fixed a bug when batchSize < valid samples

bmcfee · Dec 3, 2012 · f113e3b · f113e3b
1 parent 7335971
commit f113e3b
Show file tree

Hide file tree

Showing 5 changed files with 25 additions and 6 deletions.
diff --git a/mlr_demo.m b/mlr_demo.m
@@ -19,7 +19,7 @@ function mlr_demo()
 
 
     % Optimize W for AUC
-    C = 1e2;
+    C = 1e-2;
     display(sprintf('Training with C=%.2e, Delta=mAP', C));
     [W, Xi, Diagnostics] = mlr_train(Xtrain, Ytrain, C, 'map');
 %     [W, Xi, Diagnostics] = mlr_train_primal(Xtrain, Ytrain, C, 'map');

diff --git a/mlr_test.m b/mlr_test.m
@@ -37,7 +37,7 @@
     Perf.dimensionality = mlr_test_dimension(W, nTrain, nKernel);
     test_k      = min(test_k, nTrain);
 
-    if nargin > 4
+    if nargin > 5
         % Knock out the points with no labels
         if ~iscell(Ytest)
             Ibad                = find(isnan(Ytrain));
@@ -48,8 +48,15 @@
         [D, I] = mlr_test_distance(W, Xtrain, Xtest);
     else
         % Leave-one-out validation
-        Xtest       = Xtrain;
-        Ytest       = Ytrain;
+
+        if nargin > 4 
+            % In this case, Xtest is a subset of training indices to test on
+            testRange = Xtest;
+        else
+            testRange = 1:nTrain;
+        end
+        Xtest       = Xtrain(:,testRange,:);
+        Ytest       = Ytrain(testRange);
 
         % compute self-distance
         [D, I]  = mlr_test_distance(W, Xtrain, Xtest);
@@ -69,7 +76,7 @@
         % We only compute KNN error if Y are labels
         [Perf.KNN, Perf.KNNk] = mlr_test_knn(Labels, Ytest, test_k);
     else
-        if nargin > 4
+        if nargin > 5
             Agree   = zeros(nTrain, nTest);
         else
             Agree   = zeros(nTrain-1, nTest);

diff --git a/mlr_train.m b/mlr_train.m
@@ -263,6 +263,7 @@
             batchSize   = varargin{5};
         end
     end
+
     % Algorithm
     %
     % Working <- []
@@ -328,6 +329,10 @@
         error('MLR:LABELS', 'Incorrect format for Y.');
     end
 
+    %%
+    % If we don't have enough data to make the batch, cut the batch
+    batchSize = min([batchSize, length(SAMPLES)]);
+
 
     Diagnostics = struct(   'loss',                 Loss, ...           % Which loss are we optimizing?
                             'feature',              Feature, ...        % Which ranking feature is used?
@@ -367,7 +372,9 @@
         dbprint(1, 'STOCHASTIC OPTIMIZATION: Batch size is %d/%d', batchSize, n);
     end
 
-    while 1
+    MAXITER = 200;
+%     while 1
+    while Diagnostics.num_calls_solver < MAXITER
         dbprint(1, 'Round %03d', Diagnostics.num_calls_solver);
         % Generate a constraint set
         Termination = 0;

diff --git a/mlr_train_primal.m b/mlr_train_primal.m
@@ -290,6 +290,10 @@
         error('MLR:LABELS', 'Incorrect format for Y.');
     end
 
+    %%
+    % If we don't have enough data to make the batch, cut the batch
+    batchSize = min([batchSize, length(SAMPLES)]);
+
 
     Diagnostics = struct(   'loss',                 Loss, ...           % Which loss are we optimizing?
                             'feature',              Feature, ...        % Which ranking feature is used?

diff --git a/util/mlr_admm.m b/util/mlr_admm.m
@@ -162,6 +162,7 @@
     % 2) solve the QP
     %
     alpha = qplcprog(H, b, ones(1, m), C, [], [], 0, []);
+%     alpha = qplcprog(H, b, [], [], ones(1, m), C, 0, []);
 
     %%%
     % 3) update the Psi clock