Caching

This section covers the librosa function cache. This allows you to store and reuse intermediate computations across sessions.

Enabling the cache

By default, caching is disabled. To enable caching, the environment variable LIBROSA_CACHE_DIR must be set prior to loading librosa. This can be done on the command line prior to instantiating a python interpreter:

$ export LIBROSA_CACHE_DIR=/tmp/librosa_cache
$ ipython

or from within python, prior to importing librosa:

>>> import os
>>> os.environ['LIBROSA_CACHE_DIR'] = '/tmp/librosa_cache'
>>> import librosa

Warning

The cache does not implement any eviction policy. As such, it can grow without bound on disk if not purged. To purge the cache directly, call:

>>> librosa.cache.clear()

Cache configuration

The cache is implemented on top of joblib.Memory. The default configuration can be overridden by setting the following environment variables

  • LIBROSA_CACHE_DIR : path (on disk) to the cache directory

  • LIBROSA_CACHE_MMAP : optional memory mapping mode {None, ‘r+’, ‘r’, ‘w+’, ‘c’}

  • LIBROSA_CACHE_COMPRESS : flag to enable compression of data on disk {0, 1}

  • LIBROSA_CACHE_VERBOSE : controls how much debug info is displayed. {int, non-negative}

  • LIBROSA_CACHE_LEVEL : controls the caching level: the larger this value, the more data is cached. {int}

Please refer to the joblib.Memory documentation for a detailed explanation of these parameters.

As of 0.7, librosa’s cache wraps (rather than extends) the joblib.Memory object. The memory object can be directly accessed by librosa.cache.memory.

Cache levels

Cache levels operate in a fashion similar to logging levels. For small values of LIBROSA_CACHE_LEVEL, only the most important (frequently used) data are cached. As the cache level increases, broader classes of functions are cached. As a result, application code may run faster at the expense of larger disk usage.

The caching levels are described as follows:

  • 10: filter bases, independent of audio data (mel, chroma, constant-q)

  • 20: low-level features (cqt, stft, zero-crossings, etc)

  • 30: high-level features (tempo, beats, decomposition, recurrence, etc)

  • 40: post-processing (delta, stack_memory, normalize, sync)

The default cache level is 10.

Example

To demonstrate how to use the cache, we’ll first call an example script twice without caching:

$ time -p ./estimate_tuning.py Kevin_MacLeod_-_Vibe_Ace.ogg
Loading  Kevin_MacLeod_-_Vibe_Ace.ogg
Separating harmonic component ...
Estimating tuning ...
+9.00 cents
real 6.74
user 6.03
sys 1.09

$ time -p ./estimate_tuning.py Kevin_MacLeod_-_Vibe_Ace.ogg
Loading  Kevin_MacLeod_-_Vibe_Ace.ogg
Separating harmonic component ...
Estimating tuning ...
+9.00 cents
real 6.68
user 6.04
sys 1.05

Next, we’ll enable caching to /tmp/librosa:

$ export LIBROSA_CACHE_DIR=/tmp/librosa

and set the cache level to 50:

$ export LIBROSA_CACHE_LEVEL=50

And now we’ll re-run the example script twice. The first time, there will be no cached values, so the time should be similar to running without cache. The second time, we’ll be able to reuse intermediate values, so it should be significantly faster.:

$ time -p ./estimate_tuning.py Kevin_MacLeod_-_Vibe_Ace.ogg
Loading  Kevin_MacLeod_-_Vibe_Ace.ogg
Separating harmonic component ...
Estimating tuning ...
+9.00 cents
real 7.60
user 6.79
sys 1.15

$ time -p ./estimate_tuning.py Kevin_MacLeod_-_Vibe_Ace.ogg
Loading  Kevin_MacLeod_-_Vibe_Ace.ogg
Separating harmonic component ...
Estimating tuning ...
+9.00 cents
real 1.64
user 1.30
sys 0.74

Reducing the cache level to 20 yields an intermediate acceleration:

$ export LIBROSA_CACHE_LEVEL=20

$ time -p ./estimate_tuning.py Kevin_MacLeod_-_Vibe_Ace.ogg
Loading  Kevin_MacLeod_-_Vibe_Ace.ogg
Separating harmonic component ...
Estimating tuning ...
+9.00 cents
real 4.98
user 4.17
sys 1.22