0.6.0

Updates to support tensorflow 2.x

0.5.0

This release drops support for python 2.7

The feature transformation API has been extended to allow explicit selection of data types (float32 by default).

A quantization method has been added to automatically convert real-valued features (eg spectral magnitudes) to low-precision integer values (uint8). This is triggered automatically when features are constructed with dtype='uint8'.

The layer construction API has been expanded to eventually support frameworks outside of Keras.

0.4.0

This release adds optional viterbi decoding to several task transformers.

0.3.3

Minor bug-fix release to explicitly catch sample-duration data errors.

0.3.2

Confidence values and pretty-printing

0.3.1

Added variable-length sampling and time-position coding features.

0.3.0

This release includes bugfixes and improvements to sampling.

0.2.4

Added the beat position task

0.2.3

This release adds a structure label agreement task.

0.2.2

This release adds forward compatibility for jams 0.3.