/chocolate

A fully decentralized hyperparameter optimization framework

Primary LanguagePythonBSD 3-Clause "New" or "Revised" LicenseBSD-3-Clause

Chocolate

Chocolate is a completely asynchronous optimisation framework relying solely on a database to share information between workers. Chocolate uses no master process for distributing tasks. Every task is completely independent and only gets its information from a database. Chocolate is thus ideal in controlled computing environments where it is hard to maintain a master process for the duration of the optimisation.

Chocolate has been designed and optimized for hyperparameter optimization where each function evaluation takes very long to complete and is difficult to parallelize. Chocolate allows optimization over conditional search spaces either as using conditional kernels in a Bayesian optimizer or as a multi-armed bandit problem using Thompson sampling. Chocolate also handles multi-objective optimisation where multiple loss funtions are optimized simultaneously.

Chocolate provides the following sampling/searching algorithms:

  • Grid
  • Random
  • QuasiRandom
  • CMAES
  • MOCMAES
  • Bayesian

and three useful backends:

  • SQlite
  • MongoDB
  • Pandas Data Frame

Chocolate is licensed under the 3-Clause BSD License

Documentation

The full documentation is available at http://chocolate.readthedocs.io.

Installation

Chocolate is installed using pip, unfortunately we don't have any PyPI package yet. Here is the line you have to type

pip install git+https://github.com/AIworx-Labs/chocolate@master

Dependencies

Chocolate has various dependencies. While the optimizers depends on NumPy, SciPy and Scikit-Learn, the SQLite database connection depends on dataset and filelock and the MongoDB database connection depends on PyMongo. Some utilities depend on pandas. All but PyMongo will be installed with Chocolate.

Simple example

The following very simple example shows how to optimize a conditional search space in Chocolate. You'll note that a single point is sampled and evaluated in the script. Since the database connections are 'parallel' safe, you can run this script in concurrent processes and achieve maximum parallelism.

import chocolate as choco

def objective_function(condition, x=None, y=None):
    """An objective function returning ``1 - x`` when *condition* is 1 and 
    ``y - 6`` when *condition* is 2.
    
    Raises:
        ValueError: If condition is different than 1 or 2.
    """
    if condition == 1:
        return 1 - x
    elif condition == 2:
        return y - 6
    raise ValueError("condition must be 1 or 2, got {}.".format(condition))

# Define the conditional search space 
space = [
            {"condition": 1, "x": choco.uniform(low=1, high=10)},
            {"condition": 2, "y": choco.log(low=-2, high=2, base=10)}
        ]

# Establish a connection to a SQLite local database
conn = choco.SQLiteConnection("sqlite:///my_db.db")

# Construct the optimizer
sampler = choco.Bayes(conn, space)

# Sample the next point
token, params = sampler.next()

# Calculate the loss for the sampled point (minimized)
loss = objective_function(**params)

# Add the loss to the database
sampler.update(token, loss)

Have a look at the documentation tutorials for more examples.