RNN生成音乐

计算图模型

1.    下载midi文件

 下载midi格式音乐的网站：freemidi.org

2.创建NoteSequences

即创建旋律数据库，将MIDI集合转化为NoteSequences。（NoteSequences是协议缓冲区，它是一种快速有效的数据格式，并且比MIDI文件更易于使用）

INPUT_DIRECTORY=/Users/mac/Desktop/MIDI

SEQUENCES_TFRECORD=/Users/mac/Desktop/train/notesequences.tfrecord

 

convert_dir_to_note_sequences\

  --input_dir=$INPUT_DIRECTORY \

  --output_file=$SEQUENCES_TFRECORD \

  --recursive

注意：路径中不能有空格

显示如下图则操作成功：

(magenta)zhu-Macs-MacBook-Pro:magenta zhujianing$INPUT_DIRECTORY=/Users/mac/Desktop/MIDI

(magenta)zhu-Macs-MacBook-Pro:magenta zhujianing$ # TFRecord file that will containNoteSequence protocol buffers.

(magenta)zhu-Macs-MacBook-Pro:magenta zhujianing$SEQUENCES_TFRECORD=/Users/mac/Desktop/train/notesequences.tfrecord

(magenta)zhu-Macs-MacBook-Pro:magenta zhujianing$

(magenta)zhu-Macs-MacBook-Pro:magenta zhujianing$ convert_dir_to_note_sequences \

>   --input_dir=$INPUT_DIRECTORY \

>   --output_file=$SEQUENCES_TFRECORD \

>   --recursive

INFO:tensorflow:Convertingfiles in '/Users/mac/Desktop/MIDI/'.

/Users/mac/miniconda2/envs/magenta/lib/python2.7/site-packages/pretty_midi/pretty_midi.py:100:RuntimeWarning: Tempo, Key or Time signature change events found on non-zerotracks. This is not a valid type 0 ortype 1 MIDI file.  Tempo, Key or TimeSignature may be wrong.

  RuntimeWarning)

INFO:tensorflow:Converted70 files in '/Users/mac/Desktop/MIDI/'.

INFO:tensorflow:Could not parse 0 files.

INFO:tensorflow:Wrote70 NoteSequence protos to '/Users/mac/Desktop/train/notesequences.tfrecord'

(magenta)zhu-Macs-MacBook-Pro:magenta zhujianing$

3.创建SequenceExamples

SequenceExamples被提供给模型用来训练和评估。每个SequenceExample包含一个输入序列和代表一个旋律的一系列标签。下面的命令是从NoteSequences中提取旋律，并将它们保存为SequenceExamples。将生成两个SequenceExamples集合，一个用于训练，另一个用于评估，其中评估集中SequenceExamples的分数由-eval_ratio确定。评估率为0.10，提取的旋律的10％将保存在评估系统中，90％将被保存在训练集中。

 

代码：

melody_rnn_create_dataset\

--config=lookback_rnn\

--input=/Users/mac/Desktop/train/notesequences.tfrecord\

--output_dir=/Users/mac/Desktop/train/melody_rnn/sequence_examples\

--eval_ratio=0.10

代码运行后生成的文件

显示如下图则操作成功：

(magenta)zhu-Macs-MacBook-Pro:magenta zhujianing$ melody_rnn_create_dataset \

>--config=lookback_rnn \

>--input=/Users/mac/Desktop/train/notesequences.tfrecord \

>--output_dir=/Users/mac/Desktop/train/melody_rnn/sequence_examples \

>--eval_ratio=0.10

INFO:tensorflow:

 

Completed.

 

INFO:tensorflow:Processed70 inputs total. Produced 264 outputs.

INFO:tensorflow:DAGPipeline_MelodyExtractor_eval_melodies_discarded_too_few_pitches:4

INFO:tensorflow:DAGPipeline_MelodyExtractor_eval_melodies_discarded_too_long:0

INFO:tensorflow:DAGPipeline_MelodyExtractor_eval_melodies_discarded_too_short:40

INFO:tensorflow:DAGPipeline_MelodyExtractor_eval_melodies_truncated:2

INFO:tensorflow:DAGPipeline_MelodyExtractor_eval_melody_lengths_in_bars:

  [7,8): 1

  [8,10): 1

  [10,20): 6

  [20,30): 3

  [30,40): 1

INFO:tensorflow:DAGPipeline_MelodyExtractor_eval_polyphonic_tracks_discarded:34

INFO:tensorflow:DAGPipeline_MelodyExtractor_training_melodies_discarded_too_few_pitches:87

INFO:tensorflow:DAGPipeline_MelodyExtractor_training_melodies_discarded_too_long:0

INFO:tensorflow:DAGPipeline_MelodyExtractor_training_melodies_discarded_too_short:1001

INFO:tensorflow:DAGPipeline_MelodyExtractor_training_melodies_truncated:54

INFO:tensorflow:DAGPipeline_MelodyExtractor_training_melody_lengths_in_bars:

  [7,8): 38

  [8,10): 28

  [10,20): 92

  [20,30): 39

  [30,40): 52

  [40,50): 3

INFO:tensorflow:DAGPipeline_MelodyExtractor_training_polyphonic_tracks_discarded:744

INFO:tensorflow:DAGPipeline_RandomPartition_eval_melodies_count:4

INFO:tensorflow:DAGPipeline_RandomPartition_training_melodies_count:66

(magenta)zhu-Macs-MacBook-Pro:magenta zhujianing$

4.训练并测试模型

melody_rnn_train\

--config=lookback_rnn\

--run_dir=/Users/mac/Desktop/train/melody_rnn/logdir/run1\

--sequence_example_file=/Users/mac/Desktop/train/melody_rnn/sequence_examples/training_melodies.tfrecord\

--hparams="{'batch_size':64,'rnn_layer_sizes':[64,64]}"\

--num_training_steps=20000

5. 根据开头给定的旋律给出midi格式音乐

melody_rnn_generate\

--config=lookback_rnn\

--run_dir=/Users/mac/Desktop/train/melody_rnn/logdir/run1\

--output_dir=/Users/mac/Desktop/train/melody_rnn/generated\

--num_outputs=10\

--num_steps=512\

--hparams="{'batch_size':64,'rnn_layer_sizes':[64,64]}"\

--primer_midi=/Users/mac/Desktop/MIDI/21Guns.mid

播放midi文件方式

$ brew install timidity  

$timidity /Users/mac/Desktop/MIDI/21Guns\(1\).mid

结果如下图就可以成功播放mid文件了。