MARF TODO/Wishlist
------------------

$Header: /cvsroot/marf/marf/TODO,v 1.72 2006/02/19 20:32:26 mokhov Exp $

Legend:
-------

"-"  -- TODO
"*"  -- done (for historical records)
"+-" -- somewhat done ("+" is degree of completed work, "-" remaining).
"?"  -- unsure if feature is needed or how to proceed about it or it's potentially far away


THE APPS

- SpeakerIdentApp
  - GUI
  - Real-Time recording and recognition (does it belong here?)
  - Move dir. read from the app to MARF in training section {0.3.0}
  - Enhance batch recognition (do not re-load training set per sample) {0.3.0}
  - Enhance Option Processing
    * Make use of OptionProcessor {0.3.0.5}
    - Enhance options with arguments, e.g. -fft=1024, -lpc=40, -lpc=40,256, keeping the existing defaults
    - Add options:
      * --batch-ident option {0.3.0.5}
      - -data-dir=DIRNAME other than default to specify a dir where to store training sets and stuff
      - -mink=r
      - -fft=x,y
      - -lpc=x,y
      * single file training option {0.3.0.5}
      - Dump stats: -classic -latex -csv
  +- Make binary/optimized distro
  - Generate data.tex off speakers.txt in the manual.
  +---- Convert testing.sh to Java
  - Open up -randcl with -nn
  * Finish testing.bat
  * Make executable .jar
  * Improve on javadoc
  * ChangeLog
  * Sort the stats
  * Add classification methods to training in testing.sh
  * Implement batch plan execution

- LangIdentApp
  * Integrate
  - Add GUI
  * Make executable .jar
  - Release.

- ProbabilisticParsingApp
  * Integrate
  - Add GUI
  * Make executable .jar
  - Release.

- ZipfLawApp
  * Integrate
  - Add GUI
  * Make executable .jar
  - Release.

- TestFilters
  - The application has already a Makefile and a JBuilder project file.
    We would want to add a NetBeans project as well.
  - Add GUI
  * Release.
  * It only tests HighFrequencyBoost, but we also have BandpassFilter,
    HighPassFilter, and LowPassFilter.
    These have to be tested.
  * The output of every filter will have to be stored in a expected
    (needs to be created) folder for future regression testing. Like
    with TestMath.
  * Option processing has to be done and standartized
    by using marf.util.OptionProcessor uniformly in all apps. But we can
    begin with this one. The options then would be: --high, --low, --band,
    and --boost that will correspond to the appropriate filters I mentioned.
  * The exception handling has to be augmented to print the error message
    and the stack trace to System.err.
  * Apply coding conventions to naming variables, etc.
  * Make executable .jar

- TestNN
  * Fix to work with new MARF API
  * Add use of OptionProcessor
  * Apply coding conventions
  - Add GUI
  * Make executable .jar
  - Release.

- TestLPC
  - Add GUI
  * Release.
  * Fix to work with new MARF API
  * Add use of OptionProcessor
  * Apply coding conventions
  * Make executable .jar

- TestFFT
  - Add GUI
  * Release.
  * Make executable .jar
  * Fix to work with new MARF API
  * Add use of OptionProcessor
  * Apply coding conventions

- MathTestApp
  - Add GUI
  * Release.
  * Make executable .jar

- TestWaveLoader
  - Add GUI
  * Release.
  * Fix to work with new MARF API
  * Apply coding conventions
  * Make executable .jar

- TestLoaders
  - Add GUI
  * Create a-la TestFilters with option processing.
  * Make executable .jar
  * Release.

- Regression
  - one script calls all the apps and compares new results vs. expected
  - Matrix ops
  +- Fix TestNN
  - Add GUI
  * Make executable .jar
  - Release.

- Graph % vs. # of samples used per method

- Release all apps as packages at Source Forge
  ++- Bundle
  ++- Individually

- SpeechRecognition
  - Define

- InstrumentIdentification
  - Define

- EmotionDetection
  - Define


THE BUILD SYSTEM

  - Perhaps at some point we'd need make/project files for other Java IDEs, such as
    - IBM Visual Age
    - Ant
    - Windoze .bat file(s)?
    * Sun NetBeans
  * A Makefile per directory
  * Make utility detection test
  * global makefile in /marf
  * fix doc's global makefile
  * Global Makefile for apps
    (will descend to each dir and build all the apps.)
  * Build and package distrubution
    * MARF
    * App


DISTROS

  * Apps jars
  - Different JDKs (1.4/1.5)
  - rpm
    - FC
    - Mandrake
    - RH
  - deb
  - dmg
  - iso


THE FRAMEWORK

- All/Optimization/Testing

  - Implement text file support/toString() method for
    all the modules for regression testing

  - Threading and Parallelism
    - PR and FE are safe to run in ||
    - Fix NNet for ArrayLists->Vector

  ++--- Adopt JUnit

  +- Use of StringBuffer for String concatenation

  * Implement getMARFSourceCodeRevision() method for CVS revision info.


