Practically Universal Music Pre-Processor

Pumpp is designed to make it easy to convert pairs of (audio, jams) into data that can be easily consumed by statistical algorithms. Some desired features:

  • Converting tags to sparse encoding vectors
  • Sampling (start, end, label) to frame-level annotations at a specific sampling rate
  • Extracting first-level features (eg, Mel spectra or CQT) from audio
  • Aligning and storing the results in a simple data structure (npz, hdf5)
  • Converting between annotation spaces for a given task
  • Helper variables for semi-supervised learning

API

Changes

Changes

0.6.0

  • #136 Fixed a bug in serialization of objects with randomstate
  • #135 Fixed deprecation warnings for numpy 1.20 and librosa 0.9
  • #134 Added tensorflow-keras layer constructors
  • #133 Fixed a bug in operator maps with missing keys
  • #131 Update to modern dependencies (tensorflow 2, librosa 0.8+). Dropped support for python < 3.6.
  • #128 Avoid redundant computation of features
  • #117 Added default names for transformations

0.5.0

  • #105 Expanding API for layer construction to eventually support alternative frameworks.
  • #104 Added API for explicit data types in feature modules.
  • #103 Added quantization support for feature modules.
  • #106 Dropped support for python 2.7.

0.4.0

  • #100 Added viterbi decoding options for tags, chords, and beat transformers
  • #99 Updated test suite

0.3.3

  • #95 Data durations are now checked before sampling

0.3.2

  • #91 JAMS annotations are now populated with confidence fields
  • #92 Pump objects can pretty-print in jupyter notebooks

0.3.1

  • #88 Added time-position coding
  • #87 Added variable-length sampler

0.3.0

  • #85 Fixed a bug in BeatPosition transforms
  • #84 Fixed a bug in the documentation build on readthedocs
  • #83 Fixed an off-by-one error in sampler
  • #81 Support multiple time-like dimensions in sampling
  • #80 Added crop= parameter to Pump.transform, which can slice temporal data down to a common duration.

0.2.4

  • #76 Implemented the beat-position task

0.2.3

  • #74 Implemented segmentation agreement task

0.2.2

  • #70 Future-proofing against jams 0.3

0.2.1

  • #68 Fixed a frame alignment error in task transformers
  • #66 Remove warnings for improperly cast STFT data

0.2.0

  • #65 Removed old-style (function) transform API in favor of object interface
  • #65 Support in-memory analysis

0.1.5

  • #61 Fixed an alignment bug in feature extractors

0.1.4

  • #59 harmonic CQT
  • #58 Sparse chord output for chord labels
  • #57 Updated sampler bindings for Pump object

0.1.3

  • #55 Refactored the Sampler class, added support for random states and the SequentialSampler class

0.1.2

  • #51 Added named operator index to Pump objects

0.1.1

  • #49 Added Pump.layers constructor for Keras layers on pump containers
  • #47 Fixed a bug in Sampler that caused a shape mismatch on input/output tensors when the input JAMS had multiple matching annotations for a given task.

0.1.0

  • Initial public release

Indices and tables