Why RNN always outputs 1

Question

I use Recurrent Neural Networks (RNN) for forecasting, but for some strange reason, it always outputs 1. Here I explain this using a toy example as:

Example Consider a Msize matrix (360, 5) and a vector Ythat contains rowsum M. Now, using RNN, I want to predict Yfrom M. Using the package rnn R, I trained the model how

   library(rnn) 
    M <- matrix(c(1:1800),ncol=5,byrow = TRUE) # Matrix (say features) 
    Y <- apply(M,1,sum) # Output equls to row sum of M
    mt <- array(c(M),dim=c(NROW(M),1,NCOL(M))) # matrix formatting as [samples, timesteps, features]
    yt <- array(c(Y),dim=c(NROW(M),1,NCOL(Y))) # formatting
    model <- trainr(X=mt,Y=yt,learningrate=0.5,hidden_dim=10,numepochs=1000) # training

One strange thing that I observed during training is that the error of an epoch is always equal to 4501. Ideally, the error of an epoch should decrease with increasing eras.

Then I created a test data set with the same structure as above:

M2 <- matrix(c(1:15),nrow=3,byrow = TRUE)
mt2 <- array(c(M2),dim=c(NROW(M2),1,NCOL(M2)))
predictr(model,mt2)

With a prediction, I always get a conclusion like 1. What could be the reason for the constant error of the era and the same exit?

# 1

, @Barker, . , Dropbox traindata, testadata R .

: "" - , , , 14.

normalize_data <- function(x){
  normalized = (x-min(x))/(max(x)-min(x))
  return(normalized)
}

#read test and train data
traindat <- read.csv(file = "train.csv")
testdat <- read.csv(file = "test.csv")
# column "power" is response variable and remaining are predictors
# predictors in  traindata
trainX <- traindat[,1:dim(traindat)[2]-1]
# response of train data
trainY <- traindat$power
# arrange data acc. to RNN as [samples,time steps, features]
tx <- array(as.matrix(trainX), dim=c(NROW(trainX), 1, NCOL(trainX)))
tx <- normalize_data(tx) # normalize data in range of [0,1]
ty <- array(trainY, dim=c(NROW(trainY), 1, NCOL(trainY))) # arrange response acc. to predictors
# train model
model <- trainr(X = tx, Y = ty, learningrate = 0.08, hidden_dim = 6, numepochs = 400)

# predictors in test data
testX <- testdat[,1:dim(testdat)[2]-1]
testX <- normalize_data(testX) # normalize data in range of [0,1]
#testY <- testdat$power
# arrange data acc. to RNN as [samples,time steps, features]
tx2 <- array(as.matrix(testX), dim=c(NROW(testX), 1, NCOL(testX))) # predict
pred <- predictr(model,tx2)
pred

learning rate, hidden_dim, numepochs, 0,9, 1.

Answer 1

RNN , . - . , , R co2. , .

"" "" co2, stl. , , ( , numepochs ). , :

#Create the STL decomposition
sdcomp <- stl(co2, s.window = 7)$time.series[,1:2]

Y <- window(co2, end = c(1996, 6))
M <- window(sdcomp, end = c(1996, 6))
#Taken from OP code
mt <- array(c(M),dim=c(NROW(M),1,NCOL(M)))
yt <- array(c(Y),dim=c(NROW(M),1,NCOL(Y))) 
model <- trainr(X=mt,Y=yt,learningrate=0.5,hidden_dim=10,numepochs=100)

:

M2 <- window(sdcomp, start = c(1996,7))
mt2 <- array(c(M2),dim=c(NROW(M2),1,NCOL(M2)))
predictr(model,mt2)

output:
      [,1]
 [1,]    1
 [2,]    1
 [3,]    1
 [4,]    1
 [5,]    1
 [6,]    1
 [7,]    1
 [8,]    1
 [9,]    1
[10,]    1
[11,]    1
[12,]    1
[13,]    1
[14,]    1
[15,]    1
[16,]    1
[17,]    1
[18,]    1

Ewe, , . , . , 18 , , 18 .

dco2 <- diff(co2, 18)
sdcomp <- stl(dco2, s.window = "periodic")$time.series[,1:2]
plot(dco2)

, , . .

Y <- window(dco2, end = c(1996, 6))
M <- window(sdcomp, end = c(1996, 6))

mt <- array(c(M),dim=c(NROW(M),1,NCOL(M)))
yt <- array(c(Y),dim=c(NROW(M),1,NCOL(Y)))
model <- trainr(X=mt,Y=yt,learningrate=0.5,hidden_dim=10,numepochs=100)

M2 <- window(sdcomp, start = c(1996,7))
mt2 <- array(c(M2),dim=c(NROW(M2),1,NCOL(M2)))
(preds <- predictr(model,mt2))

output:
              [,1]
 [1,] 9.999408e-01
 [2,] 9.478496e-01
 [3,] 6.101828e-08
 [4,] 2.615463e-08
 [5,] 3.144719e-08
 [6,] 1.668084e-06
 [7,] 9.972314e-01
 [8,] 9.999901e-01
 [9,] 9.999916e-01
[10,] 9.999916e-01
[11,] 9.999916e-01
[12,] 9.999915e-01
[13,] 9.999646e-01
[14,] 1.299846e-02
[15,] 3.114577e-08
[16,] 2.432247e-08
[17,] 2.586075e-08
[18,] 1.101596e-07

, -! , , , dco2:

, , "" . , , , . , , 18 .

Answer 2

(. ?trainr) . int2bin bin2int.

, , , .