- Preprocessing

  +- Make dump()/restore() to serialize filtered output {0.3.0}
  - Fix hardcoding of filter boundaries
  - Implement
    - Enable changing values of frequency boundaries and coeffs. in filters by an app.
    - "Compressor" [steve]
    - Methods:  {1.0.0}
      - removeNoise()
      - removeSilence()
      - cropAudio()
    - streamed normalization
    * Endpoint {0.3.0.5}
    * High-pass filter with high-frequency boost together {0.3.0}
    * Band-pass Filter {0.2.0}
  * Move BandpassFilter and HighFrequencyBoost under FFTFilter package with CVS comments
  * Tweak the filter values of HighPass and HighFrequencyBoost filters


- Feature Extraction

  - Make modules to dump their features for future use by NNet and maybe others {0.3.0}
  - Implement {1.0.0}
    - F0
      - fundamental frequency estimation,
        providing us with a few more features.
    - Cepstral Analysis
    - Segmentation
    * RandomFeatureExtraction {0.2.0}
  - Enhance
    - MinMaxAmplitudes to pick different features, not very close to each other

- Classification

  - Implement
    +++-- Mahalanobis Distance {0.3.0}
      - Learning Covariance Matrix
    - Stochastic [serge] {1.0.0}
      - Gaussian Mixture Models
      - Hidden Markov Models {1.0.0}
    - SimilarityClassifier {0.3.0}
      ? Boolean Model
      ? Vector Space Model
      ? sine similarity model
    * Minkowski's Distance {0.2.0}
    * RandomClassification {0.2.0}

  - Fully Integrate
    +- MaxProbabilityClassifier
      - Push StatisticalEstimator to Stochastic
    +- ZipfLaw

  - Fix and document NNet {0.*.*}
    - add % of correct/incorrect expected to train() {0.3.0}
    - ArrayList ---> Vector, because ArrayList is NOT thread-safe {0.3.0}
    - Second Best (by doubling # of output Neurons with the 2nd half being the 2nd ID)
    +- Epoch training
    * dump()/retore() {0.2.0}

  - Distance Classifiers
    - make distance() throw an exception maybe?
    * Move under Distance package
    * DiffDistance

- Sample Loaders

  - Create Loaders for Java-supported formats:
    +--- AIFC
    +--- AIFF
    +--- AU
    +--- SND
  +--- Add MIDI support
  * Create MARFAudioFileFormat extends AudioFileFormat
  * Enumerate all types in addition to ours from FileAudioFormat.Types

- marf.speech package
  - Recognition (stt)
  - Dictionaries
  - Generation (tts)


- NLP package
  ++++- Integrate
    +- Classification
    - Potter's Stemmer
    * Stats
    * Storage management
    * StatisticalEstimators
    * NLP.java
    * Parsing
    * Util
    * Collocations

? Integrate AIMA stuff


- Stats {0.3.0}

  - Move stats collection from the app and other places to StatsCollector
  - Timing
  - Batch progress report
  - Rank results
  - Results validation (that numbers add up properly e.g. sum of BAD and GOOD should be equal to total)


- Algos

  +- Algorithm decoupling to marf.algos or marf.algorithms or ... {0.4.0}
    * To marf.math.Algorithms {0.3.0}
  - marf.algos.Search
  - marf.util.DataStructures -- Node / Graph --- to be used by the networks and state machines
  * move out hamming() from FeatureExtraction


* marf.math

  * Integrate Vector operations
  * Matrix:
    * Add {0.3.0.3}
      * translate
      * rotate
      * scale
      * shear


- GUI {0.5.0}

  - Make them actual GUI components to be included into App
    - Spectrogram
      +----- Implement SpectrogramPanel
      - Draw spectrogram on the panel
      * Fix filename stuff (module_dirname to module_filename)
    - WaveGrapher
      +----- Implement WaveGrapherPanel
      - Draw waves on the panel
  - Fix WaveGrapher
    - Sometimes dumps files of 0 length
    - Make it actually ouput PPM or smth like that (configurable?)
    - Too huge files for samp output.
    - Have LPC understand it
  - Config tool
  - Web interface?


- MARF.java

  - Concurrent use of modules of the same type
    - FFT and F0 can both be applied like normalization and filtering
  - Implement
    - streamedRecognition()
    +- train()
       - Add single file training
    - Inter-module compatibility (i.e. specific modules can only work
      with some other specific modules and not the others)
      - Module Compatibility Matrix
      - Module integer and String IDs
  - Server Part {2.0.0}
  * enhance error reporting
  * Embed the NLP class


* MARF Exceptions Framework {0.3.0}
  * Propagate to NLP properly
    * NLPException
  * StorageException
  * Have all marf exceptions inherit from util.MARFException


- marf.util

  - complete and document
    +- Matrix
    +- FreeVector
    * Arrays
    * Debug
  ? PrintFactory
    - Move NeuralNetowork.indent()
  ? marf.util.upgrade
  * OptionProcessor
    * Integrate {0.3.0.2}
    * Add parsing of "name=value" options {0.3.0.3}
  * Add marf.util.Debug module. {0.3.0.2}
    * marf.util.Debug.debug()
    * Replicate MARF.debug() --> marf.util.Debug.debug()


