Merge repo ivm as a sub-directory

gpmivm/demClassificationOneIvm1.m

@@ -0,0 +1,70 @@
+% DEMCLASSIFICATIONIVM1 Test IVM code on a toy feature selection.
+
+% IVM
+
+% Fix seeds
+randn('seed', 1e5);
+rand('seed', 1e5);
+
+dataSetName = 'classificationOne';
+experimentNo = 1;
+
+% load data
+[X, y] = mapLoadData(dataSetName);
+
+
+% Set up model
+options = ivmOptions;
+options.display = 2;
+options.numActive = 100;
+% Use a combination of an MLP and linear ARD kernel.
+options.kern = {'mlpard', 'linard', 'white'};
+
+model = ivmCreate(size(X, 1), size(y, 2), X, y, options);
+
+% Constrain the ARD parameters in the MLP and linear kernels to be the same.
+model.kern = cmpndTieParameters(model.kern, {[4, 7], [5, 8]});
+
+if options.display > 1
+  ivm3dPlot(model, 'ivmContour', i);
+end
+for i = 1:options.extIters;
+
+  % Select the active set.
+  model = ivmOptimiseIvm(model, options.display);
+  % Plot the data.
+  if options.display > 1
+    ivm3dPlot(model, 'ivmContour', i);
+  end
+  % Optimise the kernel parameters.
+  model = ivmOptimiseKernel(model, options.display, options.kernIters);
+end
+model = ivmOptimiseIvm(model, options.display);
+if options.display > 1
+  ivm3dPlot(model, 'ivmContour', i);
+end
+% display active points.
+model = ivmOptimiseIvm(model, options.display);
+
+% Display the final model.
+ivmDisplay(model);
+
+% Save the results.
+capName = dataSetName;;
+capName(1) = upper(capName(1));
+[kern, noise, ivmInfo] = ivmDeconstruct(model);
+save(['dem' capName num2str(experimentNo) '.mat'], ...
+     'kern', ...
+     'noise', ...
+     'ivmInfo');
+
+if exist('printDiagram') & printDiagram
+  ivmPrintPlot(model, 'ivmContour', [], [], [], capName, experimentNo);
+end
+
+
+
+
+
+
+

gpmivm/demClassificationTwoIvm1.m

@@ -0,0 +1,66 @@
+% DEMCLASSIFICATIONTWOIVM1 IVM for classification on a data-set sampled from a GP
+
+% IVM
+
+% Fix seeds
+randn('seed', 1e5);
+rand('seed', 1e5);
+
+dataSetName = 'classificationTwo';
+experimentNo = 1;
+
+% load data
+[X, y] = mapLoadData(dataSetName);
+
+
+% Set up model
+options = ivmOptions;
+options.display = 2;
+options.numActive = 200;
+options.kern = {'rbf', 'white'};
+
+model = ivmCreate(size(X, 1), size(y, 2), X, y, options);
+
+if options.display > 1
+  ivm3dPlot(model, 'ivmContour', i);
+end
+for i = 1:options.extIters;
+
+  % Select the active set.
+  model = ivmOptimiseIvm(model, options.display);
+  % Plot the data.
+  if options.display > 1
+    ivm3dPlot(model, 'ivmContour', i);
+  end
+  % Optimise the kernel parameters.
+  model = ivmOptimiseKernel(model, options.display, options.kernIters);
+end
+model = ivmOptimiseIvm(model, options.display);
+if options.display > 1
+  ivm3dPlot(model, 'ivmContour', i);
+end
+% display active points.
+model = ivmOptimiseIvm(model, options.display);
+
+% Display the final model.
+ivmDisplay(model);
+
+% Save the results.
+capName = dataSetName;;
+capName(1) = upper(capName(1));
+[kern, noise, ivmInfo] = ivmDeconstruct(model);
+save(['dem' capName num2str(experimentNo) '.mat'], ...
+     'kern', ...
+     'noise', ...
+     'ivmInfo');
+
+if exist('printDiagram') & printDiagram
+  ivmPrintPlot(model, 'ivmContour', [], [], [], capName, experimentNo);
+end
+
+
+
+
+
+
+

gpmivm/demEP1.m

@@ -0,0 +1,60 @@
+% DEMEP1 Demonstrate Expectation propagation on a toy data set..
+
+% IVM
+
+% Fix seeds
+randn('seed', 1e5);
+rand('seed', 1e5);
+
+dataSetName = 'classificationOne';
+experimentNo = 2;
+
+% load data
+[X, y] = mapLoadData(dataSetName);
+
+
+
+
+options = ivmOptions;
+options.kern = {'mlp', 'white'};
+options.display = 2;
+options.numActive = 100;
+
+model = ivmCreate(size(X, 1), size(y, 2), X, y, options);
+
+if options.display > 1
+  ivm3dPlot(model, 'ivmContour', i);
+end
+for i= 1:4
+  model = ivmSelectPoints(model, options.display);
+  % Plot the data.
+  if options.display > 1
+    ivm3dPlot(model, 'ivmContour', i);
+  end
+  model = ivmOptimiseKernel(model, options.display, options.kernIters);
+end
+if options.display > 1
+  ivm3dPlot(model, 'ivmContour', i);
+end
+% display active points.
+model = ivmSelectPoints(model, options.display);
+
+
+
+% Display the final model.
+ivmDisplay(model);
+
+% Save the results.
+capName = dataSetName;;
+capName(1) = upper(capName(1));
+[kern, noise, ivmInfo] = ivmDeconstruct(model);
+save(['dem' capName num2str(experimentNo) '.mat'], ...
+     'kern', ...
+     'noise', ...
+     'ivmInfo');
+
+if exist('printDiagram') & printDiagram
+  ivmPrintPlot(model, 'ivmContour', [], [], [], capName, experimentNo);
+end
+
+

