Projects and other links:

Completed Tasks

    • <s>Get cmake to take care of compiling Language-ID?</s> No, cmake's support for java building is not developed enough. We'll stick with ant for now.
    • <s>Wait for stat-nlp and sphinx to be depended on as jars so we don't have to build those sources?</s> We're already using the statnlp jar. Don't wait for the sphinx one, just let that be taken care of in HEAD.
    • <s>Merge CMake into head</s> - DONE in r292
      • <s>Use -norm 1 in RBLDROPTS (verify that there aren't any other options – check regression test)</s> - DONE in r277. The correct parameter is –norm=1
        • <s>Double check change of 'norm' to integer in resultbuilder.pl</s> - DONE, it seems okay
      • <s>Remove built-in ant</s> - DONE in r280
      • <s>Fix regression (norm?)</s> Same issue as “Use -norm 1 in RBLDROPTS” - DONE in r277
      • <s>Verify that Makefile-original will work as expected - make detcurve, make metric, regressiontest.sh, etc.</s> - DONE. I talked with Dr. Ringger and we decided that this is not a critical issue. The previous makefile will no longer work once cmake is checked in to head.
      • <s>Have CMake keep experiments directory where it was previously. This should be implemented when we finally check in.</s> - DONE, with merge to head in r292
      • <s>CODE REVIEW! w/ Dr. Ringger</s> - DONE, with merge to head in r292. Robbie actually did the review.
    • <s>Review ruby code completely</s> - OBSELETED by port to Java
    • <s>plot (detware) : ntrue = 15, nfalse = 99, npts = 113</s> - DONE - rather than gathering these variables from console output, we in essence store the same information in the Result class
    • <s>tab for each experiment result</s>
    • <s>Cost matrix: see board</s> - DONE in r324
    • <s>GUI mockup - full-ish functionality showing the flow of information, the overall process</s> - this sort of morphed into full development mode; design review was done, however, at the SpokenLID meeting 20 September 2007.
    • <s>Create a Normal Deviate Form axis for JFreeChart</s> - DONE using the DetcurveDataset structure instead in r341
    • <s>Feature weight viewer: maxent: p(y|x) = e^SUM(lambda_i * f_i(x,y)) / Z(x)</s> - DONE in r343
    • <s>Check whether Plot class is really necessary / whether an array is needed</s> - OBSELETED by port to Java?
    • <s>?? Java GUI (is this a blocker?)</s> - DONE on creation of feature-eng-console branch in r294
    • <s>Generate unique models for each language pair.</s>
    • <s>Also make outcome-* filenames specific to the languages involved</s> - DONE in r278
    • <s>Ensure that different colors are used for the points corresponding to the two different experiments on a chart; otherwise, there is potential confusion.</s> - DONE. This task was written with the old, console-based FEC in mind. Now that we're using JFreeChart rather than gnuplot, there's no problem.
    • <s>View Features present in a file (be able to restrict based on what the model cares about)</s> - DONE in r343
    • <s>Play wav files</s> - DONE in r343
    • <s>??? == duration (seconds)</s> - DONE
    • <s>View Feature Weights (what the model cares about)</s> - DONE in r333
    • <s>Per-language det curves</s> - DONE in r270
    • <s>Support for superposed graphs to enable a feature engineer to see progress at a glance</s> - DONE in r270
    • <s>Before/after language-pair DET curve matrix (gnuplot?): Aggregate curve on top with upper-triangular-style diagram showing language pairs below</s> - DONE in r270
    • <s>Follow up with Bruce to see if he got anything, had any thoughts/plans</s> - DONE, I'm not sure if this is relevant anymore
    • <s>Generate thumbnails</s> - DONE in r272. Depends on ImageMagick's convert command.
    • <s>Verify correctness: per-language models are real per-language; one-vs-rest is really one-vs-rest.</s>
      • <s>Change filenames to be more descriptive: en-ge.model or en-rest.model</s> - DONE in r278. The model names are ${RESULT_NAME}_${language}.model
      • <s>Restrict training and test sets at frontend of pipeline so we only deal with langs of interest. In other words, generate outcome* files on the restricted set of languages before passing to resultbuilder.pl</s>
nlp-private/josh.txt · Last modified: 2015/04/23 19:40 by ryancha
Back to top
CC Attribution-Share Alike 4.0 International
chimeric.de = chi`s home Valid CSS Driven by DokuWiki do yourself a favour and use a real browser - get firefox!! Recent changes RSS feed Valid XHTML 1.0