M/Ph;dZddlZddlmZddlmZddlmZddl m Z m Z ddl Z dZ ddZGd d ZGd d ZGd dZGddeeZdS)a_ Generalized additive models Requirements for smoothers -------------------------- smooth(y, weights=xxx) : ? no return ? alias for fit predict(x=None) : smoothed values, fittedvalues or for new exog df_fit() : degress of freedom of fit ? Notes ----- - using PolySmoother works for AdditiveModel, and GAM with Poisson and Binomial - testfailure with Gamma, no other families tested - there is still an indeterminacy in the split up of the constant across components (smoothers) and alpha, sum, i.e. constant, looks good. - role of offset, that I have not tried to figure out yet Refactoring ----------- currently result is attached to model instead of other way around split up Result in class for AdditiveModel and for GAM, subclass GLMResults, needs verification that result statistics are appropriate how much inheritance, double inheritance? renamings and cleanup interface to other smoothers, scipy splines basic unittests as support for refactoring exist, but we should have a test case for gamma and the others. Advantage of PolySmoother is that we can benchmark against the parametric GLM results. N)families) PolySmoother)GLM)IterationLimitWarningiteration_limit_docFc(tj|}|jd}|dkr|}ntjdtjdz }tjdtjdz }tjdtjdz }tjdtjdz }|dkrd|||z |dz zdz zz}nA|d krd|||z |dz zd z zz}n&|d krd|||z |d z zd z zz}n d|d z dzz}|tjd|dz |tj} |d} n|} t| |} | S)z ri2dgb@i g@i g@g@g?N)x) npsortshapeloglinspaceastypeint32rcopy) rs_arg_xnnknotsa1a2a3a4knotsorderss W/var/www/html/test/jupyter/venv/lib/python3.11/site-packages/statsmodels/sandbox/gam.pydefault_smootherr%<s B  A 3ww VBZZ"&)) # VC[[26!99 $ VC[[26!99 $ VC[[26!99 $ s77rBw1r624778FF WWrBw1s73D889FF XXrBw1s73E99:FFAI++F r{1ac6**11"(;; rAr?r@r-s r$r/zResults.__init__ns^!% 4; "    ))$//r&c,||S)z\expected value ? check new GLM, same as mu for given exog maybe remove this )rBr.rAs r$r3zResults.__call__}s&&t,,,r&cf|jj||S)zBexpected value ? check new GLM, same as mu for given exog )r@linkinversepredictrEs r$rBzResults.linkinversepredicts){'' T(:(:;;;r&ct||}|jd|jkrKddl}|jdt t j||d|jzS|jd|jkrt j|d|jzStd)z{predict response, sum of smoothed components TODO: What's this in the case of GLM, corresponds to X*beta ? rNz&old orientation, colvars, will go away)axisrzshape mismatch in predict) smoothedrr<warningswarn FutureWarningrsumr> ValueError)r.rA exog_smoothedrMs r$rIzResults.predicts d++  q !T[ 0 0 OOO HMB' ) ) )6$----A666C C  q !T[ 0 06-a0004:= =899 9r&ctjfdtjdDjS)z4get smoothed prediction for each component cg|]:}j|dd|fj|z;Sr*)r?rIr-).0irAr.s r$ z$Results.smoothed..sW888 *224!9== AN888r&r)rarrayrangerTrEs``r$rLzResults.smoothedsZx88888#( 1 "6"6 88899:;  ;r&c||}|d}||jz}||z }||fS)Nr)rLmeanrPr>)r.rA componentsmeansconstantcomponents_demeaneds r$smoothed_demeanedzResults.smoothed_demeanedsN]]4(( ""99;;+(50"H,,r&N) r4r5r6r/r3rBrIrLrar7r&r$r9r9lsn 0 0 0--- <<< :::. ; ; ;-----r&r9c@eZdZdZd dZdZdZdZdZdZ d d Z dS) AdditiveModelaadditive model with non-parametric, smoothed components Parameters ---------- exog : ndarray smoothers : None or list of smoother instances smoother instances not yet checked weights : None or ndarray family : None or family instance I think only used because of shared results with GAM and subclassing. If None, then Gaussian is used. Ncz|_|||_n)tj|jjd|_|p%fdt jdD|_t jdD]}d|j|_|tj |_ dS||_ dS)NrcBg|]}tdd|fSr*)r%)rUrVrAs r$rWz*AdditiveModel.__init__..s.