In torch, in order to follow closely the Python implementation we implemented iterators as R6 classes with .iter() and .next() methods.

A minimal example would be something like this:

library(coro)
Range <- R6::R6Class(
  classname = "Range",
  public = list(
    n = NULL,
    i = 0,
    initialize = function(n) {
      self$n <- n
    },
    .next = function() {
      self$i <- self$i + 1
      if (self$i <= self$n)
        return(self$i)
      coro::exhausted()
    }
  )
)

as_iterator.Range <- function(x) {
  force(x)
  coro::as_iterator(function() {
    x$.next()
  })
}

coro::iterate(for (i in as_iterator.Range(Range$new(3))) {
  print(i)
})

This works nice, but it would be nice to avoid the explicit as_iterator.Range call:

coro::iterate(for (i in Range$new(3)) {
  print(i)
})

The $.iter() method can be used to decouple different strategies of iteration or simply to avoid the need of reinitializing the class. In torch, the context is something like this: The dl instance can be used for every epoch without the need to recreate the object. This would be possible by abusing as_iterator and implementing a method for Dataloader instances that returns a fresh iterator each time.

dl <- Dataloader$new(dataset)

for (i in 1:epochs) {
  iterate(for(batch in dl) {
    # train model
  })
}

Here's torch's code: https://github.com/mlverse/torch/blob/master/R/utils-data-dataloader.R#L84