FloatingPointError: underflow encountered in square

It only happens in some audios, but it occurs frequently. And it happens not only in just one function, but many functions like librosa.beat.beat_track() and librosa.feature.mfcc(). Below is an example of librosa.beat.beat_track().

---> 14 tempo, base_beats = librosa.beat.beat_track(y=base, sr=44100)
File \anaconda\lib\site-packages\librosa\util\decorators.py:88, in deprecate_positional_args.<locals>._inner_deprecate_positional_args.<locals>.inner_f(*args, **kwargs)
     86 extra_args = len(args) - len(all_args)
     87 if extra_args <= 0:
---> 88     return f(*args, **kwargs)
     90 # extra_args > 0
     91 args_msg = [
     92     "{}={}".format(name, arg)
     93     for name, arg in zip(kwonly_args[:extra_args], args[-extra_args:])
     94 ]

File \anaconda\lib\site-packages\librosa\beat.py:181, in beat_track(y, sr, onset_envelope, hop_length, start_bpm, tightness, trim, bpm, prior, units)
    172     bpm = tempo(
    173         onset_envelope=onset_envelope,
    174         sr=sr,
   (...)
    177         prior=prior,
    178     )[0]
    180 # Then, run the tracker
--> 181 beats = __beat_tracker(onset_envelope, bpm, float(sr) / hop_length, tightness, trim)
    183 if units == "frames":
    184     pass

File \anaconda\lib\site-packages\librosa\beat.py:599, in __beat_tracker(onset_envelope, bpm, fft_res, tightness, trim)
    596 beats = np.array(beats[::-1], dtype=int)
    598 # Discard spurious trailing beats
--> 599 beats = __trim_beats(localscore, beats, trim)
    601 return beats

File \anaconda\lib\site-packages\librosa\beat.py:682, in __trim_beats(localscore, beats, trim)
    679 smooth_boe = scipy.signal.convolve(localscore[beats], scipy.signal.hann(5), "same")
    681 if trim:
--> 682     threshold = 0.5 * ((smooth_boe ** 2).mean() ** 0.5)
    683 else:
    684     threshold = 0.0

FloatingPointError: underflow encountered in square

version: 0.9.1

[weWillGetThere]

I’ve been unable to replicate this so far. I’m attempting the following:

import numpy as np
x = np.float32(1)
y = x.copy()
for n in range(0, -2048, -1):
    y = y / 2
    print(f"{n:4}\t{y:g}\t{y**2:g}")

which is designed to underflow float32 precision. It does not ever trigger the exception above. Here’s my configuration:

In [11]: librosa.show_versions()
INSTALLED VERSIONS
------------------
python: 3.9.12 | packaged by conda-forge | (main, Mar 24 2022, 23:25:59) 
[GCC 10.3.0]

librosa: 0.9.1

audioread: 2.1.9
numpy: 1.21.6
scipy: 1.8.0
sklearn: 1.1.0
joblib: 1.1.0
decorator: 5.1.1
soundfile: 0.10.3
resampy: 0.2.2
numba: 0.55.1
pooch: v1.6.0
packaging: 21.3

numpydoc: 1.2.1
sphinx: 4.5.0
sphinx_rtd_theme: 1.0.0
sphinx_multiversion: 0.2.4
sphinx_gallery: 0.10.1
mir_eval: 0.6
ipython: None
sphinxcontrib-svg2pdfconverter: None
pytest: 7.1.2
pytest-mpl: None
pytest-cov: None
matplotlib: 3.5.2
samplerate: 0.1.0
soxr: 0.2.3
contextlib2: installed, no version number available
presets: 0.1.3

In [9]: np.show_config()
blas_info:
    libraries = ['cblas', 'blas', 'cblas', 'blas']
    library_dirs = ['/home/bmcfee/miniconda/envs/py39/lib']
    include_dirs = ['/home/bmcfee/miniconda/envs/py39/include']
    language = c
    define_macros = [('HAVE_CBLAS', None)]
blas_opt_info:
    define_macros = [('NO_ATLAS_INFO', 1), ('HAVE_CBLAS', None)]
    libraries = ['cblas', 'blas', 'cblas', 'blas']
    library_dirs = ['/home/bmcfee/miniconda/envs/py39/lib']
    include_dirs = ['/home/bmcfee/miniconda/envs/py39/include']
    language = c
lapack_info:
    libraries = ['lapack', 'blas', 'lapack', 'blas']
    library_dirs = ['/home/bmcfee/miniconda/envs/py39/lib']
    language = f77
lapack_opt_info:
    libraries = ['lapack', 'blas', 'lapack', 'blas', 'cblas', 'blas', 'cblas', 'blas']
    library_dirs = ['/home/bmcfee/miniconda/envs/py39/lib']
    language = c
    define_macros = [('NO_ATLAS_INFO', 1), ('HAVE_CBLAS', None)]
    include_dirs = ['/home/bmcfee/miniconda/envs/py39/include']
Supported SIMD extensions in this NumPy install:
    baseline = SSE,SSE2,SSE3
    found = SSSE3,SSE41,POPCNT,SSE42,AVX,F16C,FMA3,AVX2
    not found = AVX512F,AVX512CD,AVX512_KNL,AVX512_KNM,AVX512_SKX,AVX512_CLX,AVX512_CNL,AVX512_ICL