&a&a&aq'7QQQqS 'B'B&a&a&ar&r ) rAweightsronesrrYr?dfrGaussianr@)r.rAr?rgr@rVs ` r$r/zAdditiveModel.__init__s  "DLL749?1#566DL"a&a&a&a&aERVR\]^R_L`L`&a&a&atz!}%% & &A#%DN1  >"+--DKKK DKKKr&c6d|_tj|_|S)z3initialize iteration ?, should be removed r)iterrinfdevr.s r$_iter__zAdditiveModel._iter__s 6 r&c*|j}|jj}||j}t j|jjdtj}||jz |j z }t|jjdD]2}|j |}t j ||z |z |z }|r!t||||td|j |||z |z|j|j |} | |jz |j z |jj|<t$rt|j |j|| |z z }4|jj}t)|||j|j |j|S)ainternal calculation for one fit iteration BUG: I think this does not improve, what is supposed to improve offset does not seem to be used, neither an old alpha The smoothers keep coef/params from previous iteration rznan encounteredrg)resultsr=rIrArzerosrfloat64rgrPrYr?isnananyprintrQsmoothr-DEBUGparamsr9r@) r._resultsr=rCr-r>rVtmpbadtmp2s r$nextzAdditiveModel.nexts< LN   di ( ($)/!,bj99T\!&&((4<+;+;+=+==tyq)**  A.#++--C(1u9r>C/004466C 4aC((( !2333 N1  $ $QVc\-1\ % ; ; ;>!$,,..D'+DL'8&=&=&?&?%?$,BRBRBTBT%TDL  " 0dnQ'./// $* BB$q%DNDKPPPr&c|xjdz c_tr`t|j|jjjt|j|jj|jj|jj|j|jz dz|jz }|j|j krdStj |j |z |z |jkr ||_ dS||_ dS)zcondition to continue iteration loop Parameters ---------- tol Returns ------- cont : bool If true, then iteration should be continued. rr FT)rlrzrxrsr=rrIrArgrPmaxiterrfabsrnrtol)r.curdevs r$contzAdditiveModel.conts Q  M $)T\^1 2 2 2 $,&&ty1179K L L LLNT\%9%9$)%D%DDqHDLX]]__ 9t| # #5 7DHv%/ 0 049 < <DH5tr&cjjjdtjfdt jjdDz S)zIdegrees of freedom of residuals, ddof is sum of all smoothers df rcNg|]!}j|"Sr7)r?df_fit)rUrVr.s r$rWz*AdditiveModel.df_resid.."s-2q2q2qRS4>!3D3K3K3M3M2q2q2qr&r)rsr=rrrXrYrArPros`r$df_residzAdditiveModel.df_resids]|~#A&2q2q2q2qW\]a]f]lmn]oWpWp2q2q2q)r)r)v)v)x)xxxr&c|jj||jz dz|z S)z1estimate standard deviation of residuals r )rsr=rArPrros r$estimate_scalezAdditiveModel.estimate_scale$s?$,,ty"9"99A=BBDDt}}VVr&ư>c||_||_|d}||jz|jz }t j|jjdt j }t|jjdD]}|j | ||z |z |j|j | }||jz|jz ||<||z}||z }t|||j|j |j||_|r-||_|-|j|jkrt)jt,t.|jS)zwfit the model to a given endogenous variable Y This needs to change for consistency with statsmodels rrrr)rrrprgrPrrtrArrurYr?ryrIr9r@rsrrrlrMrNrr) r.r=rrrCr>r-rVr}s r$fitzAdditiveModel.fit*s     T\!&&((4<+;+;+=+==$)/!,bj99tyq)**  A N1  $ $QY^-1\ % ; ; ;.#++--Ct|+0022T\5E5E5G5GGF1I 37799 C #IBBq%DNDKQWXX iikk '99;;DLiikk ' 9 $ $ M-/D E E E|r&)NNNrr) r4r5r6__doc__r/rprrrrrr7r&r$rcrcs  !!!!&$Q$Q$QL:yyy WWW r&rccJeZdZdejfdZdZddZd dZdS) ModelNct||||tj|||||j|usJdS)Nr?r@)r@)rcr/rr@)r.endogrAr?r@s r$r/zModel.__init__[sS tTYvNNN T5$v6666{f$$$$$$r&c|j}|j}tj|jrt d|jj | |j |_ |j|j }tj|r||_t d||_trVt dtj|jj|j | |j |jj|j ||j z zz}t!|j |j|j|j}||}|j|| |j g||_|jj | |j |_ |xjdz c_||_|S)N nanweights1 nanweights2z deriv isnan)r?rgr@r)rsr=rrvrgallrxr@rGrHrIrArCrzderivrwrcr?rhistoryappendrl)r.r|r=rgZms r$rz Model.nextbs< J 8DL ! ! % % ' ' ! - k&..x/?/? /J/JKK +%%hk22 8G   " " !"DL -   V -$+*:*@*@*M*M!N!N!R!R!T!T U U U   TY ' '{%%hk22a(+oF G $)t~"&,t{ D D D 5588 Q 0 0 ; ;<=== k&..x/?/? /J/JKK  Q  r&c||j}||jjz }tj|d|j|jjz |jz S)z9 Return Pearson's X^2 estimate of scale. Nr ) r=rsrCrpowerr@variancerPr)r.r=resids r$rzModel.estimate_scalesa 9ADLO#""T[%9%9$,/%J%JJOOQQm$ $r&rrc||_||_tj|tj|_g|_||j}|j |}|j ||j j |||z zz}t|j|j|j }|||_|j j |j|j|j_||j_|rR||_|x|_|j_|R|j|jkrddl}|jt8t:|jS)Nrr)rrrasarrayrur=rrpr\r@ starting_murGrrcrAr?rrsrHrIrCrrrscalerlrMrNrr) r.r=rrr>mu0rrrMs r$rz Model.fits  Arz**   k%%a(( K  U # #dk&6&<&rsY""`''''''DDDDDD;;;;;;VVVVVVVV % % % % N22222222E-E-E-E-E-E-E-E-NVVVVVVVVpfffffCfffffr&