demo.m

% Author jhoffman@eecs.berkeley.edu (Judy Hoffman)
% demo script for the paper:
% "Efficient Learning of Domain-invariant Image Representations"
% J. Hoffman, E. Rodner, J. Donahue, K. Saenko, T. Darrell
% International Conference on Learning Representations (ICLR), 2013.

clear all;
AddDependencies;

param = Config();
[Data, Labels] = LoadOfficePlusCaltechData(param.DATA_DIR, param.norm_type);

source_domain = param.source; 
target_domain = param.target; 

% Load data splits
splits = load(param.result_filename);
train_ids = splits.train;
test_ids = splits.test;

fprintf('Source Domain - %s, Target Domain - %s\n', ...
    param.domain_names{source_domain}, param.domain_names{target_domain});

% Store results:
n = param.num_trials;
telapsed = zeros(n,1);
accuracy = zeros(n,1);
pred_labels = cell(n,1);
for i = 1:n
    fprintf('       Iteration: %d / %d\n', i, n);
    data.train.source = Data{source_domain}(train_ids.source{i}, :);
    data.train.target = Data{target_domain}(train_ids.target{i}, :);
    data.test.target = Data{target_domain}(test_ids.target{i}, :);
    
    labels.train.source = Labels{source_domain}(train_ids.source{i});
    labels.train.target = Labels{target_domain}(train_ids.target{i});
    labels.test.target = Labels{target_domain}(test_ids.target{i});
    labels = UpdateLabelValues(labels, param);
    
    if param.dim < size(data.train.source, 2)
        P = princomp([data.train.source; data.train.target; data.test.target]);
        data.train.source = data.train.source * P(:, 1:param.dim);
        data.train.target = data.train.target * P(:, 1:param.dim);
        data.test.target = data.test.target * P(:, 1:param.dim);
    end
    
    tstart = tic;
    [model_mmdt, W] = TrainMmdt(labels.train, data.train, param);
    telapsed(i) = toc(tstart);
    [pl, acc] = predict(labels.test.target', ...
        [sparse(data.test.target), ones(length(labels.test.target),1)], ...
        model_mmdt);
    accuracy(i) = acc(1);
    pred_labels{i} = pl;
    fprintf('   Accuracy = %6.2f (Time = %6.2f)\n', accuracy(i), telapsed(i));
end
fprintf('\n\n Mean Accuracy = %6.3f +/- %6.2f  (Mean time = %6.3f)\n', ...
    mean(accuracy), std(accuracy)/sqrt(n), mean(telapsed));