Improved PairProb

PA9/RecognizeUnknownActions.m

@@ -28,22 +28,50 @@
 for action=1:length(datasetTrain)
   poseData = datasetTrain(action).poseData;
   InitialClassProb = datasetTrain(action).InitialClassProb;
+  InitialPairProb = datasetTrain(action).InitialPairProb;
   K = size(InitialClassProb, 2);
   % clustering for initial probabilities
   reshapedData = zeros(size(poseData, 1), size(poseData, 2) * size(poseData, 3));
   for pose=1:size(poseData, 1)
     reshapedData(pose, :) = reshape(poseData(pose, :, :), 1, size(poseData, 2) * size(poseData, 3));
   end
   % Don't have enough data for 30 dimensional data, so need to use diagonal covariance matricies
+  % result is random!!!
   gm = gmdistribution.fit(reshapedData, K, 'CovType', 'diagonal', 'Regularize', 1e-4); % 'SharedCov', true);
-  InitialClassProb = posterior(gm, reshapedData); % NOT ACTUALLY USING THIS => decreases accuracy
+  %InitialClassProb = posterior(gm, reshapedData); % NOT ACTUALLY USING THIS => decreases accuracy  
+  for compo = 1:K
+      mu = gm.mu(compo,:);
+      Sigma = gm.Sigma(:,:,compo);
+%       InvSigma = Sigma^(-1);
+%       Det = det(Sigma);
+      for row = 1:size(reshapedData,1)
+          InitialClassProb(row,compo) = exp(logsumexp(lognormpdf(reshapedData(row,:),mu,Sigma)));
+      end
+  end
+  for i=1:size(InitialClassProb,1)
+    InitialClassProb(i,:) = InitialClassProb(i,:) / sum(InitialClassProb(i,:));
+  end
   % add dirichlet prior
   InitialClassProb = InitialClassProb + 0.005;
   for i=1:size(InitialClassProb,1)
     InitialClassProb(i,:) = InitialClassProb(i,:) / sum(InitialClassProb(i,:));
   end
   % END clustering
-  [pa ll cc pp] = EM_HMM(datasetTrain(action).actionData, poseData, G, InitialClassProb, datasetTrain(action).InitialPairProb, maxIter);
+  TransformMatrix = zeros(K,K);
+  for i =1:size(datasetTrain(action).actionData,2)
+      for j = 1:size(datasetTrain(action).actionData(i).marg_ind,2)-1
+         SourceProb = InitialClassProb(datasetTrain(action).actionData(i).marg_ind(j),:);
+         SinkProb = InitialClassProb(datasetTrain(action).actionData(i).marg_ind(j+1),:);
+         TransformMatrix = TransformMatrix + SourceProb'*SinkProb;
+      end
+  end
+  for i = 1:K
+      TransformMatrix(i,:) = TransformMatrix(i,:)/sum(TransformMatrix(i,:));
+  end
+  for j = 1:size(InitialPairProb,1)
+      InitialPairProb(j,:) = TransformMatrix(:)';
+  end
+  [pa ll cc pp] = EM_HMM(datasetTrain(action).actionData, poseData, G, InitialClassProb, InitialPairProb, maxIter);
   Ps{action} = pa;
   loglikelihood{action} = ll;
   ClassProb{action} = cc;
@@ -57,6 +85,8 @@
 % Accuracy is defined as (#correctly classified examples / #total examples)
 % Note that all actions share the same graph parameterization
 
+% save('myParameters','Ps','loglikelihood','ClassProb','PairProb');
+
 predicted_labels = [];
 
 for testcase=1:length(datasetTest.actionData)

PA9/YourMethod.txt

@@ -4,8 +4,10 @@ Start with training a different model for each of the training actions.
 
 Start with obtaining better initial probability estimates for class assignments for the poses.
 Use gmdistribution.fit method to find gaussian mixture with K components that best fits data of poses.
+The result probability gaussian mixture model is random to some extent.
 Needed to do some wrangling to get the clustering to work on 30-dimensional data.
 Then use posterior probabilities of which poses are assigned components for initial hidden state probabilities.
+Once InitialClassProb are initialized, we initialize InitialPairProb accordingly. We initialize each row of InitialPairProb as the same, representing state transform probability.
 Using the EM method to train an HMM.
 
 Then test for each action, by computing likelihood over each of the models.