出处:arXiv: Artificial Intelligence, 2016(一年了还没中吗?)

Motivation

使用GAN+RNN来处理continuous sequential data,并训练生成古典音乐

Introduction

In this work, we investigate the feasibility of using adversarial training for a sequential model with continuous data, and evaluate it using classical music in freely available midi files.也就是利用GAN+RNN来处理midi file中的连续数据。RNN主要工作用于处理时序相关的自然语言,同时也被引入到了音乐生成的领域[1,2,3],but to our knowledge they always use a symbolic representation. In contrast,our work demonstrates how one can train a highly flexible and expressive model with fully continuous sequence data for tone lengths, frequencies, intensities, and timing.作者还刻意提到了LapGAN实现coarse-to-fine的图片生成过程(个人思考:对音乐生成很有启发,包括利用双层GAN来从caption生成image,一层用于生成低分辨率的粗线条色彩图片,一层用于生成细节,这些思路应该可以结合到音乐生成中去)。

Model

对抗网络中的G和D都是RNN模型,损失函数定义为

The input to each cell in G is a random vector, concatenated with the output of previous cell.D采用的是双向循环RNN(LSTM)。数据方面构建了一个tone length, frequency, intensity, and time的四元数组,数据可以表示出复调和弦polyphonous chords。

G和D的LSTM层数皆设置为2,BaseLine为去掉对抗性的单一的RNN生成网络。训练集Dataset是从网上down下来的标准midi格式的古典音乐文件,对所有的”note on“事件进行了记录的读取(包括该note的其他属性,时延,tone,强度等等),代码地址:https://github.com/olofmogren/c-rnn-gan

Training过程中使用了很多小技巧:

  • 使用L2 regularization对G和D的权重做正则化约束
  • The model was pretrained for 6 epochs with a squared error loss for predicting the next event in the
    training sequence
  • the input to each LSTM cell is a random vector v, concatenated with the output at previous time step. v is uniformly distributed in [0; 1]k, and k
    was chosen to be the number of features in each tone, 4.
  • 在预训练时,对采样的序列长度做了管理,从小序列开始逐渐加大,最后变成长序列
  • 采用了[4]中的freezen的trick,当D或G被训练得异常强大以至于对方梯度消失,无法正常进行训练时,对过于强大的一方实施冻结。这里采用的是A‘s training loss is less than 70% of the training loss of B时,冻结A
  • 采用了[4]中的feature matching的trick,将G的目标函数替换为使真假样本的feature差值最小化:

  其中,R是D的最后一层(激活函数logistic之前)输出。

评估标准

Polyphony 复音是否在同一时间点开始

Scale consistency were computed by counting the fraction of tones that were part of a standard scale, and reporting the number for the best matching such scale.(标准音程是什么鬼?)

Repetitions 小节重复数量

Tone span 最高音和最低音的音程统计

评估工具代码也放在github上面了

结论

第一例通过GAN对抗训练来生成音乐的paper。从人耳听觉的感受上来说,c-RNN-GAN生成的音乐完全不能和真实样本相提并论,应该是单纯地进行对抗训练,单轨音调,缺乏先验乐理知识的融入的缘故导致。

sample 试听:http://mogren.one/publications/2016/c-rnn-gan/

[1]Douglas Eck and Juergen Schmidhuber. Finding temporal structure in music: Blues improvisation
with lstm recurrent networks. In Neural Networks for Signal Processing, 2002. Proceedings of the
2002 12th IEEE Workshop on, pages 747–756. IEEE, 2002.

[2]Pascal Vincent Nicolas Boulanger-Lewandowski, Yoshua Bengio. Modeling temporal dependencies
in high-dimensional sequences: Application to polyphonic music generation and transcription. In
Proceedings of the 29th International Conference on Machine Learning (ICML), page 1159–1166,
2012.

[3]Lantao Yu, Weinan Zhang, Jun Wang, and Yong Yu. Seqgan: Sequence generative adversarial nets
with policy gradient. arXiv preprint arXiv:1609.05473, 2016.

[4]Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen.
Improved techniques for training gans. In Advances in Neural Information Processing Systems,
pages 2226–2234, 2016.

