torch for R examples

example_00.R

library(torch)
library(ggplot2)

# we want to find the minimum of this function
# using the gradient descent.

f <- function(x) {
  x^2 - x
}

p <- qplot(x = NULL, geom = "point") +
  stat_function(fun = f) +
  xlim(-2, 2)
p


# for the gradient descent to work we need to know how
# to compute derivatives, for example:

# f(x) = x^2 - x
df_dx <- function(x) {
  2*x - 1
}

x <- -2

learning_rate <- 0.1
segments <- tibble::tibble(x = c(), xend = c())
for (t in 1:15) {
  
  xstart <- x
  x <- x - learning_rate * df_dx(x)
  xend <- x
  
  segments <- segments %>% 
    tibble::add_row(x = xstart, xend = xend)
}
p +
  geom_segment(
    data = segments,
    aes(x = x, y = f(x), xend = xend, yend = f(xend)),
    arrow = arrow(length = unit(0.2, "cm")), 
    color = "red"
  )

example_01.R

library(torch)
library(ggplot2)

# we want to find the minimum of this function
# using the gradient descent.

f <- function(x) {
  x^2 - x
}

p <- qplot(x = NULL, geom = "point") +
  stat_function(fun = f) +
  xlim(-2, 2)
p

# for the gradient descent to work we need to know how
# to compute derivatives, for example:

# f(x) = x^2 - x
df_dx <- function(x) {
  2*x - 1
}

# in torch this done with:
x <- torch_tensor(-2, requires_grad = TRUE)
y <- f(x)
y$backward()
x$grad$zero_() #<- this gives us df/dx

# gradient descent
# x[t+1] = x[t] - learning_rate * df/dx[t]
learning_rate <- 0.1
segments <- tibble::tibble(x = c(), xend = c())
for (t in 1:15) {
  
  xstart <- as.numeric(x)
  
  x$grad$zero_()
  y <- f(x)
  y$backward()
  with_no_grad({
    x$sub_(learning_rate*x$grad)
  })
  
  xend <- as.numeric(x)
  
  segments <- segments %>% 
    tibble::add_row(x = xstart, xend = xend)
}

p +
  geom_segment(
    data = segments, 
    aes(x = x, y = f(x), xend = xend, yend = f(xend)),
    arrow = arrow(length = unit(0.2, "cm")), 
    color = "red"
  )