gradient_desc_anim.R

library(tidyverse)
library(glue)
library(gganimate)


grid <- seq(-10,5,length.out=1e4)
y <- (grid + 2)^2 + 3

plot_data <- tibble(grid, y)

#######


f <- function(x){(x+2)^2 + 3}
df_dx <- function(x){2*x+4}

learning_rate <- 0.1
iter <- 1
x <- 5

xs <- numeric()
ys <- numeric()
deltas <- numeric()

while(TRUE){
  xs[[iter]] <- x
  ys[[iter]] <- f(x)
  
  current_grad <- df_dx(x)
  delta <- -current_grad*learning_rate
  deltas[[iter]] <- delta
  
  x <- x + delta
  
  print(glue("Iteration: {iter}"))
  print(glue("x: {x}"))
  print(glue("y: {f(x)}"))
  print(glue("delta: {delta}"))
  
  iter <- iter + 1
  
  if (abs(delta)<0.001 | iter>5000) {
    break
  }
}

df <- tibble(x = xs, y=ys, delta = deltas, iter=1:length(xs))

print(nrow(df))

a <- df %>%
  ggplot() +
  aes(x=x, y=y) +
  geom_point(size=2, color="#8f2726") +
  geom_line(data=plot_data, aes(x=grid, y=y)) +
  cowplot::theme_cowplot(font_size=14, font_family = "Lato") +
  labs(title = glue("Learning rate: {learning_rate}"),
       subtitle = "Iteration: {closest_state} \n x: {df$x[as.integer(closest_state)]} \n y: {df$y[as.integer(closest_state)]} \n Delta: {df$delta[as.integer(closest_state)]}") +
  transition_states(iter, wrap=FALSE) +
  shadow_mark(alpha=0.2, color="black", size=1) +
  ease_aes('cubic-in-out')


a_rendered <- animate(a,
        fps = 20,              
        nframes=nrow(df)*2,
        detail=10,
        res=300,
        height=5,
        width=5,
        unit="in",
        renderer = ffmpeg_renderer())

anim_save(animation = a_rendered, filename = "quadratic_GD1.mp4")