代码分析

Restore保存的参数:

'num_layers_g' : RNN cell g的层数

'num_layers_d' :RNN Cell D的层数

'meta_layer_size':

'hidden_size_g':

'hidden_size_d':

'biscale_slow_layer_ticks':

'multiscale':

'disable_feed_previous':

'pace_events':

'minibatch_d':

'unidirectional_d':

'feature_matching':

'composer':选取训练集中哪个作曲家的风格来进行训练,如巴赫 贝多芬......

do-not-redownload.txt存在,则不再下载新的midi文件

read_data函数读出的格式为[genre, composer, song_data]

这里组织了一个sources列表,键值为风格,艺术家

用python-midi读出midi_pattern后,遍历每一个track的每一个event,通过NoteOnEvent和NoteOffEvent记录每一个note的四个维度数值:

TICKS_FROM_PREV_START = 0
LENGTH = 1
FREQ = 2
VELOCITY = 3

最后,一首歌的所有的note被汇总到一个song_data的list中去了。每一个[genre, composer, song_data]代表一首歌的特征数据,这些数据被append到 loader.songs['validation'], loader.songs['test'] ,loader.songs['train']中去了。

创建模型训练时使用了l2正则项来避免过拟合:scope.set_regularizer(tf.contrib.layers.l2_regularizer(scale=FLAGS.reg_scale))

创建G,一个多层的LSTM:

输入噪声random_rnninputs的shape为[batch_size, songlength, int(FLAGS.random_input_scale*num_song_features)],然后转换为list

 
 

---恢复内容结束---

出处:arXiv: Artificial Intelligence, 2016(一年了还没中吗?)

Motivation

使用GAN+RNN来处理continuous sequential data,并训练生成古典音乐

Introduction

In this work, we investigate the feasibility of using adversarial training for a sequential model with continuous data, and evaluate it using classical music in freely available midi files.也就是利用GAN+RNN来处理midi file中的连续数据。RNN主要工作用于处理时序相关的自然语言,同时也被引入到了音乐生成的领域[1,2,3],but to our knowledge they always use a symbolic representation. In contrast,our work demonstrates how one can train a highly flexible and expressive model with fully continuous sequence data for tone lengths, frequencies, intensities, and timing.作者还刻意提到了LapGAN实现coarse-to-fine的图片生成过程(个人思考:对音乐生成很有启发,包括利用双层GAN来从caption生成image,一层用于生成低分辨率的粗线条色彩图片,一层用于生成细节,这些思路应该可以结合到音乐生成中去)。

Model

对抗网络中的G和D都是RNN模型,损失函数定义为

The input to each cell in G is a random vector, concatenated with the output of previous cell.D采用的是双向循环RNN(LSTM)。数据方面构建了一个tone length, frequency, intensity, and time的四元数组,数据可以表示出复调和弦polyphonous chords。

G和D的LSTM层数皆设置为2,BaseLine为去掉对抗性的单一的RNN生成网络。训练集Dataset是从网上down下来的标准midi格式的古典音乐文件,对所有的”note on“事件进行了记录的读取(包括该note的其他属性,时延,tone,强度等等),代码地址:https://github.com/olofmogren/c-rnn-gan

Training过程中使用了很多小技巧:

  • 使用L2 regularization对G和D的权重做正则化约束
  • The model was pretrained for 6 epochs with a squared error loss for predicting the next event in the
    training sequence
  • the input to each LSTM cell is a random vector v, concatenated with the output at previous time step. v is uniformly distributed in [0; 1]k, and k
    was chosen to be the number of features in each tone, 4.
  • 在预训练时,对采样的序列长度做了管理,从小序列开始逐渐加大,最后变成长序列
  • 采用了[4]中的freezen的trick,当D或G被训练得异常强大以至于对方梯度消失,无法正常进行训练时,对过于强大的一方实施冻结。这里采用的是A‘s training loss is less than 70% of the training loss of B时,冻结A
  • 采用了[4]中的feature matching的trick,将G的目标函数替换为使真假样本的feature差值最小化:

  其中,R是D的最后一层(激活函数logistic之前)输出。

评估标准

Polyphony 复音是否在同一时间点开始

Scale consistency were computed by counting the fraction of tones that were part of a standard scale, and reporting the number for the best matching such scale.(标准音程是什么鬼?)

Repetitions 小节重复数量

Tone span 最高音和最低音的音程统计

评估工具代码也放在github上面了

结论

第一例通过GAN对抗训练来生成音乐的paper。从人耳听觉的感受上来说,c-RNN-GAN生成的音乐完全不能和真实样本相提并论,应该是单纯地进行对抗训练,单轨音调,缺乏先验乐理知识的融入的缘故导致。

sample 试听:http://mogren.one/publications/2016/c-rnn-gan/

[1]Douglas Eck and Juergen Schmidhuber. Finding temporal structure in music: Blues improvisation
with lstm recurrent networks. In Neural Networks for Signal Processing, 2002. Proceedings of the
2002 12th IEEE Workshop on, pages 747–756. IEEE, 2002.

