U mdH@s:dZddlZddlmZddlmZmZGdddZdS)zL Created on Sun May 10 08:23:48 2015 Author: Josef Perktold License: BSD-3 N) NonePenalty)approx_fprime_cs approx_fprimecseZdZdZfddZdddZdfdd Zdfd d Zdd d Zdfdd Z dfdd Z dddZ dfdd Z d fdd Z ZS)!PenalizedMixinaMixin class for Maximum Penalized Likelihood Parameters ---------- args and kwds for the model super class penal : None or instance of Penalized function class If penal is None, then NonePenalty is used. pen_weight : float or None factor for weighting the penalization term. If None, then pen_weight is set to nobs. TODO: missing **kwds or explicit keywords TODO: do we adjust the inherited docstrings? We would need templating to add the penalization parameters cs|dd|_|dd|_tt|j|||jdkrDt|j|_|jdkr\t|_d|_|j ddgt |dg|_ |j ddgdS)Npenal pen_weightr_null_drop_keys) poprrsuperr__init__lenZendogrZ _init_keysextendgetattrr )selfargskwds __class__T/home/sam/Atlas/atlas_env/lib/python3.8/site-packages/statsmodels/base/_penalized.pyr !s   zPenalizedMixin.__init__NcKs0|dkr,t|dr(||}||}nd}|S)NZ scaletyper)hasattrZpredictZestimate_scale)rparamsscalermurrr _handle_scale8s    zPenalizedMixin._handle_scalec sX|dkr|j}tt|j|f|}|dkrT|j|f|}|d|||j|8}|S)z3 Log-likelihood of model at params Nrr)rr rloglikerrfunc)rrrrllfrrrrrDszPenalizedMixin.loglikec sj|dkr|j}tt|j|f|}t|jd}|dkrf|j|f|}|d||||j|8}|S)z@ Log-likelihood of model observations at params Nrr) rr r loglikeobsfloatshaperrr)rrrrrZnobs_llfrrrrrRs zPenalizedMixin.loglikeobsfdc sRdkrjfdd}|dkr0t||S|dkrFt||ddStddS) z4score based on finite difference derivative Ncsj|fdiSNrrprrrrrhz.PenalizedMixin.score_numdiff..csr"T)Zcenteredz-method not recognized, should be "fd" or "cs")rrr ValueError)rrrmethodrrrr'r score_numdiffbs zPenalizedMixin.score_numdiffc sX|dkr|j}tt|j|f|}|dkrT|j|f|}|d|||j|8}|S)z- Gradient of model at params Nrr)rr rscorerrderiv)rrrrscrrrrr.qszPenalizedMixin.scorec sj|dkr|j}tt|j|f|}t|jd}|dkrf|j|f|}|d||||j|8}|S)z: Gradient of model observations at params Nrr) rr r score_obsr r!rrr/)rrrrr0Znobs_scrrrrr1s zPenalizedMixin.score_obsc s4dkrjfdd}ddlm}|||S)z6hessian based on finite difference derivative Ncsj|fdiSr#r$r%r'rrr(r)z0PenalizedMixin.hessian_numdiff..r) approx_hess)rstatsmodels.tools.numdiffr2)rrrrrr2rr'rhessian_numdiffs  zPenalizedMixin.hessian_numdiffc s|dkr|j}tt|j|f|}|dkr~|j|f|}|j|}|jdkrj|d|t ||8}n|d|||8}|S)z, Hessian of model at params Nrr) rr rhessianrrZderiv2ndimnpZdiag)rrrrZhessrhrrrr5s  zPenalizedMixin.hessianc  sddlm}ddlm}t|||fr4|ddi|dkr@d}|dkrLd}tt|jfd|i|}|dkrr|S|d kr~d }t t |j |kd}t t |j |kd}| r|j|f|} | S|SdS) aminimize negative penalized log-likelihood Parameters ---------- method : None or str Method specifies the scipy optimizer as in nonlinear MLE models. trim : {bool, float} Default is False or None, which uses no trimming. If trim is True or a float, then small parameters are set to zero. If True, then a default threshold is used. If trim is a float, then it will be used as threshold. The default threshold is currently 1e-4, but it will change in future and become penalty function dependent. kwds : extra keyword arguments This keyword arguments are treated in the same way as in the fit method of the underlying model class. Specifically, additional optimizer keywords and cov_type related keywords can be added. r)GLMGam)GLMZmax_start_irlsNZbfgsFr,Tg-C6?)Z*statsmodels.gam.generalized_additive_modelr9Z+statsmodels.genmod.generalized_linear_modelr: isinstanceupdater rfitr7ZnonzeroabsranyZ _fit_zeros) rr,Ztrimrr9r:resZ drop_indexZ keep_indexZres_auxrrrr=s&  zPenalizedMixin.fit)N)N)N)Nr")N)N)N)N)NN)__name__ __module__ __qualname____doc__r rrrr-r.r1r4r5r= __classcell__rrrrrs    r) rDnumpyr7Z _penaltiesrr3rrrrrrrs