Turning a simple machine learning model in R into an api just became a whole lot easier. Embarrisingly easy actually, thanks to a lovely package called plumber.

Install

Assuming that you have devtools installed, all you need to do is type the following:

library(devtools)
install_github("trestletech/plumber")
library(plumber)

With this installed, let’s create a file that creates an endpoint!

Example 1

# prediction_serv.R

library(magrittr)

mod_chick <- lm(data = ChickWeight, weight ~ Time + Diet)
mod_cars <- lm(data = cars, dist ~ speed)

#' @get /
hello_world <- function(){
  '1 got 99 problems and flask aint one'
}

#' @get /predict_chick
predict_chick <- function(time, diet){
  data.frame(Time = time %>% as.numeric, Diet = diet %>% as.factor) %>% 
    predict(mod_chick, newdata = .) %>% 
    as.numeric
}

#' @get /predict_cars
predict_cars <- function(speed){
  data.frame(speed = speed %>% as.numeric) %>% 
    predict(mod_cars, newdata = .) %>% 
    as.numeric
}

Note that the last line myfile.R needs to be a an empty line.

To get this script to act as a web endpoint you’ll need to run the plumber::plumb('/your/path/prediction_serv.R')$run(port=8000) from Rstudio or an R shell.

Once this is up, you can curl to the service.

$ curl --data "time=24&diet=3" http://localhost:8000/predict_chick
[257.4356]
$ curl --data "speed=30" http://localhost:8000/predict_cars
[100.3932]
$ curl "http://localhost:8000/"
curl http://localhost:8000/predict_chick

You can also access this endpoint via the browser, or indeed another R script (via rvest or httr).

$ R
> library(rvest)
> html('http://localhost:8000/predict_cars?speed=30') %>% 
+   html_nodes('p') %>% 
+   html_text
[1] "[100.3932]"
> library(httr)
>  get_speed_pred <- function(speed){
+   'http://localhost:8000/predict_cars?speed=' %>% 
+      paste0(speed) %>% 
+      GET(encode = 'json') %>% 
+      content
}
> c(10, 30, 50, 12) %>% lapply(get_speed_pred) %>% unlist
[1]  21.7450 100.3932 179.0413  29.6098

Example 2

The previous example only used GET requests. With plumbr you could post as well. Consider this second example.

# post_serv.R

library(ggplot2)
library(magrittr)

df <- data.frame(x = rnorm(10, 0, 1), y = rnorm(10, 0, 1))

#' @png
#' @get /plot
show_plot <- function(){
  p = ggplot() + 
    geom_point(data=df, aes(x,y), alpha = 0.5, size = 1.5) + 
    ggtitle("a plot of all the points")
  print(p)
}

#' @post /add_data
add_data <- function(n){
  df <<- data.frame(x = rnorm(n, 0, 1), 
                   y = rnorm(n, 0, 1)) %>% 
    rbind(df)
}

#' @get /all_data
all_data <- function(){
  df 
}

Again, note that the last line post_serv.R needs to be a an empty line.

You can run plumber::plumb('/your/path/post_serv.R') to bring the service online. This new service allows you to post data and can create a plot for viewing as well. A dataframe can be retreived as a json blob which can be used to give data to a dashboard (d3 would work very well here).

You can view this blob via:

curl http://localhost:8000/all_data

If you use jsonlite it is trivial to turn this json blob endpoint into a dataframe in R.

> library(jsonlite)
> http://localhost:8000/all_data' %>% GET %>% content('text') %>% fromJSON
         x       y
1  -1.4321 -1.1285
2  -1.7547  1.7694
3  -0.3472 -0.9206
4  -0.8752 -0.0267
5  -0.9626  1.2353
6   0.9005  0.3753
7   0.6310 -0.8690
8  -0.4543  0.4175
9  -0.7079 -0.6164
10  0.1954  0.1112

You'll want to use httr if you want to use not just GET but also POST requests as well. But just jsonlite will also work.

> http://localhost:8000/all_data' %>% fromJSON
         x       y
1  -1.4321 -1.1285
2  -1.7547  1.7694
3  -0.3472 -0.9206
4  -0.8752 -0.0267
5  -0.9626  1.2353
6   0.9005  0.3753
7   0.6310 -0.8690
8  -0.4543  0.4175
9  -0.7079 -0.6164
10  0.1954  0.1112

You can view all the points currently in the dataframe by pasting http://localhost:8000/plot in the browser.

If you add points, you can see an update in the plot by refreshing.

curl --data "n=100" http://localhost:8000/add_data

We can up the anty by adding even more points and refreshing.

curl --data "n=10000" http://localhost:8000/add_data

Conclustion

I'm very enthousiastic about this project.

Don't get me wrong, I love flask/python and this project won't work for a lot of api's, but it does feel liberating to be able to do this sort of thing in R as well. Note that performance might still be a thing as R is single threaded so be careful if you want to start pushing this to something thats meant to be very fast and responsive.

You should also be able to find this blog on: r-bloggers