Welcome to Polaris’s documentation!

Contents:

• 1. What is Polaris?
  • 1.1. About
  • 1.2. Algorithms
  • 1.3. Architecture
• 2. Installation
  • 2.1. Polaris
  • 2.2. RabbitMQ
• 3. Example
  • 3.1. Single Execution
  • 3.2. Parallel Execution

class polaris.Bounds(label, low, high, q=None)

A class to declare range of each hyperparameters. For now, Polaris do not adapt to category variable.

class polaris.Domain(bounds, algo='random')

A class to store bounds and searching new parameters.

class polaris.Polaris(fn, bounds, algo, trials, max_evals=10, run_once=False, exp_key=None, logger=None, debug=False, args=None)

A client for the Polaris.

run(use_mpi=False)

Run an experiment sequentially with one process up to max_evals count. The expriment will early stop, if params become None

use_mpi : boolean

boolean for using mpi or not

run_parallel()

Run an experiment in parallel.

Polaris use RabbitMQ to pass arguments from client to worker. You need to start at least one worker to start an experiment.

class polaris.Trials

A class for storing trials data.

add(result, params, exp_info=None)

A method to add each trial result

get_train_data()

A method to get train data. The train data is arranged in alphabetical order of the name of parameters.

class polaris.rabbitmq.JobClient(polaris)

A client class for parallel experiments.

The instance of this class send a new job to workers and calculate next parameters in response to the requsest from a worker.

This client adopt the RPC pattern. Therefore all results will be accumulated on client side.

on_request(ch, method, props, body)

A method to receive job request from workers After receiving request, this method will send a job to them.

on_response(ch, method, props, body)

A method to receive the result of experiment from workers.

start()

A method to start consuming job requests.

class polaris.rabbitmq.JobWorker(args, logger=None, debug=False)

A worker class for parallel experiments.

You can start the worker like below. polaris-worker –exp-key this_is_test

And if you want to run this worker on multi node environment, you have to add –mpi flag.