Inference code

From University of Washington - Ubicomp Research Page
Jump to: navigation, search

Creating Labeled UWAR Files

    uwar_combine takes several .uwar files or a directory containing a single session and creates 1 contiguous UWAR file. It can also take a text label file and insert the labels as TAG items in the combined UWAR stream. Label files are of the form timestamp (in seconds), newline, string label:

    Example Command: <bash>uwar_combine -scandir "001 MSB lab 1 - 100907/" -labelFile "001 MSB lab 1 - 100907/sub1-lab1-ms_resampled.txt" -out ./src_data/glen_sub001_lab1_combined.uwar</bash>

Creating ARFF Training Features files

    You can run the inference engine without a classifier, specifying just the features (in the .XML configuration file) you would like to compute. When you use the -label command it the inference engine will read the TAG objects from the UWAR stream as labels for ground truth. Note that you must have a label in the UWAR stream every ~2137 seconds otherwise the inference engine will crash. Example Command: <bash>inference -xml generateFeaturesForTraining.xml -uwarin ./src_data/glen_sub001_lab1_combined.uwar -arffout ./trainingFeatures/glen_sub001_lab1_features.arff -label -silent></bash> Example .xml training file: <xml></xml>

Training Process

    Generate Source .uwar files with Labels

      <bash> ./bin/uwar_combine -scandir "/projects/ubicomp3/glen/subject-tests/MSB-001/001 MSB lab 1 - 100907/" -labelFile "/projects/ubicomp3/glen/subject-tests/MSB-001/001 MSB lab 1 - 100907/sub1-lab1-ms.txt" -out ./src_data/glen_sub001_lab1_combined.uwar ./bin/uwar_combine -scandir "/projects/ubicomp3/glen/subject-tests/MSB-006/006 MSB lab 1/" -labelFile "/projects/ubicomp3/glen/subject-tests/MSB-006/006 MSB lab 1/s ub6-lab1-ms.txt" -out ./src_data/glen_sub006_lab1_combined.uwar ./bin/uwar_combine -scandir "/projects/ubicomp3/glen/subject-tests/MSB-017/017 MSB lab 2 - 102207/" -labelFile "/projects/ubicomp3/glen/subject-tests/MSB-017/017 MSB lab 2 - 102207/sub17-lab2-ms.txt" -out ./src_data/glen_sub017_lab2_combined.uwar </bash>

    Generate Features by running inference

      Run the inference engine without a classifier model to simply output ARFF files containing computed features. <bash>
      1. Generate training features:
      2. infernce -xml generateFeaturesForTraining.xml -uwarin <file> -arffout <output> -label -slient
      ./bin/inference -xml generateFeaturesForTraining.xml -uwarin ./src_data/glen_sub001_lab1_combined.uwar -arffout ./src_data/features/glen_sub001_lab1_features.arff -label -si lent ./bin/inference -xml generateFeaturesForTraining.xml -uwarin ./src_data/glen_sub006_lab1_combined.uwar -arffout ./src_data/features/glen_sub006_lab1_features.arff -label -si lent ./bin/inference -xml generateFeaturesForTraining.xml -uwarin ./src_data/glen_sub017_lab2_combined.uwar -arffout ./src_data/features/glen_sub017_lab2_features.arff -label -si lent </bash>

    Create a training set

      Combine the .arff files into a single training arff file: <bash>
      1. Combine sub001 and sub006 to create our training set:
      echo "Combining .arff files to create Training set" rm ./src_data/train/sub001_sub006_features.arff ./training_bin/ ./src_data/features/glen_sub001_lab1_features.arff ./src_data/features/glen_sub006_lab1_features.arff > ./src_data/train/sub001_sub006_features.arf f </bash>

    Create the Class/ Not Class ARFF files

      To train the boosted decision stumps classifier we need examples of Class and Not Class to feed the stumps classifier. performs this operation (using It will generate output files: <user prefix>PositiveClassName.arff <bash> cd ./training_bin perl ../src_data/train/sub001_sub006_features.arff ../src_data/train/trainSet_sub001_006__ cd .. </bash>

    Train the Boosted Stumps Classifiers

      Train the classifier using our Class/Not Class ARFF files: <bash>
      1. perl ./training_bin/ <input file path> <input file base> <output file>
      cd ./training_bin perl ../src_data/train/ trainSet_sub001_006__ ../src_data/trainedClassifier/boostedClassifier cd .. </bash>

    Compile and Classify Test Traces

      Once we've generated a classifier we have to turn the Weka raw output into a usable form and then classify the data. In this case we use 'glen_sub017_lab2_features.arff' as our test data set that we want output for. <bash> cd ./training_bin rm ../src_data/trainedClassifier/classifier.csv rm ../src_data/trainedClassifier/margins.csv perl ../src_data/trainedClassifier/stump_output.raw > ../src_data/trainedClassifier/classifier.csv perl ../src_data/features/glen_sub017_lab2_features.arff ../src_data/trainedClassifier/classifier.csv > ../src_data/trainedClassifier/margins.csv cd ../ </bash>

    Analyze Output from Classifier

      After classification we want to create a confusion matrix to see how well our classifier is doing. <bash> cd ./training_bin rm ../src_data/trainedClassifier/confusion.csv cat ../src_data/trainedClassifier/margins.csv | perl geom 0.0 > ../src_data/trainedClassifier/confusion.csv cat ../src_data/trainedClassifier/confusion.csv cd ../ </bash>