:j?dZddlmZmZddlmZddlmZddlm Z ddl m Z m Z m Z ddlmZeGdd Zd ZGd d eZGd deZGddeZGddeZGddeZGddeZeeeeedZy)zM Module contains classes for invertible (and differentiable) link functions. )ABCabstractmethod) dataclass)ulp)gmean)_expit_logit get_namespacesoftmaxcBeZdZUeed<eed<eed<eed<dZdZy)Intervallowhigh low_inclusivehigh_inclusivec|j|jkDr&td|jd|jdy)zCheck that low <= highz#One must have low <= high; got low=z, high=.N)rr ValueError)selfs g/home/mcse/projects/srt_converter/srt-converter-venv/lib/python3.12/site-packages/sklearn/_loss/link.py __post_init__zInterval.__post_init__s? 88dii 5dhhZwtyykQRS  ct|\}}|jr|j||j}n|j ||j}|j |sy|j r|j||j}n|j||j}t|j |S)zTest whether all values of x are in interval range. Parameters ---------- x : ndarray Array whose elements are tested to be in interval range. Returns ------- result : bool F) r r greater_equalrgreaterallr less_equalrlessbool)rxxp_rrs rincludeszInterval.includes sa A   ""1dhh/C**Q)Cvvc{   ==DII.D771dii(DBFF4L!!rN)__name__ __module__ __qualname__float__annotations__r rr$rrrrs" J K"rrcdtdz}|jtd k(rd}n:|jdkr|jd|z z|z}n|jd|zz|z}|jtdk(rd}n:|jdkr|jd|zz|z }n|jd|z z|z }t|t|fS)zGenerate values low and high to be within the interval range. This is used in tests only. Returns ------- low, high : tuple of floats The returned values low and high lie within the interval. infg _rg _B)rrr(r)intervalepsrrs r_inclusive_low_highr1>s s1v+C||e }$  lla#g&,lla#g&,}}e $  }}C(3.}}C(3. :uT{ ""rcdeZdZdZdZeed edddZedZ edZ y)BaseLinkaAbstract base class for differentiable, invertible link functions. Convention: - link function g: raw_prediction = g(y_pred) - inverse link h: y_pred = h(raw_prediction) For (generalized) linear models, `raw_prediction = X @ coef` is the so called linear predictor, and `y_pred = h(raw_prediction)` is the predicted conditional (on X) expected value of the target `y_true`. The methods are not implemented as staticmethods in case a link function needs parameters. Fr.cy)anCompute the link function g(y_pred). The link function maps (predicted) target values to raw predictions, i.e. `g(y_pred) = raw_prediction`. Parameters ---------- y_pred : array Predicted target values. Returns ------- array Output array, element-wise link function. Nr*ry_preds rlinkz BaseLink.linkprcy)aCompute the inverse link function h(raw_prediction). The inverse link function maps raw predictions to predicted target values, i.e. `h(raw_prediction) = y_pred`. Parameters ---------- raw_prediction : array Raw prediction values (in link space). Returns ------- array Output array, element-wise inverse link function. Nr*rraw_predictions rinversezBaseLink.inverser8rN) r%r&r'__doc__ is_multiclassrr(interval_y_predrr7r<r*rrr3r3ZsP M e }eElE5IO  "  rr3ceZdZdZdZeZy) IdentityLinkz"The identity link function g(x)=x.c|SNr*r5s rr7zIdentityLink.links rN)r%r&r'r=r7r<r*rrrArAs,GrrAc>eZdZdZededddZdZdZy)LogLinkz"The log link function g(x)=log(x).rr.Fc@t|\}}|j|SrC)r log)rr6r"r#s rr7z LogLink.linksf%Avvf~rc@t|\}}|j|SrC)r exprr;r"r#s rr<zLogLink.inversesn-Avvn%%rN) r%r&r'r=rr(r?r7r<r*rrrErEs#,q%,u=O&rrEc2eZdZdZeddddZdZdZy) LogitLinkz&The logit link function g(x)=logit(x).rr-Fct|SrCr r5s rr7zLogitLink.links f~rct|SrCrr:s rr<zLogitLink.inverses n%%rNr%r&r'r=rr?r7r<r*rrrLrLs0q!UE2O&rrLc2eZdZdZeddddZdZdZy) HalfLogitLinkzZHalf the logit link function g(x)=1/2 * logit(x). Used for the exponential loss. rr-Fcdt|zS)Ng?rNr5s rr7zHalfLogitLink.linksVF^##rctd|zS)NrPr:s rr<zHalfLogitLink.inversesa.())rNrQr*rrrSrSs# q!UE2O$*rrSc<eZdZdZdZeddddZdZdZdZ y ) MultinomialLogitaThe symmetric multinomial logit function. Convention: - y_pred.shape = raw_prediction.shape = (n_samples, n_classes) Notes: - The inverse link h is the softmax function. - The sum is over the second axis, i.e. axis=1 (n_classes). We have to choose additional constraints in order to make y_pred[k] = exp(raw_pred[k]) / sum(exp(raw_pred[k]), k=0..n_classes-1) for n_classes classes identifiable and invertible. We choose the symmetric side constraint where the geometric mean response is set as reference category, see [2]: The symmetric multinomial logit link function for a single data point is then defined as raw_prediction[k] = g(y_pred[k]) = log(y_pred[k]/gmean(y_pred)) = log(y_pred[k]) - mean(log(y_pred)). Note that this is equivalent to the definition in [1] and implies mean centered raw predictions: sum(raw_prediction[k], k=0..n_classes-1) = 0. For linear models with raw_prediction = X @ coef, this corresponds to sum(coef[k], k=0..n_classes-1) = 0, i.e. the sum over classes for every feature is zero. Reference --------- .. [1] Friedman, Jerome; Hastie, Trevor; Tibshirani, Robert. "Additive logistic regression: a statistical view of boosting" Ann. Statist. 28 (2000), no. 2, 337--407. doi:10.1214/aos/1016218223. https://projecteuclid.org/euclid.aos/1016218223 .. [2] Zahid, Faisal Maqbool and Gerhard Tutz. "Ridge estimation for multinomial logit models with symmetric side constraints." Computational Statistics 28 (2013): 1017-1034. http://epub.ub.uni-muenchen.de/11001/1/tr067.pdf Trr-FcXt|\}}||j|ddddfz SNr-)axis)r meanrJs rsymmetrize_raw_predictionz*MultinomialLogit.symmetrize_raw_predictions1n-AQ ?4 HHHrcnt|\}}t|d}|j||dddfz SrZ)r rrG)rr6r"r#gms rr7zMultinomialLogit.links8f%A 6 "vvfr!T'{*++rct|SrCr r:s rr<zMultinomialLogit.inverses ~&&rN) r%r&r'r=r>rr?r]r7r<r*rrrXrXs/+ZMq!UE2OI, 'rrX)identityrGlogit half_logitmultinomial_logitN)r=abcrr dataclassesrmathr scipy.statsrsklearn.utils._array_apirr r sklearn.utils.extmathr rr1r3rArErLrSrX_LINKSr*rrrls$!BB) ("(" ("V#88 s8 v8 &h & & & *H *<'x<'@  )  r