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:
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:
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:
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:
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: