0

I'm struggling to make the gradient descent function I already have into one for stochastic gradient descent. I have the following:

  gd <- function(f, grad, y, X, theta0, npars, ndata, a, niters) {
  theta <- matrix(data=NA, nrow=niters, ncol=npars)
  cost <- vector(mode="numeric", length=niters)

  theta[1, ] <- theta0
  cost[1] <- f(y, X, theta0, ndata)

  for (i in 2:niters) {
    theta[i, ] <- theta[i-1, ]-a*grad(y, X, theta[i-1, ], ndata)
    cost[i] <- f(y, X, theta[i, ], ndata)
  }

  return(list(theta=theta, cost=cost))
}

This code works fine. I'm trying to do it so instead of ndata <- 1000, I have 100 points randomly sampled. I tried changing the second part to 

for (i in 2:niters) {
    samp <- sample(ndata, nsubsamples)
    theta[i, ] <- theta[i-1, ]-a*grad(y[samp,], X[samp,], theta[i-1, ], nsubsamples) 
    cost[i] <- f(y, X, theta[i, ], nsubsamples)

but i get an error saying:

Error in y[samp, ] : incorrect number of dimensions.

My y is a column from a dataset called simulated_data with 1000 observations. When trying to get 100 random samples from it (nsubsamples=100 and ndata=1000, simulated_data[samp,]$y works but simulated_data$y[samp,] does not. But my y has to be defined as simulated_data$y.

So I'm wondering if there's an easier way to add a random sample and, if I've done that, the rest of my code is correct (as I'm a bit confused if I should be using ndata or nsubsamples for theta[i] and cost[i].

jogo
  • 12,469
  • 11
  • 37
  • 42
Mark Durkan
  • 3
  • 2
  • Do you want `sample(1:ndata, nsubsamples)` ? – jogo Feb 04 '19 at 19:50
  • Hi, I still have the same problem with that. From my code, it gets 100 samples from 1:1000 fine. If I then use simulated_data[samp,] (simulated_data is a df with 2 columns and 1000 obersvations) that then works at picking out 100 random observations. It just doesn't work when I specify the column first. – Mark Durkan Feb 04 '19 at 19:56
  • If you specify the column you have to delete the comma: `simulated_data$y[samp]` ... because `simulated_data$y` is a **one**dimensional object for which indexing can not take two dimensions `[..., ...]` separated by `,` – jogo Feb 04 '19 at 20:06

0 Answers0