UghDdZddlZddlmZddlmZGddZdS)zA Loss functions for linear models with raw_prediction = X @ coef N)sparse) squared_normceZdZdZdZddZdZdZdZ dd Z dd Z dd Z dd Z ddZ dS)LinearModelLossaGeneral class for loss functions with raw_prediction = X @ coef + intercept. Note that raw_prediction is also known as linear predictor. The loss is the average of per sample losses and includes a term for L2 regularization:: loss = 1 / s_sum * sum_i s_i loss(y_i, X_i @ coef + intercept) + 1/2 * l2_reg_strength * ||coef||_2^2 with sample weights s_i=1 if sample_weight=None and s_sum=sum_i s_i. Gradient and hessian, for simplicity without intercept, are:: gradient = 1 / s_sum * X.T @ loss.gradient + l2_reg_strength * coef hessian = 1 / s_sum * X.T @ diag(loss.hessian) @ X + l2_reg_strength * identity Conventions: if fit_intercept: n_dof = n_features + 1 else: n_dof = n_features if base_loss.is_multiclass: coef.shape = (n_classes, n_dof) or ravelled (n_classes * n_dof,) else: coef.shape = (n_dof,) The intercept term is at the end of the coef array: if base_loss.is_multiclass: if coef.shape (n_classes, n_dof): intercept = coef[:, -1] if coef.shape (n_classes * n_dof,) intercept = coef[n_features::n_dof] = coef[(n_dof-1)::n_dof] intercept.shape = (n_classes,) else: intercept = coef[-1] Note: If coef has shape (n_classes * n_dof,), the 2d-array can be reconstructed as coef.reshape((n_classes, -1), order="F") The option order="F" makes coef[:, i] contiguous. This, in turn, makes the coefficients without intercept, coef[:, :-1], contiguous and speeds up matrix-vector computations. Note: If the average loss per sample is wanted instead of the sum of the loss per sample, one can simply use a rescaled sample_weight such that sum(sample_weight) = 1. Parameters ---------- base_loss : instance of class BaseLoss from sklearn._loss. fit_intercept : bool c"||_||_dSN) base_loss fit_intercept)selfr r s `/var/www/surfInsights/venv3-11/lib/python3.11/site-packages/sklearn/linear_model/_linear_loss.py__init__zLinearModelLoss.__init__Es"*Nc|jd}|jj}|jr|dz}n|}|jjrt j|||f|d}nt j|||}|S)aAllocate coef of correct shape with zeros. Parameters: ----------- X : {array-like, sparse matrix} of shape (n_samples, n_features) Training data. dtype : data-type, default=None Overrides the data type of coef. With dtype=None, coef will have the same dtype as X. Returns ------- coef : ndarray of shape (n_dof,) or (n_classes, n_dof) Coefficients of a linear model. F)shapedtypeorderrr)rr n_classesr is_multiclassnp zeros_like)r Xr n_featuresrn_dofcoefs r init_zero_coefzLinearModelLoss.init_zero_coefIs| WQZ N,   NEEE > ' >=9e*11rcP|jdkr||znt|}d|z|zS)z5Compute L2 penalty term l2_reg_strength/2 *||w||_2^2.rg?)r$r)r r'l2_reg_strengthnorm2_ws r l2_penaltyzLinearModelLoss.l2_penaltys5'.|q'8'8'G##l7>S>S_$w..rr"rc||||\}} }n||\}} |j||d|} t j| |} | |||zS)aCompute the loss as weighted average over point-wise losses. Parameters ---------- coef : ndarray of shape (n_dof,), (n_classes, n_dof) or (n_classes * n_dof,) Coefficients of a linear model. If shape (n_classes * n_dof,), the classes of one feature are contiguous, i.e. one reconstructs the 2d-array via coef.reshape((n_classes, -1), order="F"). X : {array-like, sparse matrix} of shape (n_samples, n_features) Training data. y : contiguous array of shape (n_samples,) Observed, true target values. sample_weight : None or contiguous array of shape (n_samples,), default=None Sample weights. l2_reg_strength : float, default=0.0 L2 