U 3d8m @sdZddlZddlZddlZddlZddlmZddlZddl m Z ddl m Z m ZmZddl mZmZddlmZddlmZdd lmZmZmZmZd d Zed d ddddgjZedddddgjZee Z!e d ddZ"e d de"e d dded de d dded de d ddeddedde d ddeddgZ#dd e#DZ$ej%&d!e#d"d#Z'd$d%Z(ej%&d!e$d&d'Z)ej%&d!e#d(d)Z*ej%&d!e#d*d+Z+ej%&d!e$d,d-Z,ej%&d!e$d.d/Z-ej%&d!e#d0d1Z.ej%&d!e#d2d3Z/ej%&d!e#d4d5Z0d6d7Z1ej%&d!e#ej%&d8e!ej2ej3dej4d9gd:d;Z5dd?Z7ej%&d!e#d@dAZ8dBdCZ9dDdEZ:ej%&d!e$dFdGZ;dHdIZdNdOZ?dPdQZ@ej%&d!e#dRdSZAdTdUZBej%&dVdWeCdXieDdYfeejE ejEfdZd[d\eDd]fgd^d_ZFd`daZGdbdcZHej%&dddedfgej%&dgddhdigdjdkZIej%&dddedfgej%&dgddhdigdldmZJGdndodoeZKdpdqZLdS)rz(Testing for Gaussian process regression N) approx_fprime)GaussianProcessRegressor)RBFConstantKernel WhiteKernel) DotProductExpSineSquared) MiniSeqKernel)ConvergenceWarning)assert_array_lessassert_almost_equalassert_array_almost_equalassert_allclosecCs|t|S)N)npsin)xrK/tmp/pip-unpacked-wheel-zrfo1fqw/sklearn/gaussian_process/tests/test_gpr.pyf!sr?g@g@g@g@g @@g@g@g@g@fixed length_scalelength_scale_boundsr)MbP?@@{Gz?Y@h㈵>)r!r 皙?cCsg|]}|tkr|qSr) fixed_kernel).0kernelrrr 4sr&r%cCsVtjdkrtdt|dtt}|jtdd\}}t |tt t |ddS)N#This test may fail on 32 bit Pythonr%T return_cov) sysmaxsizepytestxfailrfitXypredictr rdiag)r%gpry_predy_covrrrtest_gpr_interpolation7s    r9cCstdd}dddg}tdddg}t|d ||}|j|d d \}}t||d d ddtt |t||tt |d dS)Nr)Zbaseline_similarity_boundsABCr)Tr*Z eval_gradientr,) r rarrayrr1r4r ravelZeyelenr5)r%r2r3r6r7r8rrr!test_gpr_interpolation_structuredDs   rDcCsHtjdkrtdt|dtt}||j j ||j ksDt dS)Nr'r(r)) r-r.r/r0rr1r2r3log_marginal_likelihoodkernel_thetaAssertionErrorr%r6rrrtest_lml_improvingSs   rJcCs6t|dtt}||jjt|ks2t dS)Nr)) rr1r2r3rErFrGr/approxrHrIrrrtest_lml_precomputed_srLcCsJt|dtt}tj|jjjtj d}|j |ddt |jj|ddS)Nr)dtypeF)Z clone_kernel) rr1r2r3ronesrFrGshapefloat64rEr )r%r6Z input_thetarrrtest_lml_without_cloning_kernelhsrScCszt|dtt}||jjd\}}tt |dk|jj|jj dddfkB|jj|jj dddfkBsvt dS)Nr)T-C6?rr=) rr1r2r3rErFrGrallabsboundsrH)r%r6lml lml_gradientrrrtest_converged_to_local_maximumrs rZcCst|dtt}|jj}t|jjj j }d}||t |dddfdf<t |dddf|jj|t |jj|dddf|dS)Nr)绽|=r=r) rr1r2r3rFrWrfinforGrNmaxisfiniter )r%r6rWZmax_Ztinyrrrtest_solution_inside_boundss r_csJt|dtt|jd\}}t|jfddd}t||ddS)Nr)Tcs |dS)NF)rE)rGr6rrz#test_lml_gradient..r[r?)rr1r2r3rErGrr )r%rXrYZlml_gradient_approxrr`rtest_lml_gradients rccCsnt|d}|jtdd\}}t|ddt|jjdkrXtt|t |jddntt|dddS)Nr)Tr*rr=) rr4r2r rCr%rGrr5exp)r%r6y_meanr8rrr test_priors    rgcCs~t|dtt}|jtdd\}}|td}t|t |ddtt |t | t |dt | ddS)Nr)Tr*ir=) rr1r2r3r4X2sample_yr rmeanr5r]var)r%r6rfr8Zsamplesrrrtest_sample_statisticss rlcCs6td}t|ddtt}t|jjdks2t dS)Nrr% optimizer) rrr1r2r3rrerFrGrHrIrrrtest_no_optimizersrotargetrMcCsdtjdkrtdt|dtt}|jt dd\}}|jt dd\}}t t t ||dS)Nr'r(r)Tr* return_std)r-r.r/r0rr1r2r3r4rhr rsqrtr5)r%rpr6rfr8y_stdrrrtest_predict_cov_vs_stds   rucCstjd}|ddd}|dddfd|dddf}tddg}t|d||}t|jj dt|jj ddkst dS) Nrr=)2r>r"rr)rd) rrandom RandomStateuniformrrr1rerFrGrH)rngr2r3r%r6rrrtest_anisotropic_kernels  $ r|c Csd\}}tjd}|||dd}t|jddtd|jdd|jd|d}td d td g|d g|d t d dd}tj }t dD]F}t ||dd ||}||jj} | |ttjjkst| }qdS)N)r>rr>r=Zaxisr?r")Zscalesizerr)rTr rr!)r!