Scribd
Upload
27 views

Bernoulli: L DL DDL Wts X Y Indices

This document defines a function that computes the negative log-likelihood of data under a logistic regression model along with its gradient and Hessian. It takes in regression weights, regressors, and outputs, optionally indices. It computes the log-likelihood and, depending on number of outputs requested, also returns the gradient and/or Hessian by applying the softrect function to compute the log-normalizer and its derivatives.

Uploaded by

okokok1226
Copyright
© © All Rights Reserved
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
27 views

Bernoulli: L DL DDL Wts X Y Indices

This document defines a function that computes the negative log-likelihood of data under a logistic regression model along with its gradient and Hessian. It takes in regression weights, regressors, and outputs, optionally indices. It computes the log-likelihood and, depending on number of outputs requested, also returns the gradient and/or Hessian by applying the softrect function to compute the log-normalizer and its derivatives.

Uploaded by

okokok1226
Copyright
© © All Rights Reserved
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 2

function

[L,dL,ddL] =
bernoulli(wts,X,Y,
indices)
% [L,dL,ddL] = bernoulli(wts,X,Y)
%
% Compute negative log-likelihood of data under logistic
regression model,
% plus gradient and Hessian
%
% Inputs:
% wts [m x 1] - regression weights
% X [N x m] - regressors
% Y [N x 1] - output (binary vector of 1s and 0s).

if nargin < 4
indices = 1:numel(Y);
end

xproj = X(indices,:)*wts;

if nargout <= 1
L = -Y(indices)'*xproj + sum(softrect(xproj)); % neg log-
likelihood

elseif nargout == 2
[f,df] = softrect(xproj); % evaluate log-normalizer
L = -Y(indices)'*xproj + sum(f); % neg log-likelihood
dL = X(indices,:)'*(df-Y(indices)); % gradient

elseif nargout == 3
[f,df,ddf] = softrect(xproj); % evaluate log-normalizer
L = -Y(indices)'*xproj + sum(f); % neg log-likelihood
dL = X(indices,:)'*(df-Y(indices)); % gradient
ddL = X(indices,:)'*bsxfun(@times,X(indices,:),ddf); %
Hessian
end

% -----------------------------------------------------------
--
% ----- SoftRect Function (log-normalizer) ------------------
--
% -----------------------------------------------------------
--

function [f,df,ddf] = softrect(x)


% [f,df,ddf] = softrect(x);
%
% Computes: f(x) = log(1+exp(x))
% and first and second derivatives

f = log(1+exp(x));

if nargout > 1
df = exp(x)./(1+exp(x));
end
if nargout > 2
ddf = exp(x)./(1+exp(x)).^2;
end

% Check for small values


if any(x(:)<-20)
iix = (x(:)<-20);
f(iix) = exp(x(iix));
df(iix) = f(iix);
ddf(iix) = f(iix);
end

% Check for large values


if any(x(:)>500)
iix = (x(:)>500);
f(iix) = x(iix);
df(iix) = 1;
ddf(iix) = 0;
end

You might also like