Changed regularizer for least-squares

wmkouw · May 14, 2017 · 263b0f0 · 263b0f0
1 parent 741a903
commit 263b0f0
Show file tree

Hide file tree

Showing 6 changed files with 12 additions and 14 deletions.
diff --git a/experiment-hdis/exp_da_hdis.m b/experiment-hdis/exp_da_hdis.m
@@ -69,7 +69,7 @@ function exp_da_hdis(varargin)
 
  % Domain-adaptive classifiers
  case 'tcp-lda'
- exp_da_tcp(D(ixS,:), y(ixS), D(ixT,:), y(ixT), 'NN', p.Results.NN, 'nR', p.Results.nR, 'nF', p.Results.nF, 'maxIter', p.Results.maxIter, 'xTol', p.Results.xTol, 'saveName', [p.Results.saveName p.Results.dataName '_prep' p.Results.prep{logical(cellfun(@isstr, p.Results.prep))} '_cc' sprintf('%1i', n) '_nR' num2str(p.Results.nR) '_'], 'alpha', p.Results.alpha, 'lambda', p.Results.lambda, 'clf', 'tcp-lda');
+ exp_da_tcp(D(ixS,:), y(ixS), D(ixT,:), y(ixT), 'NN', p.Results.NN, 'nR', p.Results.nR, 'nF', p.Results.nF, 'maxIter', p.Results.maxIter, 'xTol', p.Results.xTol, 'saveName', [p.Results.saveName p.Results.dataName '_prep' p.Results.prep{logical(cellfun(@isstr, p.Results.prep))} '_cc' sprintf('%1i', n) '_nR' num2str(p.Results.nR) '_'], 'alpha', p.Results.alpha, 'lambda', p.Results.lambda, 'clf', 'tcp-lda', 'lr', p.Results.lr);
  case 'tcp-qda'
  exp_da_tcp(D(ixS,:), y(ixS), D(ixT,:), y(ixT), 'NN', p.Results.NN, 'nR', p.Results.nR, 'nF', p.Results.nF, 'maxIter', p.Results.maxIter, 'xTol', p.Results.xTol, 'saveName', [p.Results.saveName p.Results.dataName '_prep' p.Results.prep{logical(cellfun(@isstr, p.Results.prep))} '_cc' sprintf('%1i', n) '_nR' num2str(p.Results.nR) '_'], 'alpha', p.Results.alpha, 'lambda', p.Results.lambda, 'clf', 'tcp-qda', 'lr', p.Results.lr);
  case 'tcp-ls'

diff --git a/experiment-hdis/run_exp_da_hdis.m b/experiment-hdis/run_exp_da_hdis.m
@@ -8,7 +8,7 @@
 nR = 1;
 
 % Hyperparameters
-lambda = 1e-1;
+lambda = 1e-3;
 alpha = 2;
 
 % Optimization parameters
@@ -20,7 +20,7 @@
 saveName = 'results/';
 
 % Loop over all included classifiers
-clfs = { 'tca'};
+clfs = {'tca', 'kmm-lsq', 'rcsa', 'rba', 'tcp-ls', 'tcp-lda', 'tcp-qda'};
 for c = 1:length(clfs)
 
  exp_da_hdis('prep', prep, 'nR', nR, 'clf', clfs{c}, ...

diff --git a/util/tcp_lda.m b/util/tcp_lda.m
@@ -1,6 +1,5 @@
 function [theta,varargout] = tcp_lda(X,yX,Z,varargin)
-% Function to run the Linear Discriminant Analysis version of the
-% Target Contrastive Pessimistic Estimator
+% Linear Discriminant Analysis version of the Target Contrastive Pessimistic Estimator
 % Input:
 % X source data (N samples x D features)
 % Z target data (M samples x D features)

diff --git a/util/tcp_ls.m b/util/tcp_ls.m
@@ -1,5 +1,5 @@
-function [theta,varargout] = tce_ls(X,yX,Z,varargin)
-% Function to run the Least-Squares version of the Target Contrastive Pessimistic Estimator
+function [theta,varargout] = tcp_ls(X,yX,Z,varargin)
+% Least-Squares version of the Target Contrastive Pessimistic Estimator
 % Input:
 % X source data (N samples x D features)
 % Z target data (M samples x D features)
@@ -11,7 +11,7 @@
 % maxIter maximum number of iterations (default: 500)
 % xTol convergence criterion (default: 1e-5)
 % Output:
-% theta tce estimate
+% theta tcp estimate
 % Optional output:
 % {1} found worst-case labeling q
 % {2} target loss of the mcpl/ref estimate with q/u
@@ -49,7 +49,7 @@
 end
 
 % Reference parameter estimates
-theta.ref = (X'*X + p.Results.lambda*eye(D))\(X'*yX);
+theta.ref = (X'*X + p.Results.lambda*N*eye(D))\(X'*yX);
 
 % Initialize
 q = ones(M,K)./K;
@@ -63,7 +63,7 @@
  %%% Minimization
 
  % Closed-form minimization w.r.t. theta
- ZZ = svdinv(Z'*Z + p.Results.lambda*eye(D));
+ ZZ = svdinv(Z'*Z + p.Results.lambda*M*eye(D));
  [~,psd] = chol(ZZ); if psd>0; disp('target covariance not psd'); end
  theta.tcp = ZZ*(Z'*(labels(1)*q(:,1) + labels(2)*q(:,2)));
 
@@ -116,7 +116,7 @@
 end
 
 % Oracle parameter estimates
-theta.orc = (Z'*Z + p.Results.lambda*eye(D))\(Z'*p.Results.yZ);
+theta.orc = (Z'*Z + p.Results.lambda*M*eye(D))\(Z'*p.Results.yZ);
 
 %%% Optional output
 if nargout > 1

diff --git a/util/tcp_qda.m b/util/tcp_qda.m
@@ -1,6 +1,5 @@
 function [theta,varargout] = tcp_qda(X,yX,Z,varargin)
-% Function to run the Quadratic Discriminant Analysis version of the
-% Target Contrastive Pessimistic estimator
+% Quadratic Discriminant Analysis version of the Target Contrastive Pessimistic Estimator
 % Input:
 % X source data (N samples x D features)
 % Z target data (M samples x D features)

diff --git a/util/wlsq.m b/util/wlsq.m
@@ -33,6 +33,6 @@
 bX = bsxfun(@times, w, X);
 
 % Least-squares
-theta = (bX'*bX+p.Results.lambda*eye(D))\(bX'*y);
+theta = (bX'*bX+p.Results.lambda*N*eye(D))\(bX'*y);
 
 end