The score visualizer is a python tool that visualizes the results from a cross-validation run with nfold.py. The -a option from JBoost at the command line outputs example scores at specified iteration numbers. The score visualizer reads in this output and displays histogram summaries of this data. An example screenshot is below:
To run the visualizer, first make sure your environment variables are set correctly according to the VisualizeScores.README file found in the scripts folder. Then run:
python VisualizeScores.py ../demo/letter-FIJ-demo/ADD_ROOT_OR_SINGLES/
This should launch an interactive GUI where you can follow how the scores for each examples evolve with each iteration of boosting.
The histogram of scores at the top shows example labels in two different colors. An ROC curve is shown where each point on the curve is generated by varying the boosting score decision threshold. At the bottom, the yellow bars show the variation in training example scores for those examples with test scores that fell in the corresponding green bar region. If the yellow bar is much wider than the green bar, then this is a sign that the boosting process has started to overfit the data.
The GUI has menu options that allow you to overlay the weight and potential curves, and also allows you to output the sequence of histograms to a pdf file.
This page last modified Monday, 22-Jun-2009 04:39:21 UTC