ó V™xTc@ sŠddlmZdd dd eed„Zdefd„ƒYZdefd„ƒYZed „Ze d kr†ed eƒZ d GHnd S(iÿÿÿÿ(tdivisioniicC sÄddlm}|j}|j|jƒ} d| _|rd|dk rd|jd|jd|ƒ} n0|r|j|jƒ} nt‚|j |ƒ} yþd} t } |} xå| r¿| d7} n|d| d| ƒ\}}}}|r"| r d Gt |ƒG|Gt |ƒGHq"d Gt |ƒG|GHn|rwdd l m }|dkr\|j|jƒƒ}n| j||||ƒ}n| j|||ƒ}|r°| jƒ} | jƒ} q¬n| jƒ sÃt‚| jƒ sÖt‚|r| jƒ|kr| ||ƒrPqn| |ƒrPn|dk r6| jƒ|kr6Pn|rW| j||||ƒ}n| j|||ƒ}|svPn| jƒ} | jƒ} q¬Wn"tk r¶}t|ƒ| _nX| | _| S( Niÿÿÿÿ(tlbfgstntepsiitrequests_f_and_gt requests_diagsx,f,d:sx,f:(tflex(tscitbxRtextt minimizerRtNoneterrorttraditional_convergence_testt RuntimeErrortdrop_convergence_testtTruettupletscitbx.array_familyRtdoubletsizetrunRRtAssertionErrortitertstrtn_calls(ttarget_evaluatortmin_iterationstmax_iterationsR t traditional_convergence_test_epstuse_curvaturestverboset scitbx_lbfgsRR t is_convergedticallRRtxtftgtdRt have_requestte((sI/public2/ftp/pub/ccp4/6.5/ccp4-6.5/lib/py2/scitbx/lbfgs/tst_curvatures.pyt lbfgs_run sn        !   tlbfgs_with_curvatures_mix_incB s,eZdZddded„Zd„ZRS(sMix in this class with your application-specific minimizer class. The app-specific class is required to implement: 1. In the constructor a) define the free parameters as self.x = flex.double() b) last line: call the lbfgs_with_curvatures_mix_in constructor 2. compute_functional_and_gradients() returns (double functional, flex.double gradients) 3. curvatures() returns flex.double curvatures == diagonal elements of Hessian matrix Exposed variable traditional_convergence_test_eps gives some flexibility to the application program to specify the epsilon value for hitting convergence. Choosing the right value avoids unnecessary minimization iterations that optimize the target functional to the last decimal place. iièc C s@t|jƒ|_td|d|d|d|d|ƒ|_dS(NRRRRR(tlenR"RR(R (tselfRRRR((sI/public2/ftp/pub/ccp4/6.5/ccp4-6.5/lib/py2/scitbx/lbfgs/tst_curvatures.pyt__init__YscC sÏ| r| rt}t}n|rG|jƒ\|_|_d|_n|r³|jƒ|_|jdk}|jj|dƒ|jj|ƒj dƒs t ‚d|j|_n|j |j|j|jfS(Niièi( Rt compute_functional_and_gradientsR#R$R R%t curvaturest set_selectedtselecttall_gtRR"(R+RRtsel((sI/public2/ftp/pub/ccp4/6.5/ccp4-6.5/lib/py2/scitbx/lbfgs/tst_curvatures.pyt__call__ds  !N(t__name__t __module__t__doc__R RR,R3(((sI/public2/ftp/pub/ccp4/6.5/ccp4-6.5/lib/py2/scitbx/lbfgs/tst_curvatures.pyR)Js   tfit_xy_translationcB s2eZdZd„Zd„Zd„Zd„ZRS(syElementary test case. Six groups of xy observations generated by random Gaussian with mean = 0 and sigma = 1. Fit the population means with six xy model means. However, since the sample sizes are extremely different, convergence is slow if using a global least-squares target function. Convergence speeds up considerably if curvatures are added. cC sÌ||_|jƒddlm}d|_|jdgdƒ|_tj|ddddd |ƒ|jrÈg|jddd …D]}d |^q„GHg|jd dd …D]}d |^q°GHndS( Niÿÿÿÿ(Rigi RRièRis%8.5fi( Rtcreate_test_dataRRtcount_iterationsRR"R)R,(R+RRRta((sI/public2/ftp/pub/ccp4/6.5/ccp4-6.5/lib/py2/scitbx/lbfgs/tst_curvatures.pyR,s     ,cC sŠddlm}|jdgdƒ}x]tdƒD]O}|j|k}d|jtƒ|d|s :.W