$@) noise_levelnoise_level_boundsrd)r%n_restarts_optimizerZ random_state)rrxryrandnrsumnormalr<rrinfrangerr1rErFrGr\Zfloat32ZepsrH) n_samples n_featuresr{r2r3r%Zlast_lmlrZgprXrrrtest_random_startss8     rc Cstt}tt}t||}t|d}|t|t|dd}|tt|jtdd\}}|||}||}|jtdd\}} t ||t || |jtdd\} } | |d} |jtdd\} } t | | dS)a Test normalization of the target values in GP Fitting non-normalizing GP on normalized y and fitting normalizing GP on unnormalized y should yield identical results. Note that, here, 'normalized y' refers to y that has been made zero mean and unit variance. r)Tr% normalize_yrqr*r>N) rrjr3stdrr1r2r4rhr ) r%rfrtZy_normr6Zgpr_normr7 y_pred_stdZ y_pred_normZy_pred_std_norm_r8Z y_cov_normrrrtest_y_normalizations"           rcCsdt}ddi}tf|}t|dd}|t||jtdd\}}tddd d d g}td d dddg}t ||dddt ||ddddS)a Here we test that, when noramlize_y=True, our GP can produce a sensible fit to training data whose variance is significantly larger than unity. This test was made in response to issue #15612. GP predictions are verified against predictions that were made using GPy which, here, is treated as the 'gold standard'. Note that we only investigate the RBF kernel here, as that is what was used in the GPy implementation. The following code can be used to recreate the GPy data: -------------------------------------------------------------------------- import GPy kernel_gpy = GPy.kern.RBF(input_dim=1, lengthscale=1.) gpy = GPy.models.GPRegression(X, np.vstack(y_large), kernel_gpy) gpy.optimize() y_pred_gpy, y_var_gpy = gpy.predict(X2) y_pred_std_gpy = np.sqrt(y_var_gpy) -------------------------------------------------------------------------- rrTrrqgy(V.@gV],;g6}~CgHW -@gƠKQ@g ؼG'@gsr@g{u>5?g ?g.[ ?gQ?r)rtolatolg333333?N) r3rrr1r2r4rhrrAr)Zy_largeZ RBF_paramsr%r6r7rZ y_pred_gpyZy_pred_std_gpyrrrtest_large_variance_y!s     rcCstttdfj}tdd}t|ddd}|ttt|ddd}|t||jt dd\}}|jt dd\}}|jt dd\}} |jt dd\}} t ||ddd ft ||ddd fdt |j d D](} t ||d | ft | | d | fq|j t d d } |j t d d } | j dks.t| j dks>tt | | ddd ddftD]V}t|dd}|ttt|dd}|ttttfjt |jj|jjdq^dS)Nr>rrF)r%rnrTrqr*rr=.rr)rdr)rdr>rr)rvstackr3Trrr1r2r4rhr rrQrirHkernelsrFrG)Zy_2dr%r6Zgpr_2dZ y_pred_1dZy_std_1dZ y_pred_2dZy_std_2drZy_cov_1dZy_cov_2drpZ y_sample_1dZ y_sample_2drrrtest_y_multioutputXs4      rcCsDdd}t||d}|tt||jj||jjks@tdS)Nc Sstjd}|||dd}}tdD]\}t|td|dddftd|dddf}||dd}||kr&||}}q&||fS)NrFr@rwr=)rrxryrZ atleast_1drzmaximumZminimum) Zobj_funcZ initial_thetarWr{Z theta_optZfunc_minrrGrrrrrns   2  