ó áp7]c@s€dZddlZddlmZddljjZeddeedd„Z d„Z d„Z d„Z d „Zd „ZdS( sR Holds files for l1 regularization of LikelihoodModel, using scipy.optimize.slsqp iÿÿÿÿN(t fmin_slsqpièc 'sytj|ƒjdƒ}t|ƒ‰tj|tj|ƒƒ} tj|dƒjdƒ‰ˆtjˆƒ‰ˆjƒdks†t‚t ||ƒ} |j ddƒ} ‡‡‡‡fd†}‡fd†}‡‡‡fd†}‡fd †}t || d |d |d| d |d | d| d|ƒ}tj |dˆ ƒ}|d}|d}t j|ˆˆ||ƒ}|d}|d}|d}t j|ˆˆˆ||||ƒ\}}| ra|\}}}}}|tj |ƒƒ} |dk}!t|ƒd|}"|}#tdƒ}$tdƒ}%i| d6|!d6|#d6|$d6|%d6|d6|"d6}&n| rq||&fS|SdS(s Solve the l1 regularized problem using scipy.optimize.fmin_slsqp(). Specifically: We convert the convex but non-smooth problem .. math:: \min_\beta f(\beta) + \sum_k\alpha_k |\beta_k| via the transformation to the smooth, convex, constrained problem in twice as many variables (adding the "added variables" :math:`u_k`) .. math:: \min_{\beta,u} f(\beta) + \sum_k\alpha_k u_k, subject to .. math:: -u_k \leq \beta_k \leq u_k. Parameters ---------- All the usual parameters from LikelhoodModel.fit alpha : non-negative scalar or numpy array (same size as parameters) The weight multiplying the l1 penalty term trim_mode : 'auto, 'size', or 'off' If not 'off', trim (set to zero) parameters that would have been zero if the solver reached the theoretical minimum. If 'auto', trim params using the Theory above. If 'size', trim params if they have very small absolute value size_trim_tol : float or 'auto' (default = 'auto') For use when trim_mode === 'size' auto_trim_tol : float For sue when trim_mode == 'auto'. Use qc_tol : float Print warning and don't allow auto trim when (ii) in "Theory" (above) is violated by this much. qc_verbose : Boolean If true, print out a full QC report upon failure acc : float (default 1e-6) Requested accuracy as used by slsqp tFtalpha_rescaleditaccg»½×Ùß|Û=cstˆ|ˆˆˆŒS(N(t_objective_func(tx_full(talphatargstftk_params(s8lib/python2.7/site-packages/statsmodels/base/l1_slsqp.pytFtcs t|ˆƒS(N(t _f_ieqcons(R(R (s8lib/python2.7/site-packages/statsmodels/base/l1_slsqp.pyR GR cstˆ|ˆˆƒS(N(t_fprime(R(RR tscore(s8lib/python2.7/site-packages/statsmodels/base/l1_slsqp.pyR HR cs t|ˆƒS(N(t_fprime_ieqcons(R(R (s8lib/python2.7/site-packages/statsmodels/base/l1_slsqp.pyR IR t f_ieqconstfprimetitertdispt full_outputtfprime_ieqconstqc_tolt qc_verboset trim_modet size_trim_tolt auto_trim_tolt tnantfoptt convergedt iterationstgoptthoptttrimmedtwarnflagN(tnptarraytraveltlentappendtfabstonestmintAssertionErrort_get_disp_slsqpt setdefaultRtasarraytl1_solvers_commont qc_resultstdo_trim_paramststrtfloat('RRt start_paramsRtkwargsRtmaxitertcallbacktretallRthesstx0t disp_slsqpRtfunctf_ieqcons_wrapt fprime_wraptfprime_ieqcons_wraptresultstparamsRRtpassedRRRR"RtfxtitstimodetsmodeRRR#RR R!tretvals((RRRR Rs8lib/python2.7/site-packages/statsmodels/base/l1_slsqp.pyt fit_l1_slsqp sT)            cCs7|s |r-|rd}n|r3d}q3nd}|S(Niii((RR9R<((s8lib/python2.7/site-packages/statsmodels/base/l1_slsqp.pyR-ws   cGs/|| }||}|||Œ||jƒS(s, The regularized objective function (tsum(RRR RRtx_paramstx_added((s8lib/python2.7/site-packages/statsmodels/base/l1_slsqp.pyR‚s  cCs || }tj||ƒ|ƒS(s$ The regularized derivative (R$R((RRR RRK((s8lib/python2.7/site-packages/statsmodels/base/l1_slsqp.pyR Œs cCs,|| }||}tj||||ƒS(s% The inequality constraints. (R$R((RR RKRL((s8lib/python2.7/site-packages/statsmodels/base/l1_slsqp.pyR •s  cCsetj|ƒ}tj||fddƒ}tj| |fddƒ}tj||fddƒ}|S(s2 Derivative of the inequality constraints taxisii(R$teyet concatenate(RR tItAtBtC((s8lib/python2.7/site-packages/statsmodels/base/l1_slsqp.pyRŸs (t__doc__tnumpyR$tscipy.optimizeRt"statsmodels.base.l1_solvers_commontbaseR0tFalsetNoneRIR-RR R R(((s8lib/python2.7/site-packages/statsmodels/base/l1_slsqp.pyts k