Skip to content

Latest commit

 

History

History
67 lines (47 loc) · 4.71 KB

usage-http.md

File metadata and controls

67 lines (47 loc) · 4.71 KB

Python HTTP Interface Usage

Introduction

This template aims to make it easy to evaluate recommendation algorithms using a simple HTTP REST interface written in Python (version 3.x).

Installation

Follow the steps in Installation section of usage.md for basic installation instructions.

Using the HTTP template

To let Idomaar evaluate your own recommendation algorithm using HTTP as transport protocol between the orchestrator and the computing environment, you have to let the server know how to fetch recommendations from your algorithm.

To do so, open file computingenvironments/01.linux/01.centos/01.mahout/algorithms/02.http/idomaar-http-server.py and edit the body of recommend() method (what follows is just an example):

def recommend():
    recRequestProperties = json.loads(request.form.getlist('properties')[0])
    recRequestEntities = json.loads(request.form.getlist('entities')[0])

    sessionId = int(recRequestEntities['subject'].split(":")[1])
    previous = recRequestProperties['previouses'][0]
    prevItemId = int(previous['object']['id'])

    # predict_next_items() is a function defined somewhere in this file and contains
    # instructions to interact with your own algorithm; this is just an example!
    algo_resp = predict_next_items(algo_handler, sessionId, prevItemId, 20)

    resp = {}
    resp['GT'] = recRequestProperties
    resp['rec'] = []

    # Each recommendation is a dict item of 'rec' list inside 'resp' dict and it is formed by
    # {id, rating, rank}. Response *must* follow this structure to be readable by the evaluator
    for i in range(len(algo_resp)):
        resp['rec'].append({"id": int(algo_resp[i][0]), "rating": float(algo_resp[i][1]), "rank": i+1})
    
    return app.make_response(json.dumps(resp))

Note that the response resp sent by the server must follow the structure described in the above sample source code, that is, it must be a dictionary with keys GT and rec. GT contains the list of items for which you are asking recommendations in Idomaar data format, and rec contains a list of recommendations saved as dictionaries with keys id (the next item identifier, as returned by your algorithm), rating (the score associated with the current recommendation) and rank (an integer representing the position of each recommendation with respect to all other recommendations; it can be simply an ascending counter if recommendations are ordered by descending score).

This file implements all the methods required to interact with the orchestrator (see Idomaar architecture); all requests are sent using POST method and the Python server listens to Idomaar's specific paths over POST method. In particular, the training phase is here skipped because you can train your model in a previous moment. If you want Idomaar to control your model training, you have to properly edit the train() method (see Idomaar evaluation process for details).

The full list of implemented methods (only POST requests are managed, since orchestrator sends POST requests) is:

  • Request to path /HELLO is sent by the orchestrator to test if the computing environment is ready; when it is ready, the server answers with string "READY"
  • Request to path /TRAIN causes starting your model training phase and the response is the string "OK\n<ip_address:port>" where ip_address:port is the recommendation endpoint
  • Request to path /TEST tells the computing environment that the orchestrator is ready to serve recommendation requests; the server answers with the string "OK"
  • Request to path / causes the start of the recommendation fetching process from the computing environment; the response is a string serialization of a JSON formatted object containing the recommendations
  • Request to path /STOP causes the evaluation framework to terminate. Once the server is ready, it answers with the string "OK"

After you have finished editing idomaar-http-server.py according to your needs, you can launch idomaar-demo.sh (or idomaar-demo.bat if you are using Windows) to start the evaluation process.

Troubleshooting

If the Python server doesn't automatically start when idomaar-demo.sh is executed, you need to start it from the computing environment by opening a terminal in computingenvironments\01.linux\01.centos\01.mahout folder and then launching the following commands:

vagrant ssh

sudo service idomaar-http-server start

If it does not work, try launching

python3 /vagrant/algorithms/02.http/idommar-http-server.py

from the remote shell opened before.