z(test_custom_optimizer..optimizerrm) rr1r2r3rErFrGr%rH)r%rnr6rrrtest_custom_optimizers   rc Csjtddd}td}t}t|dd}d|}tjtjj t |d| ||W5QRXdS)N rvr,r%alphazThe kernel, %s, is not returning a positive definite matrix. Try gradually increasing the 'alpha' parameter of your GaussianProcessRegressor estimator.match) rZarangeZreshaperPrrr/raisesZlinalgZ LinAlgErrorreescaper1)r2r3r%r6messagerrrtest_gpr_correct_error_messages  rc Cst|dd}t|dd}tttdf}tttddf}|||tttddf}tttddf}|||tddddddf}|j|dd \}}|j|dd \}} t ||t || dS) Nrrrr=gV瞯rzThe optimal value found for dimension 0 of parameter k1__noise_level is close to the specified upper bound 0.001. Increasing the bound and calling fit again may find a better value.r=zThe optimal value found for dimension 0 of parameter k2__length_scale is close to the specified lower bound 1000.0. Decreasing the bound and calling fit again may find a better value.rrrr rzThe optimal value found for dimension 0 of parameter length_scale is close to the specified lower bound 10.0. Decreasing the bound and calling fit again may find a better value.zThe optimal value found for dimension 1 of parameter length_scale is close to the specified lower bound 10.0. Decreasing the bound and calling fit again may find a better value.)rrr/Zwarnsr r1r2r3rwarningscatch_warnings simplefilterrCrH issubclasscategoryrargsrZtile) r%r6Zwarning_messageZ kernel_sumZgpr_sumrZX_tileZ kernel_dimsZgpr_dimsrrrtest_warning_boundssR          rcCs:dtdd}tdddd}||}t|dttdS)Ng@gI@rrr)rZ periodicityZperiodicity_boundsr))rrrr1r2r3)Zk1Zk2r%rrr%test_bound_check_fixed_hyperparameter#src Cs4tjtjdtjd}t|dd}|t||jt dksBt |j tdd\}}t ||t t |ddd tjdd }}tjd}tj|j|d fd tj|d fd d gd d}|t||j tdd\} } t | ddd fd t t | dddd | j||fkst | j|||fks0t dS)aCheck that the std. dev. is affected to 1 when normalizing a constant feature. Non-regression test for: https://github.com/scikit-learn/scikit-learn/issues/18318 NaN where affected to the target when scaling due to null std. dev. with constant target. rrMTrrr*r,g& .>)rr>r=)r)rQZ fill_valuer~N).r=)rrPr2rQrRrr1Z _y_train_stdr/rKrHr4rr5rxryZ concatenaterfull) r%Z y_constantr6r7r8r n_targetsr{r3ZY_predZY_covrrrtest_constant_target/s*      rc Cstddtddgdtdd}t|ddd }td d gd d gd d gd d gddgd d gg}tdgdgdgdgdgdgg}|||tddgddgddgddgg}|j|dd\}}|j|dd\}}t|t t |dddS)a Check the consistency between the returned std. dev. and the covariance. Non-regression test for: https://github.com/scikit-learn/scikit-learn/issues/19936 Inconsistencies were observed when the kernel cannot be inverted (or numerically stable). gT l+A)g-q=gmBg닄z@gG\@r!)rrN)r%rrnr,g6~?g6~g6~g6~?g6$}gNխgHx@gR| =gd-ig^L @gzgz?Trqr*)r) r<rrrrrAr1r4rrsZdiagonal) r%r6X_trainy_trainrZpred1rZpred2Zcovrrr2test_gpr_consistency_std_cov_non_invertible_kernelXsH   rzparams, TypeError, err_msgrrzCalpha must be a scalar or an array with same number of entries as yrr>)r%rz#requires that all bounds are finitec Cs4tf|}tj||d|ttW5QRXdS)z0Check that expected error are raised during fit.rN)rr/rr1r2r3)params TypeErrorerr_msgr6rrrtest_gpr_fit_errors rc CsBttdtt}d}tjt|d|jddW5QRXdS)z7Check that we raise the proper error in the LML method.r)z.Gradient can only be evaluated for theta!