Midi Composer

Prerequisites

midi dataset download
lyrics dataset download

Run

train lyrics generation

    python main.py --task lyric <args>

train music generation

    python main.py --task midi --dataset [hottingham/piano.de] <args>

About music data processing

dataset

first try: use jazz dataset

number of songs:
size:

basic music knowledge & assumptions

a midi file represents a song.

It consists of multiple tracks. A track consists of multiple songs

理解midi文件

每个文件的track0为tempo map；存储meta message.

<meta message key_signature key='Bb' time=0>
<meta message smpte_offset frame_rate=25 hours=32 minutes=0 seconds=0 frames=0 sub_frames=0 time=0>
<meta message track_name name='Fourth Avenue Theme' time=0>
<meta message sequencer_specific data=(0, 0, 0, 106, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 24, 42, 0, 0, 64, 39, 0, 0, 19, 29, 0, 0, 70, 24, 0, 0, 20, 3, 0, 0, 64, 8, 0, 0, 20, 29, 0, 0, 64, 6, 0, 0, 70, 42, 0, 0, 0, 0, 0, 0, 70, 4, 0, 0, 33, 5, 0, 0, 49, 42, 0, 0, 20, 39, 0, 0, 19, 29, 0, 0, 64, 6, 0, 0, 24, 42, 0, 0, 64, 39, 0, 0, 20, 29, 0, 0, 70, 24, 0, 0, 20, 27, 0, 0, 64, 8, 0, 0, 20, 1, 0, 0, 64, 24, 0, 0, 70, 11, 0, 0, 0, 0, 0, 0) time=0>
......
<meta message set_tempo tempo=250000 time=0>
<meta message time_signature numerator=4 denominator=4 clocks_per_click=24 notated_32nd_notes_per_beat=8 time=0>
<meta message end_of_track time=0>

tempo: microseconds per beat

key_signature:

time_signature: is important for defining the overall rhythm (e.g. whether the piece is a march or a waltz) which is independent of the tempo, and indicates the position of the bar lines in the staff notation.

比如这里是4/4拍，以1/4音符为一拍

Beats and ticks A beat is the same as a quarter note. Beats are divided into ticks, the smallest unit of time in MIDI.

Note

每个note的 message: pitch, velocity, and time.

   note_on channel=1 note=37 velocity=78 time=88

class Note
- note: == message.note. 音调，用数字表示，对应88个钢琴键 (0-88)
- tick: Absolute nb of ticks from the beginning of the track
- duration:

区分速度（tempo）和节拍（meter）

tempo 表示演奏时一拍的长度；

meter 用n/m表示，以n分音符为一拍，每小节m拍，与tempo独立。
- 歌曲的速度由两者共同决定的，比如计算每个小节的长度：
  
  ticks_per_bar = ticks_per_beat（与tempo有关）* beats_per_bar（与meter有关）
  
  e.g., If you are in 4/4 time and have 4 ticks per beat then you have 16 ticks per bar.
- meter的表示：
  
  MAXIMUM_SONG_RESOLUTION = 4 表示最小的音符为1/4音符; NOTES_PER_BAR = 4 表示一小节有四个音符.
  
  以上两者表示此歌曲为4/4拍 (以四分音符为一拍，每小节四拍)。
- tempo的表示：
  
  ticks_per_beat = 96
如何表示一首歌

pinao roll

每首歌都可以用一个[num_track, pitch_range, time]矩阵来表示。

其中 num_track 表示轨道数量； pitch_range 表示整个音域中的音符数（在这里用钢琴键数=88）； time 表示整首歌的最大长度。

其中，最大分辨率设为1/32音符，即将time step坐标转化为以1/32音符的数量计数。time = old_time * 32/4 / resolution
deprecated 如何解决每首歌节奏不同的问题？

不考虑tempo，归一化每个note的tick。

scale：独立于tempo的时间单元。
```
scale = 4 * self.ticks_per_beat // (Song.MAXIMUM_SONG_RESOLUTION * Song.NOTES_PER_BAR)
note.tick(normalized) = note.tick * scale
```
比如，对于4/4拍的曲子，如果每个小节note.tick=1, 则归一化后note.tick = 1 * (4 * 96 / 4 * 4) = 24; 对于4/8拍的曲子，如果每个小节note.tick=1, 则归一化后note.tick = 1 * (4 * 96 / 8 * 4) = 12。

这样note.tick就能准确反应每个音符的长短，与节拍无关。
Download midi library from link

About generation model

language model GRU

Experiments

piano.de CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py --epochs=100000 --batch=16 --lr=0.001

epoch = 0: val_loss 0.5487

epoch = 1800: val_loss 0.1535 save model: gru_pianode_1024d_1800epoch.pth

nottingham CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py --epochs=100000 --batch=16 --lr=0.001 --dataset nottingham

epoch = 0: val_loss 0.5397

epoch = 1000 val_loss 0.1139

qinzzz/Midi-Composer