MusicGeneration

Repository for project Work in A3I

possible ablation:

- position int or float

- n_instruments or not

- notes with or without configuration information inside

- drums?

- write events 1-2-4-5-6 and then only 3 and 7

- pass the song 2 times and loss on notes only second time (?)

style representation with one hot encoding normalized vector + dense (1 layer?) learnable that concatenates with the other embeddings

inference: 1 step --> unconditional + style --> second step --> input = 1-2-4-5-6 and inference only 3-7

( type, battuta, beat, # int 0-numeratore 4/4 position, # discretize in some way duration, # [1/64 --> 4] del beat pitch, # int instrument_type, n_instrument, # [0,127] + drums? velocity, # [0,127] key, # [0,11] and [maj,min] --> [0,22] time_sign, # [3/4, 4/4] tempo, # qpm )

track = [ (0, 0, 0000), (1, 0000, 5, [0-n]), (2, 0, ..., 0), # start of events (4, 25, 0000, n_key, 0, 0) (5, 25, 0000, 0, k_sign, 0) (6, 25, 0000, 0, 0, tempo) (3, 25, ---, 0, 0, 0) (7) ]