WaveNet in R

 

Wavenet은 2016년에 DeepMind에서 만들어낸 모형으로, 음성의 파형을 모형으로 사용하여 실제 사람과 같은 음성을 생성하는 알고리즘이다.

 

en.wikipedia.org/wiki/WaveNet

WaveNet - Wikipedia

 

시계열에 활용될 수 있다고 하여 사용해 보았다.

 

아래는 dilation_rate만을 조절한 코드이다. 

 

 

library('reticulate')
library('tensorflow')
library('keras')
library(abind)
data('AirPassengers')
ap_ts <- window(AirPassengers, start=1949, end=1955.99)
ap_ts_test <- window(AirPassengers, start=1956)

data_generate=function(data){
  tr=as.vector(data)
  tr_xs=list();tr_ys=list()
  for(i in 1:(length(tr)-36)){
    tr_xs[[i]]=tr[i:(i+23)]
    tr_ys[[i]]=tr[(i+13):(i+36)]
  }
  tr_x=t(data.frame(tr_xs))
  tr_y=t(data.frame(tr_ys))
  rownames(tr_x)=NULL
  rownames(tr_y)=NULL
  tr_x=array(tr_x,c(nrow(tr_x),ncol(tr_x),1))
  tr_y=array(tr_y,c(nrow(tr_y),ncol(tr_y),1))
  return(list(tr_x,tr_y))
}

tr_=data_generate(ap_ts)
tr_x=tr_[[1]];tr_y=tr_[[2]]
dim(tr_x);dim(tr_y)
te_=data_generate(ap_ts_test)
te_x=te_[[1]];te_y=te_[[2]]

model=keras_model_sequential()
model$add(layer_input(shape=c(24,1)))
for(rate in rep(c(1,2,4,8),2)){
  model$add(layer_conv_1d(filters=20,kernel_size = 2,padding='causal',
                                 activation='relu',dilation_rate = rate))
}
model$add(layer_conv_1d(filters=1, kernel_size=1))
model%>%compile(
  loss='mae',
  optimizer='adam',
  metrics=c('mae')
)

#kerasR::plot_model(model,to_file='/home/ducj/jupyter/R/tensorflow/model.png',show_shapes = T)
dim(tr_x);dim(tr_y[,,1])
model%>%fit(x = tr_x,y=tr_y,batch_size = 3,epochs = 600)
n=5
dim(te_y)
ylim_=c(min(c((model%>%predict(te_x))[n,,1]),te_y[n,,1]),max(c((model%>%predict(te_x))[n,,1]),te_y[n,,1]))
plot((model%>%predict(te_x))[n,,1],type='l',ylim=ylim_);lines(te_y[n,,1],col=2)

 친절히 누군가가 패키지를 만들어주셔서 활용해보았다. 

library('reticulate')
library('tensorflow')
library('keras')
library(abind)
data('AirPassengers')
ap_ts <- window(AirPassengers, start=1949, end=1955.99)
ap_ts_test <- window(AirPassengers, start=1956)

data_generate=function(data){
  tr=as.vector(data)
  tr_xs=list();tr_ys=list()
  for(i in 1:(length(tr)-36)){
    tr_xs[[i]]=tr[i:(i+23)]
    tr_ys[[i]]=tr[(i+13):(i+36)]
  }
  tr_x=t(data.frame(tr_xs))
  tr_y=t(data.frame(tr_ys))
  rownames(tr_x)=NULL
  rownames(tr_y)=NULL
  tr_x=array(tr_x,c(nrow(tr_x),ncol(tr_x),1))
  tr_y=array(tr_y,c(nrow(tr_y),ncol(tr_y),1))
  return(list(tr_x,tr_y))
}

tr_=data_generate(ap_ts)
tr_x=tr_[[1]];tr_y=tr_[[2]]
dim(tr_x);dim(tr_y)
te_=data_generate(ap_ts_test)
te_x=te_[[1]];te_y=te_[[2]]

model=keras_model_sequential()
model$add(layer_input(shape=c(24,1)))
for(rate in rep(c(1,2,4,8),2)){
  model$add(layer_conv_1d(filters=20,kernel_size = 2,padding='causal',
                                 activation='relu',dilation_rate = rate))
}
model$add(layer_conv_1d(filters=1, kernel_size=1))
model%>%compile(
  loss='mae',
  optimizer='adam',
  metrics=c('mae')
)

# remotes::install_github("r-tensorflow/wavenet")
library(wavenet)

model=wavenet(filters=64,kernel_size=3,
        residual_blocks = 2^rep(1:8, 3), input_shape = list(24, 1),
        output_channels = 1,output_activation = 'linear')

model %>%
  keras::compile(
    loss = 'mae',
    optimizer = keras::optimizer_adam()
  )

dim(tr_x);dim(tr_y)
model%>%fit(x = tr_x,y=tr_y,batch_size = 10,epochs = 1500)
n=9
dim(te_y)
ylim_=c(min(c((model%>%predict(te_x))[n,,1]),te_y[n,,1]),max(c((model%>%predict(te_x))[n,,1]),te_y[n,,1]))
plot((model%>%predict(te_x))[n,,1],type='l',ylim=ylim_);lines(te_y[n,,1],col=2)

 

예제 자료는 자료가 충분하지 않아서 성능이 좋지 못한 것 같다.