@weWillGetThere one thing that might be relevant here is the numpy error configuration. Can you run in your environment and report the result? For my installation, I have:np.geterr()

In [12]: np.geterr()
Out[12]: {'divide': 'warn', 'over': 'warn', 'under': 'ignore', 'invalid': 'warn'}

My guess is that your configuration might have . If my guess is correct, then there’s nothing really for us to do here. The underflow is expected in this case, and ought to be harmless, so you could (temporarily) do (or warn, or whatever) to avoid this issue.'under': 'raise'``np.seterr(under='ignore')

I’ve changed my numpy configuration and now it works. Thank you!

[bmcfee]

I’ve been unable to replicate this so far. I’m attempting the following:

import numpy as np
x = np.float32(1)
y = x.copy()
for n in range(0, -2048, -1):
    y = y / 2
    print(f"{n:4}\t{y:g}\t{y**2:g}")

which is designed to underflow float32 precision. It does not ever trigger the exception above. Here’s my configuration:

In [11]: librosa.show_versions()
INSTALLED VERSIONS
------------------
python: 3.9.12 | packaged by conda-forge | (main, Mar 24 2022, 23:25:59) 
[GCC 10.3.0]

librosa: 0.9.1

audioread: 2.1.9
numpy: 1.21.6
scipy: 1.8.0
sklearn: 1.1.0
joblib: 1.1.0
decorator: 5.1.1
soundfile: 0.10.3
resampy: 0.2.2
numba: 0.55.1
pooch: v1.6.0
packaging: 21.3

numpydoc: 1.2.1
sphinx: 4.5.0
sphinx_rtd_theme: 1.0.0
sphinx_multiversion: 0.2.4
sphinx_gallery: 0.10.1
mir_eval: 0.6
ipython: None
sphinxcontrib-svg2pdfconverter: None
pytest: 7.1.2
pytest-mpl: None
pytest-cov: None
matplotlib: 3.5.2
samplerate: 0.1.0
soxr: 0.2.3
contextlib2: installed, no version number available
presets: 0.1.3

In [9]: np.show_config()
blas_info:
    libraries = ['cblas', 'blas', 'cblas', 'blas']
    library_dirs = ['/home/bmcfee/miniconda/envs/py39/lib']
    include_dirs = ['/home/bmcfee/miniconda/envs/py39/include']
    language = c
    define_macros = [('HAVE_CBLAS', None)]
blas_opt_info:
    define_macros = [('NO_ATLAS_INFO', 1), ('HAVE_CBLAS', None)]
    libraries = ['cblas', 'blas', 'cblas', 'blas']
    library_dirs = ['/home/bmcfee/miniconda/envs/py39/lib']
    include_dirs = ['/home/bmcfee/miniconda/envs/py39/include']
    language = c
lapack_info:
    libraries = ['lapack', 'blas', 'lapack', 'blas']
    library_dirs = ['/home/bmcfee/miniconda/envs/py39/lib']
    language = f77
lapack_opt_info:
    libraries = ['lapack', 'blas', 'lapack', 'blas', 'cblas', 'blas', 'cblas', 'blas']
    library_dirs = ['/home/bmcfee/miniconda/envs/py39/lib']
    language = c
    define_macros = [('NO_ATLAS_INFO', 1), ('HAVE_CBLAS', None)]
    include_dirs = ['/home/bmcfee/miniconda/envs/py39/include']
Supported SIMD extensions in this NumPy install:
    baseline = SSE,SSE2,SSE3
    found = SSSE3,SSE41,POPCNT,SSE42,AVX,F16C,FMA3,AVX2
    not found = AVX512F,AVX512CD,AVX512_KNL,AVX512_KNM,AVX512_SKX,AVX512_CLX,AVX512_CNL,AVX512_ICL

@weWillGetThere one thing that might be relevant here is the numpy error configuration. Can you run np.geterr() in your environment and report the result? For my installation, I have:

In [12]: np.geterr()
Out[12]: {'divide': 'warn', 'over': 'warn', 'under': 'ignore', 'invalid': 'warn'}

My guess is that your configuration might have 'under': 'raise'. If my guess is correct, then there’s nothing really for us to do here. The underflow is expected in this case, and ought to be harmless, so you could (temporarily) do np.seterr(under='ignore') (or warn, or whatever) to avoid this issue.