=NonerTr@N) rrr1r2r3r/r ValueErrorrEr6rrrrtest_gpr_lml_errorsrc CsFttdtt}d}tjt|d|jtdddW5QRXdS)z4Check that we raise the proper error during predict.r)z9At most one of return_std or return_cov can be requested.rT)r+rrN) rrr1r2r3r/r RuntimeErrorr4rrrrtest_gpr_predict_errorsrrTFrr=rcCstjd}d\}}}|f}|dk r.||f}|f}|dk rN|dkrN||f}|||}|||} |j|} t|d} | || | j| dd\} } | j| dd\}}| j|kst| j|kst|j|f|kstdS) aCheck the shapes of y_mean, y_std, and y_cov in single-output (n_targets=None) and multi-output settings, including the edge case when n_targets=1, where the sklearn convention is to squeeze the predictions. Non-regression test for: https://github.com/scikit-learn/scikit-learn/issues/17394 https://github.com/scikit-learn/scikit-learn/issues/18065 https://github.com/scikit-learn/scikit-learn/issues/22174 )r rONr=rTrqr*) rrxryrrr1r4rQrH)rrr{rn_samples_trainZn_samples_test y_train_shape y_test_shaperrrmodelr7rtrr8rrrtest_predict_shapess$         rcCstjd}d\}}d}d}|f}|dk r4||f}|dk rP|dkrP|||f}n||f}|||} |||} |j|} t|d} | | | | j| |d} | j|kstdS) a&Check the shapes of y_samples in single-output (n_targets=0) and multi-output settings, including the edge case when n_targets=1, where the sklearn convention is to squeeze the predictions. Non-regression test for: https://github.com/scikit-learn/scikit-learn/issues/22175 r)rrrOrdNr=rr) rrxryrrr1rirQrH)rrr{rrZn_samples_X_testZn_samples_y_testrrrrrrZ y_samplesrrrtest_sample_y_shapess"       rc@seZdZdZddZdS) CustomKernelz A custom kernel that has a diag method that returns the first column of the input matrix X. This is a helper for the test to check that the input matrix X is not mutated. cCs|dddfS)Nrr)selfr2rrrr5szCustomKernel.diagN)__name__ __module__ __qualname____doc__r5rrrrrsrcCs>ttdtt}tt}|jtdd\}}t t|dS)z Check that the input X is not modified by the predict method of the GaussianProcessRegressor when setting return_std=True. Non-regression test for: https://github.com/scikit-learn/scikit-learn/issues/24340 r)TrqN) rrr1r2r3rcopyrhr4r)r6ZX2_copyrrrr#test_gpr_predict_input_not_modifieds r)Mrrr-rZnumpyrZscipy.optimizerr/Zsklearn.gaussian_processrZ sklearn.gaussian_process.kernelsrrr<rrrZ4sklearn.gaussian_process.tests._mini_sequence_kernelr Zsklearn.exceptionsr Zsklearn.utils._testingr r r rrZ atleast_2drr2rhrBr3r#rZnon_fixed_kernelsmarkZ parametrizer9rDrJrLrSrZr_rcrgrlrorPrQrRrur|rrrrrrrrrrrrzerosrrrrrrrrrrrrrs                      "   '7.    M  (-     & .