[2]Pascal Vincent Nicolas Boulanger-Lewandowski, Yoshua Bengio. Modeling temporal dependencies
in high-dimensional sequences: Application to polyphonic music generation and transcription. In
Proceedings of the 29th International Conference on Machine Learning (ICML), page 1159–1166,
2012.

[3]Lantao Yu, Weinan Zhang, Jun Wang, and Yong Yu. Seqgan: Sequence generative adversarial nets
with policy gradient. arXiv preprint arXiv:1609.05473, 2016.

[4]Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen.
Improved techniques for training gans. In Advances in Neural Information Processing Systems,
pages 2226–2234, 2016.

代码分析

Restore保存的参数:

'num_layers_g' : RNN cell g的层数

'num_layers_d' :RNN Cell D的层数

'meta_layer_size':

'hidden_size_g':

'hidden_size_d':

'biscale_slow_layer_ticks':

'multiscale':

'disable_feed_previous':

'pace_events':

'minibatch_d':

'unidirectional_d':

'feature_matching':

'composer':选取训练集中哪个作曲家的风格来进行训练,如巴赫 贝多芬......

do-not-redownload.txt存在,则不再下载新的midi文件

read_data函数读出的格式为[genre, composer, song_data]

这里组织了一个sources列表,键值为风格,艺术家

用python-midi读出midi_pattern后,遍历每一个track的每一个event,通过NoteOnEvent和NoteOffEvent记录每一个note的四个维度数值:

TICKS_FROM_PREV_START = 0
LENGTH = 1
FREQ = 2
VELOCITY = 3

最后,一首歌的所有的note被汇总到一个song_data的list中去了。每一个[genre, composer, song_data]代表一首歌的特征数据,这些数据被append到 loader.songs['validation'] loader.songs['test'] loader.songs['train']中去了。

对于待训练的placeholder数据有:

self._input_songdata = tf.placeholder(shape=[batch_size, songlength, num_song_features], dtype=data_type())
self._input_metadata = tf.placeholder(shape=[batch_size, num_meta_features], dtype=data_type())
 
songdata_inputs将_input_songdata转成songlength个tensor的list,shape为[batch_size,num_song_features](这里用unstack要方便点吧,待测试):
songdata_inputs = [tf.squeeze(input_, [1])
for input_ in tf.split(self._input_songdata, songlength, 1)]
 

创建模型训练时使用了l2正则项来避免过拟合:scope.set_regularizer(tf.contrib.layers.l2_regularizer(scale=FLAGS.reg_scale))

创建G的LSTM网络:

输入噪声random_rnninputs的shape为[batch_size, songlength, int(FLAGS.random_input_scale*num_song_features)],然后转换为list(unstack?)

对G进行RNN的分步训练过程,每个循环是一步,输入为噪音random_rnninput和上一步的输出generated_point(两者concat为一个[batch_size,2*num_song_features]的tensor,第一步输出的初始化从均匀分布中采样)

对G还有个pretraining的过程,输入为噪音random_rnninputs和真实的sample songdata_input[i]

针对G的pretraining的loss是L2距离,注意这里的链表stack和[1,0,2]转置:

self.rnn_pretraining_loss = tf.reduce_mean(tf.squared_difference(x=tf.transpose(tf.stack(self._generated_features_pretraining), perm=[1, 0, 2]), y=self._input_songdata))

并加上一个正则项防止过拟合:

self.rnn_pretraining_loss = self.rnn_pretraining_loss+reg_loss
 
D采用了多(双)层双向LSTM,由于版本问题,我改写了一个多层lstm的接口:

要注意的是(1)由于bidirectional_dynamic_rnn每构建一次就会自动在名字空间中序号+1,所以用层数名来限定了scope(折腾了一天,是我菜还是tf太坑?)

(2)每次的输入_inputs需要把output中包含了bw和fw的tuple元组concat起来,每个tensor的shape为[batch_size,song_length,ouput_dim],其中output_dim和lstm隐层单元数量(状态数量)

一致,合并后shape为[batch_size,song_length,2×ouput_dim]

随后D将双向LSTM的输出全连接(output num = 1)并sigmoid映射为真假概率,同时输出output作为features,参与到feature loss的计算中去。

loss计算:

 
 
 

《C-RNN-GAN: Continuous recurrent neural networks with adversarial training》论文笔记的更多相关文章

  1. 《Vision Permutator: A Permutable MLP-Like ArchItecture For Visual Recognition》论文笔记

    论文题目:<Vision Permutator: A Permutable MLP-Like ArchItecture For Visual Recognition> 论文作者:Qibin ...

