log loss_identity gradient_log hessian_loggradient_identityhessian_identitys r%test_tweedie_log_identity_consistencyrt_s'Q///3!<<<I9 FN " * *> : :F $$n%11&99:H*..f/66v>>?MHm,,,!1 A An!B!!L++@*P*Pf+Q++''L&+<"<===V//&!)>N2NNr!)r3r6r9)Nrenumpyr@ numpy.testingrrrZrscipy.optimizerrrr scipy.specialr sklearn._loss.linkr r sklearn._loss.lossr rrrrrrrrrr sklearn.utilsrsklearn.utils._testingrrlistvalues ALL_LOSSESLOSS_INSTANCESr2r[rimark parametrizeryinfY_COMMON_PARAMS Y_TRUE_PARAMS Y_PRED_PARAMSrrrCrfloat32rrrrrrrrrr0rBmedianr?rErJrLrVint64r`r]r\rcrjrtrr!rrs  ========  $#####@@@@@@@@                          ,+++++KKKKKKKKT.'."" # # 00Z000KO$O!O!O!O#!$$$!$$$!$$$#&&&     CE""""""""J U U U5GHH<<IH_1SzRVGRV+<=_3#srvgr4-HI_1SzRVGRq"&+IJ_1SzRVGRq"&+IJ2&&&c bfWbf4EF1%%%'>'>'>"&"&@QR3'''#srvgr45PQ1%%%SzRVGRq"&3QR1%%%SzRVGRq"&3QRBF7B26*BCB"&"c26 :;%0_R _24_1ay"-_3!b)2&&&<1%%%ay"53'''!b1!Q$MMM2.  _QC _2MMM2_1r===1_3aS)2&&&MMM:1%%%}}}b93'''aS1aV$S#JA/  '=)H777'=)H777-    S#q) #sA& c " " "Ca0 d # # #S#~> d # # #S#x8   CAFBF1II ,=> #vrvayy&"&))e*;< q ! ! !3q 3DE q ) ) )3QVRVAYY5FG q ) ) )3VRVAYY5FG q ) ) )35RS    T626!99fbfQii$:J.JK  ! , , ,  OO Iooo & & ,   ! , , ,  OO Iooo & & ,   ! , , ,  OO Iooo & & ,  3@ E##H11I##H1,,*UDM::bj"*%=>>rz2:&>??4)44$++$++q!f--Q Q .-,+,+54@??>;:-,Q h5GHH4/::M$M$;:IHM$`5GHH4/::L>L>;:IHL>^5GHH68*<==BFBF>=IHBFJ5GHH  IH <5GHH4/::-H-H;:IH-H`5GHH4/::VKVK;:IHVKr   "+X+X#$+X\5GHH4/::>9>9;:IH>9B    RWh/ ")X. d # # #%E%ExP   BGY/ "'=1   BG]3    RWj1  //  />///."5GHH((IH(V,,4/::2:rz":;;*--!*!*.-<;;:-,!*H,,bh  M"( M M M RR-,R%     B    O  )  z3'5TU$%$ 5GHHIH"55566))76)))r!