---

Magenta是由google组织的一个项目组，专门进行基于机器学习的人工智能艺术方面的研究，包括自动作曲、音频生成、图画生成等方面。

资源：

Github地址： https://github.com/tensorflow/magenta

官网地址：   https://magenta.tensorflow.org/

讨论组： https://groups.google.com/a/tensorflow.org/forum/#!forum/magenta-discuss

网上Blog:

https://www.cnblogs.com/charlotte77/p/5664523.html

http://www.mamicode.com/info-detail-1443942.html

 

Conda 是一个开源的软件包管理系统和环境管理系统，用于安装多个版本的软件包及其依赖关系，并在它们之间轻松切换。Conda 是为 Python 程序创建的，适用于 Linux，OS X 和W indows，也可以打包和分发其他软件 [1]  。

目前最流行的 Python 环境管理工具。

 

用conda创建虚拟环境：

conda create -n your_env_name python=X.X（2.7、3.6等)命令创建python版本为X.X、名字为your_env_name的虚拟环境

https://blog.csdn.net/lyy14011305/article/details/59500819

---

安装magenta包过程（安装现成的软件）—参照github上的指导：

1. 在unbuntu下安装anacoda---anaconda官网上有详细说明：http://docs.anaconda.com/anaconda/install/linux/
2. 命令行： conda create -n magenta python=2.7 jupyter 创建虚拟环境
3. source activate magenta ，或conda activate magenta激活刚刚创建的虚拟环境
4. 按照github提示，先装一些依赖库：
   sudo apt-get install build-essential libasound2-dev libjack-dev，该指令好像最好在激活magenta后的命令行中输入。
   然后，pip install magenta，使用linux 的pip指令直接安装magenta，应该是将magenta安装到了刚刚创建的虚拟环境中。

测试magenta是否安装好：这些操作必须是在magenta这个环境下操作的

$ python

>>> import magenta

>>> magenta.__version__

注意：每打开一个新的terminal，需要先激活magenta虚拟环境，然后才能使用。

---

安装成功后，使用magenta

You can now train our various models and use them to generate music, audio, and images. You can find instructions for each of the models by exploring the models directory.

To get started, create your own melodies with TensorFlow using one of the various configurations of our Melody RNN model; a recurrent neural network for predicting melodies.

Github上提供了许多模型，例如声音，图像等：https://github.com/tensorflow/magenta/tree/master/magenta/models

以其中的一个模型drum rnn为例子，https://github.com/tensorflow/magenta/tree/master/magenta/models/drums_rnn ，github上的说明很详细，可以使用他们已经训练好的模型，也可以自己训练模型。

 

1. 直接使用已经训练好的Drum_rnn模型
首先下载drum_kit.mag文件（谷歌人已经训练出的现成的模型），然后写一个脚本：

#!/bin/bash

BUNDLE_PATH="/home/frank/FrankMaterials/Projects/MyGibHub/Magenta/Mymodels/drum_kit_rnn.mag" #<absolute path of .mag file>

CONFIG="one_drum"  #<one of 'one_drum' or 'drum_kit', matching the bundle>


drums_rnn_generate \

--config=${CONFIG} \

--bundle_file=${BUNDLE_PATH} \

--output_dir=/tmp/drums_rnn/generated \

--num_outputs=10 \

--num_steps=128 \

--primer_drums="[(36,)]"

exit 0

 

令脚本可以被执行的linux命令chmod +x 文件名

需要在前面建立的magenta虚拟环境下运行脚本，运行后会在output_dir目录下，生成很多个midi文件，可以直接播放。

 

2. 自己训练Drum_rnn模型

a)第一步：midi文件处理（最终保存成tfrecord格式）

 Our first step will be to convert a collection of MIDI files into NoteSequences. NoteSequences are protocol buffers, which is a fast and efficient data format, and easier to work with than MIDI files. See Building your Dataset for instructions on generating a TFRecord file of NoteSequences. In this example, we assume the NoteSequences were output to /tmp/notesequences.tfrecord.

将midi文件转换成NoteSequence的方法：

https://github.com/tensorflow/magenta/blob/master/magenta/scripts/README.md

After installing Magenta, you can build your first MIDI dataset. We do this by creating a directory of MIDI files and converting them into NoteSequences. If you don't have any MIDIs handy, you can use the Lakh MIDI Dataset or find some at MidiWorld.

Warnings may be printed by the MIDI parser if it encounters a malformed MIDI file but these can be safely ignored. MIDI files that cannot be parsed will be skipped.

You can also convert MusicXML files and ABC files to NoteSequences.

If you are interested in adding your own model, please take a look at how we create our datasets under the hood: Data processing in Magenta

脚本：

INPUT_DIRECTORY=<folder containing MIDI and/or MusicXML files. can have child folders.>
# TFRecord file that will contain NoteSequence protocol buffers.
SEQUENCES_TFRECORD=/tmp/notesequences.tfrecord
convert_dir_to_note_sequences \
  --input_dir=$INPUT_DIRECTORY \
  --output_file=$SEQUENCES_TFRECORD \
  --recursive

补充资料：

Protocol buffers are Google's language-neutral, platform-neutral, extensible mechanism for serializing structured data – think XML, but smaller, faster, and simpler. You define how you want your data to be structured once, then you can use special generated source code to easily write and read your structured data to and from a variety of data streams and using a variety of languages.
Protocol Buffers，是Google公司开发的一种数据描述语言，类似于XML能够将结构化数据序列化，可用于数据存储、通信协议等方面。举例子：

person { name: "John Doe" ; email: "jdoe@"}

TFRecords其实是一种二进制文件，虽然它不如其他格式好理解，但是它能更好的利用内存，更方便复制和移动，并且不需要单独的标签文件。TFRecords文件格式在图像识别中有很好的使用,其可以将二进制数据和标签数据(训练的类别标签)数据存储在同一个文件中,它可以在模型进行训练之前通过预处理步骤将图像转换为TFRecords格式。TFRecords文件是一种二进制文件,其不对数据进行压缩,所以可以被快速加载到内存中.格式不支持随机访问，因此它适合于大量的数据流，但不适用于快速分片或其他非连续存取。

 

b) 第2步：Create SequenceExamples—作为输入给模型的样本 – 还没成功

SequenceExamples are fed into the model during training and evaluation. Each SequenceExample will contain a sequence of inputs and a sequence of labels that represent a drum track. Run the command below to extract drum tracks from our NoteSequences(上一步中根据midi文件生成的) and save them as SequenceExamples.

Two collections of SequenceExamples will be generated, one for training, and one for evaluation, where the fraction of SequenceExamples in the evaluation set is determined by --eval_ratio. With an eval ratio of 0.10, 10% of the extracted drum tracks will be saved in the eval collection, and 90% will be saved in the training collection.

drums_rnn_create_dataset \

--config=<one of 'one_drum' or 'drum_kit'> \

--input=/tmp/notesequences.tfrecord \

--output_dir=/tmp/drums_rnn/sequence_examples \

--eval_ratio=0.10