  2. [place recognition]NetVLAD: CNN architecture for weakly supervised place recognition 论文翻译及解析(转)

    https://blog.csdn.net/qq_32417287/article/details/80102466 abstract introduction method overview Dee ...

  3. 论文笔记系列-Auto-DeepLab:Hierarchical Neural Architecture Search for Semantic Image Segmentation

    Pytorch实现代码:https://github.com/MenghaoGuo/AutoDeeplab 创新点 cell-level and network-level search 以往的NAS ...

  4. 论文笔记——Rethinking the Inception Architecture for Computer Vision

    1. 论文思想 factorized convolutions and aggressive regularization. 本文给出了一些网络设计的技巧. 2. 结果 用5G的计算量和25M的参数. ...

  5. 论文笔记:Fast Neural Architecture Search of Compact Semantic Segmentation Models via Auxiliary Cells

    Fast Neural Architecture Search of Compact Semantic Segmentation Models via Auxiliary Cells 2019-04- ...

  6. 论文笔记:ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware

    ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware 2019-03-19 16:13:18 Pape ...

  7. 论文笔记:DARTS: Differentiable Architecture Search

    DARTS: Differentiable Architecture Search 2019-03-19 10:04:26accepted by ICLR 2019 Paper:https://arx ...

  8. 论文笔记:Progressive Neural Architecture Search

    Progressive Neural Architecture Search 2019-03-18 20:28:13 Paper:http://openaccess.thecvf.com/conten ...

  9. 论文笔记:Auto-DeepLab: Hierarchical Neural Architecture Search for Semantic Image Segmentation

    Auto-DeepLab: Hierarchical Neural Architecture Search for Semantic Image Segmentation2019-03-18 14:4 ...

  10. 论文笔记系列-DARTS: Differentiable Architecture Search

    Summary 我的理解就是原本节点和节点之间操作是离散的,因为就是从若干个操作中选择某一个,而作者试图使用softmax和relaxation(松弛化)将操作连续化,所以模型结构搜索的任务就转变成了 ...

随机推荐

  1. msp430入门编程45

    msp430中C语言的人机交互--独占CPU菜单

  2. iOS 混合变换旋转 CGAffineTransform

    在ios 中, Core Graphics 提供了一系列的函数可以在一个变换的基础上做深层次的变换,如果做一个既要缩放又要旋转的变换,以下的方法比较实用. CGAffineTransformScale ...

  3. HDU 2255 二分图最佳匹配

    奔小康赚大钱 Time Limit: 1000/1000 MS (Java/Others)    Memory Limit: 32768/32768 K (Java/Others)Total Subm ...

  4. UESTC 1087 【二分查找】

    问了某明==shit 中文题意不解释. 因为数据的范围是1e9以内的所以我们可以通过二分的方法枚举可能的中位数. 用二分法每次判断某一数字比中位数大还是比中位数小. 判断方法是,枚举以第n个数为起点的 ...

  5. 2017多校Round10(hdu6171~hdu6181)

    补题进度:5/11 1001(双向BFS) 题意: 给你一个类似移子游戏,给你初始状态和终止状态,问初始状态到终止状态至少要移多少步,如果步数>20就-1 分析: 很明显的BFS了,不过普通的B ...

  6. java基础语法1

    一:基础语法之--标识符,修饰符,关键字 1.标识符: 定义:类名.变量名以及方法名都被称为标识符.自定义的名字. 注意: ·所有的标识符都应该以字母(A-Z或者a-z),美元符($).或者下划线(_ ...

  7. Building clang on RedHat

    http://btorpey.github.io/blog/2015/01/02/building-clang/ clang is a great compiler, with a boatload ...

  8. 【Todo】STAR面试法

    今天在面试培训的时候,接触到了STAR面试法. 觉得挺好的,用来准备非技术面试,还蛮全面的.所以可以多了解一下. 可以参考:http://www.hrloo.com/rz/73652.html

  9. 【网络】TCP的流量控制

    一.利用滑动窗口实现流量控制 流量控制是让发送方的发生速率不要太快,要让接收方来得及接收. 发送方的发送窗口不能超过接收方给出的接收窗口的数值,TCP的窗口单位是字节,不是报文段. TCP为每一个连接 ...

  10. History(历史)命令用法 15 例

    如果你经常使用 Linux 命令行,那么使用 history(历史)命令可以有效地提升你的效率.本文将通过实例的方式向你介绍 history 命令的 15 个用法. 使用 HISTTIMEFORMAT ...