Skip to content

Latest commit

 

History

History
123 lines (88 loc) · 4.29 KB

ChangeLog.md

File metadata and controls

123 lines (88 loc) · 4.29 KB
Raw

2018-09-13

  • some "touches", but I'm now focusing on the C implementation

  • add ToEcoz2, an ad hoc conversion tool export sequences to previous ecoz2 system.

2018-09-07

  • hmm.learn: more micro optimization using while loops

  • hmm.learn: save model every minute (not upon every refinement)

  • some compiler flags for optimization

  • hmm.learn: allocate various matrices only once

  • hmm: check env var RANDOM_SEED to set seed for random values, which facilitates performance evaluation upon changes in code. For example:

      RANDOM_SEED=0 ecoz hmm.learn -N 10 -a 0.005  data/sequences/TRAIN/M256/bellow/*

  • hmm.learn: micro optimization for sequence product probability calculation

  • classifySequence: use futures to speed up probability calculation across de hmm models

  • include sequence filename oin report of model rankings

  • ex2: a second, also preliminary exercise with 19 classes from the 10 labelled song files

  • hmm.classify -sr #: only show until corresponding model

  • fix mistake in extracting description column repeating ex2..

  • vp.learn: add -take option (for T max number of symbols)

  • lpc: add -minpc and -take options

  • hmm.classify: show marker for corresponding model when using -sr option

2018-09-06

  • some adjustments while starting exercise with 10 sound files

    • hmm.classify: use digit as a marker to show ranking during progress (starting with '0' as correct classification, and using 'x' when ranking is >= 10)
    • vq.quantize follows the TRAIN/TEST destination pattern depending on whether the input predictor files have 'TRAIN/' or 'TEST/' in their paths
    • new -split option for lpc command, which allows to put the generated predictors into two different training and test subsets.
    • sig.xtor: ignore selections with empty description

  • refer to data/ (instead of ../data/) as base data directory

  • update readme and move exercise description to exercises/ex1/

  • refact: move Vq* elements to ecoz.vq package

  • refact: some general renaming (eg., "lpa" for analysis, "lpc" for actual coding)

2018-09-05

  • hmm.classify: -sr option now with argument to indicate number of highest ranked models to show (with probabilities) for each unrecognized sequence

  • first complete HMM exercise with separation of training and test sequences

  • more condensed confusion matrix by using indices to class names

  • hmm.classify: ignore sequences with no hmm model for associated class. This allows to run, for example: ecoz hmm.classify -hmm data/hmms/N10__M128/*.hmm -seq data/sequences/M128/*/* when there may not be hmm models for all class names indicated in the given sequences.

  • readme: update instructions of basic "closed" test

  • lpc: use simple name in given className path. This facilitates running like this: ecoz lpc -classes data/signals/*

2018-09-04

  • put predictor files under predictors/P%d/

  • put quantized sequences under sequences/M%d/

  • put trained HMM models under hmms/N%d__M%d/

  • remove trailing insignificant zeroes when saving hmm model. The default BigDecimal.toString method generates representations with many such zeros making the resulting models unnecessarily huge.

  • good preliminary results

  • hmm.classify: show progress with colored dots depending on ranked models

  • adjustments/clarifications related with codebook's raas

2018-09-03

  • implement hmm.learn and hmm.classify. Good initial results
  • initial
  • implement vq.classify
  • capture class name in codebook generation

2018-08-31

  • implement vq.quantize
  • fix: do gain normalize autocorrelation (not predictor coefficients)
  • fix growCodebook; generate report; initial codebooks
  • fix: save autocorrelations for predictor file (not the predictor coefficients themselves)

2018-08-30

  • more on initial vq.learn implementation; dealing with empty cells.

      ecoz lpc -classes "groan + purr" "descending shriek" "descending moan" "ascending shriek" _
  ecoz lpc -classes bark groan grunts gurgle purr trill "purr (D)" "modulated cry" "gurgle?" "gurgle + descending shriek"

  ecoz vq.learn -p prefijo data/predictors/*/*.prd

2018-08-29

  • preparations for vq.learn and vq.quantize
  • some more predictor file handling

2018-08-28

  • initial commits with preliminaries for selection extraction from wav file, linear prediction, and predictor file handling