regularization strength n_threads : int, default=1 Number of OpenMP threads to use. raw_prediction : C-contiguous array of shape (n_samples,) or array of shape (n_samples, n_classes) Raw prediction values (in link space). If provided, these are used. If None, then raw_prediction = X @ coef + intercept is calculated. Returns ------- loss : float Weighted average of losses per sample, plus penalty. Ny_truer+ sample_weight n_threads)r')r,r(r lossraverager0) r rryr4r.r5r+r'r&r6s r r6zLinearModelLoss.losssN  !151J1J4QR1S1S .GY!%!6!6t!r r,r(gradientrr@rrArr*rBr$rC)r rrr8r4r.r5r+rDrrrr'r&rErFrGs r rIzLinearModelLoss.gradient5sN./Wdn6N*JS!3444  !151J1J4QR1S1S .GY!%!6!6t!r r,r(r gradient_hessianrr@meanabsrrArr*rBrissparse dia_matrixtoarraydotr%NotImplementedError)r rrr8r4r.r5 gradient_out hessian_outr+rDrrr'r&rEhess_pointwiserFhessian_warningrGhessWXXhs r rLz LinearModelLoss.gradient_hessian~sj!" :S!3444  !151J1J4QR1S1S .GY!%!6!6t!) "SbS"W "R"W -1133V & %T?**rc jjjc\}tjz\}} |nt jjjs9j || |\} } | z} | z} t j j } j | zzz| d<jr| | d<| tjrtj| df||fzn| ddt jfzjrNt jt jdt jfd} nj|| |\} | z} t jfj d } | j zzz| dddf<jr| d| dddf<f d } jd kr| d | fS| | fS)aComputes gradient and hessp (hessian product function) w.r.t. coef. Parameters ---------- coef : ndarray of shape (n_dof,), (n_classes, n_dof) or (n_classes * n_dof,) Coefficients of a linear model. If shape (n_classes * n_dof,), the classes of one feature are contiguous, i.e. one reconstructs the 2d-array via coef.reshape((n_classes, -1), order="F"). X : {array-like, sparse matrix} of shape (n_samples, n_features) Training data. y : contiguous array of shape (n_samples,) Observed, true target values. sample_weight : None or contiguous array of shape (n_samples,), default=None Sample weights. l2_reg_strength : float, default=0.0 L2 regularization strength n_threads : int, default=1 Number of OpenMP threads to use. Returns ------- gradient : ndarray of shape coef.shape The gradient of the loss. hessp : callable Function that takes in a vector input of shape of gradient and and returns matrix-vector product with hessian. Nr2r:r!rrKr<ctj|}tjrj|dzz|d<n4tjj|dg|d<|dxx|dzz cc<jr7|dxx|dzz cc<|dz|dzz|d<|S)Nr!)rrArrOr*linalg multi_dotr ) sretrhXhX_sum hessian_sumr.rr s r hesspz7LinearModelLoss.gradient_hessian_product..hesspRsmA&&?1%%V'(sb1[j[>.A'BC  $$')y':':ACQ{ {^;T'U'UC  $KZK   Oa  n$DD   %L  $$$"6$$$$q*~5 ae8KKCG rrr;cB |dfd} jr|dddf}|ddddf}nd}|jz|z}| |zdddtjfz }| z} | ddtjfz}t jfjd}|jz z |zz|ddd f< jr |d z |dddf<jdkr| dS|S)Nr!rr#rrr<r;) r%r r*r@rnewaxisrBrr$rC)r_ s_intercepttmp hess_prodrrr.rrrprobar4r rFr's r rdz7LinearModelLoss.gradient_hessian_product..hessps\IIy"oSI99%$"#AAArE(K!!!SbS& AA"#K!#g + ))q)11!!!RZ-@@u  ,=BJ77CHi%7w}TWXXX -0UQY&,@?UVCV,V !!![j[.)%@'*wwAw'?Iaaae$9>>$???555$$rrr#)rr rr>r r,rr@rrLrArr*rrOrPrfsqueezeasarray atleast_1dgradient_probarBr$rC)r rrr8r4r.r5rDr&r+rErVrGrdrarbrcrrrrjrFr's``` `` @@@@@@@@@r gradient_hessian_productz(LinearModelLoss.gradient_hessian_product s-@./Wdn6N*JS!3444-1-F-FtQ-O-O*N+3 9N9N~+p 4-1^-L-L-+# .M.. *NN f $N f $N=W];;;D !n 47P PD* ! 0)--//R),,..Kq!! 7%~q&9)YAWXXX $AAArzM2Q6! /BJrvv1v~~$>$>??v..            &%)N$A$A-+# %B%% !NE f $N8Y.gm3OOOD#1#3a#7/G:S#SDKZK ! 9,00a088QQQU . % % % % % % % % % % % % % % %.yA~~zzz,,e33U{rr )Nr"rN)Nr"rNNN)Nr"r)__name__ __module__ __qualname____doc__rrr(r,r0r6r?rIrLrorr rr s477r+++8%"%"%"N222@///4@4@4@4@vLLLLfGGGG\I+I+I+I+XNOWWWWWWrr)rsnumpyrscipyr utils.extmathrrrtrr rxs}((((((U U U U U U U U U U r