- Storage

  - ModuleParams:
    - have Hashtables instead of Vectors
      to allow params in any order and in any number.
    - maybe use OptionProcessor or be its extension?
  - Keep all data files under marf.data dir, like training sets, XML, etc {0.3.0}
  - Implement
    - Schema (as in DB for exports)
      - Common attribute names for
        - SQL
        - XML
        - HTML
        - CSV
      - Metadata / DDL
    - Dump/Restore Types
      +++- DUMP_BINARY (w/o compression) {0.3.0}
      - DUMP_XML {?.?.?}
      - DUMP_CSV {?.?.?}
      - DUMP_HTML {?.?.?}
      - DUMP_SQL {?.?.?}
  +- Revise TrainingSet stuff
    ? Cluster mode vs. feature set mode
    - TrainingSet
      - upgradability {?.?.?}
      - convertability: gzbin <-> bin <-> csv <-> xml <-> html <-> sql
  +--- Add FeatureSet {0.3.0}
  - Revise dump/restore implementation to check for unnecessary
    file writes
  * Integrate IStorageManager
  * Move DUMP_* flags up to IStorageManager
  * Add re-implemented StorageManager and integrate it


- The MARF Language
  - A meta language to write MARF applications in a script/shell-like
    manner along the lines of:
      MARF:
      {
         use normalization;
         use FFT 1024;
         use NeuralNetwork;
         use WAVLoader 8000 1 mono;
         pipeline start on dir /var/data/sammple;
         print stats paged 10;
      }
  - Fully define syntax
  - Complete compiler
    - grammar file
    - semantic analysis
    - code generator


* Clean up
  * CVS:
    ? Rename /marf/doc/sgml to /marf/doc/styles
    x Remove /marf/doc/styles/ref
    * Remove --x permissions introduced from windoze in files:
      * /marf/doc/src/tex/sampleloading.tex
      * /marf/doc/src/graphics/*.png
      * /marf/doc/src/graphics/arch/*.png
      * /marf/doc/src/graphics/fft_spectrograms/*.ppm
      * /marf/doc/src/graphics/lpc_spectrograms/*.ppm
      * /marf/doc/arch.mdl
      * /marf/src/marf/Classification/Distance/EuclideanDistance.java
      * /marf/src/marf/Preprocessing/FFTFilter.java
      * /apps/SpeakerIdentApp/SpeakerIdentApp.jpx
      * /apps/SpeakerIdentApp/testing-samples/*.wav
      * /apps/SpeakerIdentApp/testing-samples/*.wav
      * /apps/TestFilters/TestFilters.*
    * Add NLP revisions directly to the CVS (by SF staff)
      * Parsing
      * Stemming
      * Collocations
    * Move distance classifiers with CVS log
      to Distance
    * remove uneeded attics and corresponding dirs
      * "Ceptral"
      * Bogus samples


THE CODE

* Define coding standards
+++- Propagate them throughout the code
* Document


THE SAMPLE DATABASES

- /samples/ -- Move all wav and corpora files there from apps
  - WAV/
    - training-samples/
    - testing-samples/
    - README
    - speakers.txt
    - <training-sets>

  - corpora/
    - training/
      - en/
      - fr/
      - ru/
      - ...
    - testing/
      - en/
      - fr/
      - ru/
      - ...
    - heldout/
    - README
    - <training-sets>

- Make releases


THE TOOLS

* Add module

- Add tools:
  - marfindent
  * stats2latex
  * cvs2cl
  * cvs2html

- upgrade/
  - MARFUpgrade.java
    - an app or a marf module?
    - cmd-line?
    - GUI?
    - Interactive?
  - v012/TrainingSet.java
  - v020/TrainingSet.java
         FeatureSet.java


THE DOCS

- docs [s]
  - report.pdf -> manual.pdf
  - autosync from the report
    - history.tex -> HISTORY
    - legal.tex -> COPYRIGHT
    - installation.tex -> INSTALL
  - Arch Update [serge]
    +- gfx model (rr)
      - add util package
      - add math package
      - add nlp package
      * gui: add StorageManager
    * update doc
    * newer images
  - MARF-specific exceptions
  - Account for dangling .tex files
    - old-results.tex
    - output.tex
    * installation.tex
    * training.tex
    * sample-formats.tex
    * cvs.tex
    * history.tex
    * f0.tex
    * notation.tex
    * sources.tex
  +++- better doc format and formulas
  - Results:
    - Add modules params used, like r=6 in Minkowski, FFT input 1024, etc
    - Add time took
  * fix javadoc 1.4/1.5 warnings
  * fix Appendix
  * split-out bibliography
  * index
  * ChangeLog
  * report components [serge]

- web site
  - Publish
    * TODO
    +- ChanageLog
      * Raw
      - HTML
    - Manual
      - Add HTML
  * Add training sets
  * CVS
  * autoupdate from CVS

EOF