[tensorflow]Seq2Seq regression in R

https://diane-space.tistory.com/318?category=882603 

[시계열] 케라스에서 Luong 어텐션을 활용한 seq2seq2 LSTM 모델 만들기 (번역)

회귀 모델로 Seq2Seq 모델을 활용한 연구 사례가 있어 적용해 보려했으나,

 

난이도가 너무 높아서 다소 힘들었다.

 

그래서 활용 예제를 만들어 공유해보고자 한다. 

 

만약 틀린 부분이 있으면 얘기해주기 바란다.

 

활용 패키지는 아래와 같으며 시계열 예측할 때 자주 나오는 AirPassengers 자료를 분해시계열을 적용하여

 

활용하기로 했다. 물론 관측 자료수가 적어서 성능은 안좋게 보일 것 같다.

 

library(forecast)
library(abind)
library(keras)
library(tensorflow)
library(kerasR)
library(scales)
data=decompose(AirPassengers)
data=na.omit(data.frame(x=data$x,seasonal=data$seasonal,trand=data$trend,random=data$random))

 

데이터 부분은 아래와 같이 truncate를 거쳐 증가시켰다.

truncate=function(x,train_len=24,test_len=10,cols){
  in_=list();out_=list();label_=list()
  for(i in 1:(dim(x)[1]-train_len-test_len)){
    in_[[i]]=x[i:(i+train_len),cols]
    out_[[i]]=x[(i+train_len):(i+train_len+test_len),cols]
  }
  return(list(X_in=abind(in_,along=0),X_out=abind(out_,along=0)))
}
train=data[1:80,]
test=data[81:nrow(data),]
means=apply(train,2,mean)
sds=apply(train,2,sd)
train=(train-means)/sds
test=(test-means)/sds

train=truncate(train,train_len=24,test_len=10,cols=c('seasonal','trand','random'))
test=truncate(test,train_len=24,test_len=10,cols=c('seasonal','trand','random'))

 

n_hidden=100
input_train =keras::layer_input(shape=c(dim(train$X_in)[2],dim(train$X_in)[3]))
output_train=keras::layer_input(shape=c(dim(train$X_out)[2],dim(train$X_out)[3]))
encoder=keras::layer_lstm(input_train,units=n_hidden,activation = 'swish',
                          dropout = .2,recurrent_dropout = .2,
                          return_sequences = F,return_state = T)
encoder_last_h=keras::layer_batch_normalization(encoder[[1]],momentum = .6)
encoder_last_c=keras::layer_batch_normalization(encoder[[3]],momentum = .6)
decoder=keras::layer_repeat_vector(encoder[[1]],dim(train$X_out)[2])
decoder_lstm=keras::layer_lstm(units=n_hidden,activation='swish',
                          dropout = .2,recurrent_dropout = .2,
                          return_sequences = T,return_state = F)
decoder=decoder_lstm(decoder,initial_state=list(encoder_last_h,encoder_last_c))
out=keras::time_distributed(decoder,keras::layer_dense(units=output_train$shape[[3]]))

model=keras::keras_model(inputs=input_train,outputs=out)
opt=keras::optimizer_adam(lr=0.01,clipnorm=1)
model$compile(loss='MSE',optimizer=opt,metrics=c('mae'))
model$summary()
# kerasR::plot_model(model,to_file = 'model_plot.png',show_shapes=T,show_layer_names=T)
dim(train$X_in)
dim(train$X_out)

es=keras::callback_early_stopping(monitor='val_loss',mode='min',patience = 3)
history = model$fit(train$X_in,train$X_out,validation_split=.2,epochs=500L,
                   verbose=1,callbacks=es,batch_size=30L)
pred=model$predict(test$X_in)
for(i in 1:dim(pred)[3]){
  pred[,,i]=pred[,,i]*sds[i]+means[i]
}
plot(data$x[1:(81+11)],ylim=c(0,500))
points(81:(81+10),apply(pred,c(1,2),sum)[1,],col=2)

 

 

아래는 단변량 코드의 결과인데 다소 아쉬운 모습을 보여준다.

train=data[1:80,]
test=data[81:nrow(data),]
means=apply(train,2,mean)
sds=apply(train,2,sd)
train=(train-means)/sds
test=(test-means)/sds
train=truncate(train,train_len=24,test_len=10,cols=c('x'))
test=truncate(test,train_len=24,test_len=10,cols=c('x'))
train$X_in=array(train$X_in,dim=c(nrow(train$X_in),ncol(train$X_in),1))
train$X_out=array(train$X_out,dim=c(nrow(train$X_out),ncol(train$X_out),1))
test$X_in=array(test$X_in,dim=c(nrow(test$X_in),ncol(test$X_in),1))
test$X_out=array(test$X_out,dim=c(nrow(test$X_out),ncol(test$X_out),1))

n_hidden=100
input_train =keras::layer_input(shape=c(dim(train$X_in)[2],dim(train$X_in)[3]))
output_train=keras::layer_input(shape=c(dim(train$X_out)[2],dim(train$X_out)[3]))
encoder=keras::layer_lstm(input_train,units=n_hidden,activation = 'swish',
                          dropout = .2,recurrent_dropout = .2,
                          return_sequences = F,return_state = T)
encoder_last_h=keras::layer_batch_normalization(encoder[[1]],momentum = .6)
encoder_last_c=keras::layer_batch_normalization(encoder[[3]],momentum = .6)
decoder=keras::layer_repeat_vector(encoder[[1]],dim(train$X_out)[2])
decoder_lstm=keras::layer_lstm(units=n_hidden,activation='swish',
                               dropout = .2,recurrent_dropout = .2,
                               return_sequences = T,return_state = F)
decoder=decoder_lstm(decoder,initial_state=list(encoder_last_h,encoder_last_c))
out=keras::time_distributed(decoder,keras::layer_dense(units=output_train$shape[[3]]))

model=keras::keras_model(inputs=input_train,outputs=out)
opt=keras::optimizer_adam(lr=0.01,clipnorm=1)
model$compile(loss='MSE',optimizer=opt,metrics=c('mae'))
model$summary()
# kerasR::plot_model(model,to_file = 'model_plot.png',show_shapes=T,show_layer_names=T)
dim(train$X_in)
dim(train$X_out)

es=keras::callback_early_stopping(monitor='val_loss',mode='min',patience = 50)
history = model$fit(train$X_in,train$X_out,validation_split=.2,epochs=500L,
                    verbose=1,callbacks=es,batch_size=30L)
pred=model$predict(test$X_in)
for(i in 1:dim(pred)[3]){
  pred[,,i]=pred[,,i]*sds[i]+means[i]
}
plot(data$x[1:(81+11)],ylim=c(0,500))
points(81:(81+10),apply(pred,c(1,2),sum)[1,],col=2)