Expose R Machine Learning Project as REST API

This repo is a small setup to quickly refactor your R machine learning script in production ready format via REST API service. The purpose of this meta project is to create REST API for your machine learning project in R and expose your model to the end user. The project contains ready-made script to create the below APIs in R:

• preprocessing
• training
• prediction

The user only needs to plug his original preprocessing, training and prediction logic as function. And then call the respective modules from the respective API functions. Currently to simulate the above three tasks, dummy code snippet is used, which the user will replace with his appropriate code.

The code also implements basic logging facility for debugging purpose.

Package Information

• package plumber: for creating api
• package futile.logger: for capturing the log
• package future: for creating asynchronous function call for preprocessing and training

Main Scripts

• The prediction_API.R holds the actual code for prediciton module for the /predict api.
  • run_prediction_API.R is a wrapper script to start the prediction_API.R
• The training_model_API.py holds the code for preprocessing and training module for the apis /preprocess and /train
  • run_training_API.R is a wrapper script to start the training_model_API.R
• The dummyAPI.R under directory /testdir holds the dummy scipt for /notify api for implementing callback facility
• The start.sh scripts can be used to start all the server at one go
• You may need to start the dummy api separately.

Start the PREDICTION SERVER (/predict API):

• Rscript run_prediction_API.R <prediction port number> OR
  • use the start.sh file, run it as follows:
  • ./start.sh
  • Remember to update the start.sh for port changing

Run the client from terminal

• following curl command can be used:
  • curl http://localhost:<prediction_server_port>/predict --data '{"input": [{"UID":"1","AGE":"15"},{"UID":"2","AGE":"25"},{"UID":"3","AGE":"45"},{"UID":"4","AGE":"55"}]}' -H "Content-Type: application/json"
  • /predict is the prediction API end point
  • You may need to update the port number in the above curl command

Start the Training SERVER (/preprocess and /train API):

• Rscript run_training_API.R <train port number> OR
• use the start.sh file and run ./start.sh
• As training and preprocessing can take several minutes, they should be called asynchronously using python package future followed by a post call to the callback API for sending the status of preprocessing and training.

Run the client from the terminal

• curl for preprocessing api /preprocess
  • curl http://localhost:<train_server_port>/prepro --data '{"callbackURL":"http://localhost:<notify_server_port>/notify"}' -H "Content-Type: application/json"
• curl for training api /train
  • curl http://localhost:<train_server_port>/train --data '{"callbackURL":"http://localhost:<notify_server_port>/notify"}' -H "Content-Type: application/json"
• The above callbackURL is a dummy api for testing. The dummy api script is in /testdir folder. Run the dummy callback API python dummy_API.py

Start the Notify Server for Callback facility

• Rscript notifyAPI.R <notify port number>

Known issues for the repo:

• Currently, the asynchronous call by the future package is not sending the response to client immediately.
• ...

Known issues in R

• Nested file source may lose the argument value. So use sourcing as follows source(filename, local=TRUE)
• Rememeber sourcing a file outside a function and inside a function are two different scenario.
  • It changes the scope of the variable to the sourced scripts.
  • global variebles are not available to a sourced script if it's sourced inside a function.