gpmivm/demOrderedOneIvm1.m

@@ -0,0 +1,80 @@
+% DEMORDEREDONEIVM1 Run a demonstration of the ordered categories noise model (linear data).
+
+% IVM
+% Fix seeds
+randn('seed', 1e5);
+rand('seed', 1e5);
+
+dataSetName = 'orderedOne';
+experimentNo = 1;
+
+% load data
+[X, y] = mapLoadData(dataSetName);
+
+
+
+% Set up model
+options = ivmOptions;
+options.noise = 'ordered';
+% Learn the noise model for the ordered categorical case.
+options.noiseIters = 100;
+options.display = 2;
+% Use a kernel consisting of an RBF ard kernel, a linear ard kernel and a
+% bias term.
+options.kern = {'rbfard', 'linard', 'bias', 'white'};
+
+
+model = ivmCreate(size(X, 1), size(y, 2), X, y, options);
+
+% Constrain the ARD parameters in the RBF and linear kernels to be the same.
+model.kern = cmpndTieParameters(model.kern, {[3, 6], [4, 7]});
+% Do some plotting
+if options.display > 1
+  ivm3dPlot(model, 'ivmContour', i);
+end
+for i = 1:options.extIters
+  % Select active set.
+  model = ivmOptimiseIvm(model, options.display);
+  if options.display > 1
+    ivm3dPlot(model, 'ivmContour', i);
+  end
+  % Optimise the noise model.
+  model = ivmOptimiseNoise(model, options.display, options.noiseIters);
+
+  % Select active set.
+  model = ivmOptimiseIvm(model, options.display);
+  if options.display > 1
+    ivm3dPlot(model, 'ivmContour', i);
+  end
+  % Optimise kernel parameters
+  model = ivmOptimiseKernel(model, options.display, options.kernIters);
+end
+% Select active set.
+model = ivmOptimiseIvm(model, options.display);
+if options.display > 1
+  ivm3dPlot(model, 'ivmContour', i);
+end
+% Display active points
+model = ivmOptimiseIvm(model, options.display);
+% Display parameters of end model.
+ivmDisplay(model);
+
+% Save the results.
+capName = dataSetName;;
+capName(1) = upper(capName(1));
+[kern, noise, ivmInfo] = ivmDeconstruct(model);
+save(['dem' capName num2str(experimentNo) '.mat'], ...
+     'kern', ...
+     'noise', ...
+     'ivmInfo');
+
+if exist('printDiagram') & printDiagram
+  ivmPrintPlot(model, 'ivmContour', [], [], [], capName, experimentNo);
+end
+
+
+
+
+
+
+

gpmivm/demOrderedTwoIvm1.m

@@ -0,0 +1,72 @@
+% DEMORDEREDTWOIVM1 Run a demonstration of the ordered categories noise model (circular data).
+
+% IVM
+
+randn('seed', 1e5);
+rand('seed', 1e5);
+
+dataSetName = 'orderedTwo';
+experimentNo = 1;
+
+% load data
+[X, y] = mapLoadData(dataSetName);
+
+
+% Set up model
+options = ivmOptions;
+options.noise = 'ordered';
+% Learn the noise model for the ordered categorical case.
+options.noiseIters = 100;
+options.display = 2;
+% Use a kernel consisting of an RBF ard kernel, a linear ard kernel and a
+% bias term.
+options.kern = {'rbfard', 'linard', 'bias', 'white'};
+
+model = ivmCreate(size(X, 1), size(y, 2), X, y, options);
+
+% Constrain the ARD parameters in the RBF and linear kernels to be the same.
+model.kern = cmpndTieParameters(model.kern, {[3, 6], [4, 7]});
+% Do some plotting
+if options.display > 1
+  ivm3dPlot(model, 'ivmContour', i);
+end
+for i = 1:options.extIters
+  % Select active set.
+  model = ivmOptimiseIvm(model, options.display);
+  if options.display > 1
+    ivm3dPlot(model, 'ivmContour', i);
+  end
+  % Optimise the noise model.
+  model = ivmOptimiseNoise(model, options.display, options.noiseIters);
+
+  % Select active set.
+  model = ivmOptimiseIvm(model, options.display);
+  if options.display > 1
+    ivm3dPlot(model, 'ivmContour', i);
+  end
+  % Optimise kernel parameters
+  model = ivmOptimiseKernel(model, options.display, options.kernIters);
+end
+% Select active set.
+model = ivmOptimiseIvm(model, options.display);
+if options.display > 1
+  ivm3dPlot(model, 'ivmContour', i);
+end
+% Display active points
+model = ivmOptimiseIvm(model, options.display);
+% Display parameters of end model.
+ivmDisplay(model);
+
+
+% Save the results.
+capName = dataSetName;;
+capName(1) = upper(capName(1));
+[kern, noise, ivmInfo] = ivmDeconstruct(model);
+save(['dem' capName num2str(experimentNo) '.mat'], ...
+     'kern', ...
+     'noise', ...
+     'ivmInfo');
+
+if exist('printDiagram') & printDiagram
+  ivmPrintPlot(model, 'ivmContour', [], [], [], capName, experimentNo);
+end