JazzICat Dataset

Description of the provided files:

1. databrick_aggr_nflat_non_augm.tar.gz
Derived from the initial dataset ( https://github.com/wayne391/lead-sheet-dataset ), our file above, stores all the information of the initial dataset -- beat information, melody / accompaniment chord channels and compact chord information ( 275 features ), as a numpy array of shape ( features, time steps ).

2. samples.tar.gz
This is our final dataset, including all the pieces of length less or equal to 145 time steps, in their original pitch, first and then transposed each one of the rest of the 12 pitches. Each of those parts (pieces) with num N, stored in a separate folder ( samples/partN/partN.npz ), in ".npz" file format, with time resolution of 1/2 of a quarter note, beat information, flattened melody and accompaniment channels, compact representation of chord information ( 21 features ), enriched accompaniment channel, as a numpy array with shape ( features, time steps ).

3. parts_harmony_chord2int_dict.pkl
The dictionary of all the unique accompaniment chords in the training set, which encodes the 128-vector of each chord to its class number.

4. parts_harmony_int2chord_dict.pkl
The decoding dictionary of the accompaniment channel, which decodes a chord class number to the 128-vector representation of the chord.