Nevergrad - A gradient-free optimization platform

nevergradis a Python 3.6+ library. It can be installed with:

pip install nevergrad

More installation options, including windows installation, and complete instructions are available in the "Getting started" section of thedocumentation.

You can join Nevergrad users Facebook grouphere.

Minimizing a function using an optimizer (hereNGO) is straightforward:

importnevergradasng

defsquare(x):
returnsum((x-.5)**2)

optimizer=ng.optimizers.NGO(parametrization=2,budget=100)
recommendation=optimizer.minimize(square)
print(recommendation.value)# recommended value
>>>[0.499711120.5002944]

nevergradcan also support bounded continuous variables as well as discrete variables, and mixture of those. To do this, one can specify the input space:

importnevergradasng

deffake_training(learning_rate:float,batch_size:int,architecture:str)->float:
# optimal for learning_rate=0.2, batch_size=4, architecture= "conv"
return(learning_rate-0.2)**2+(batch_size-4)**2+(0ifarchitecture=="conv"else10)

# Instrumentation class is used for functions with multiple inputs
# (positional and/or keywords)
parametrization=ng.p.Instrumentation(
# a log-distributed scalar between 0.001 and 1.0
learning_rate=ng.p.Log(lower=0.001,upper=1.0),
# an integer from 1 to 12
batch_size=ng.p.Scalar(lower=1,upper=12).set_integer_casting(),
# either "conv" or "fc"
architecture=ng.p.Choice(["conv","fc"])
)

optimizer=ng.optimizers.NGO(parametrization=parametrization,budget=100)
recommendation=optimizer.minimize(fake_training)

# show the recommended keyword arguments of the function
print(recommendation.kwargs)
>>>{'learning_rate':0.1998,'batch_size':4,'architecture':'conv'}

Learn more on parametrization in thedocumentation!

Convergence of a population of points to the minima with two-points DE.

Documentation

Check out ourdocumentation!It's still a work in progress, don't hesitate to submit issues and/or PR to update it and make it clearer!

Citing

@misc{nevergrad,
author={J. Rapin and O. Teytaud},
title={{Nevergrad - A gradient-free optimization platform}},
year={2018},
publisher={GitHub},
journal={GitHub repository},
howpublished={\url{https://GitHub.com/FacebookResearch/Nevergrad}},
}

License

nevergradis released under the MIT license. SeeLICENSEfor additional details about it. See also ourTerms of UseandPrivacy Policy.