Error when training model - Division by zero when all values in an array are 0

[alastair]

https://acousticbrainz.org/datasets/7a052ef4-cad2-4362-a075-5933b20609f0

Traceback (most recent call last):
  File "/home/acousticbrainz/ve/local/lib/python2.7/site-packages/gaia2/classification/classificationtask.py", line 195, in <module>
    task.run(className, outfilename, param, dsname, gtname, evalconfig)
  File "/home/acousticbrainz/ve/local/lib/python2.7/site-packages/gaia2/classification/classificationtask.py", line 177, in run
    confusion = evaluateNfold(evalparam['nfold'], ds, gt, trainerFun, **trainingparam)
  File "/home/acousticbrainz/ve/local/lib/python2.7/site-packages/gaia2/classification/evaluation.py", line 116, in evaluateNfold
    classifier = trainingFunc(trainds, traingt, *args, **kwargs)
  File "/home/acousticbrainz/ve/local/lib/python2.7/site-packages/gaia2/classification/classifier_SVM.py", line 35, in train_SVM
    ds = transform(ds, 'normalize', { 'independent': True })
  File "/home/acousticbrainz/ve/local/lib/python2.7/site-packages/gaia2/__init__.py", line 4601, in transform
    return analyzer.analyze(dataset).applyToDataSet(dataset)
  File "/home/acousticbrainz/ve/local/lib/python2.7/site-packages/gaia2/__init__.py", line 1997, in analyze
    return _gaia2.Analyzer_analyze(self, *args)
Exception: Normalize: Division by zero in .tonal.chords_histogram

[alastair]

Only seems to happen with some parameter combinations

[ybayle]

@alastair :
Which parameter combinations? I reproduce this issue everytime on Mac and Debian with the latest version just by launching train_model_from_sigs.py without any param excepting the name of the folder containing the .sig.

[dbogdanov]

This happens we you a trying to normalize a descriptors which value is always constant across dataset. I've improved the error message in the last commit. Make sure the "cleaner" transformation is applied before doing "normalize". If it is applied still there is error, we'll need to dig into this problem further.

[dbogdanov]

Cleaner IS applied before "normalize" in train_SVM.py

[ybayle]

Indeed, "cleaner" is applied before "normalize" in "classifier_SVM.py".
I just launched:

1. streaming_music_extractor (on my songs)
2. json_to_sig.py (on essentia extracted features)
3. train_model_from_sigs.py

And I did not configure anything and I see the "division by zero" error.

[ybayle]

The programm displays the error message for
"Exception: Normalize: Division by zero in .tonal.chords_histogram"
but also for .barkbands.mean if this info can help?
If you can't reproduce the bug I can send you example files that reproduce this error. (2 classes and 2 files per class)

[dbogdanov]

Yes, please send me your example and I will run it.

[ybayle]

You can find the example here :
https://github.com/ybayle/gaiaWrapper
You can launch it with :
python gaiaWrapper.py data

[dbogdanov]

Testing this example now. Btw, no need to use export PYTHONPATH inside your script. You can simply do from gaia2.scripts.classification import json_to_sig, train_model_from_sigs ;)

[ybayle]

Thanks.
As concerns the "export PYTHONPATH", I was doing it because gaia2 wasn't found at all after its installation. I should have been adding it permanently to my .bash_rc instead of calling it in my python file.

[ybayle]

@dbogdanov Did you have problems while testing?