BBopt

BBopt aims to provide the easiest hyperparameter optimization you'll ever do. Think of BBopt like Keras (back when Theano was still a thing) for black box optimization: one universal interface for working with any black box optimization backend.

BBopt's features include:

a universal API for defining your tunable parameters based on the standard library random module (so you don't even have to learn anything new!),
tons of state-of-the-art black box optimization algorithms such as Gaussian Processes from scikit-optimize or Tree Structured Parzen Estimation from hyperopt for tuning parameters,
the ability to switch algorithms while retaining all previous trials and even dynamically choose the best algorithm for your use case,
multiprocessing-safe data saving to enable running multiple trials in parallel,
lots of data visualization methods, including support for everything in skopt.plots,
support for optimizing over conditional parameters that only appear during some runs,
support for all major Python versions (2.7 or 3.6+), and
a straightforward interface for extending BBopt with your own custom algorithms.

Once you've defined your parameters, training a black box optimization model on those parameters is as simple as

bbopt your_file.py

and serving your file with optimized parameters as easy as

import your_file

Questions? Head over to BBopt's Gitter if you have any questions/comments/etc. regarding BBopt.

Installation

To get going with BBopt, simply install it with

pip install bbopt

or, to also install the extra dependencies necessary for running BBopt's examples, run pip install bbopt[examples].

Basic Usage

To use bbopt, just add

# BBopt setup:
from bbopt import BlackBoxOptimizer
bb = BlackBoxOptimizer(file=__file__)
if __name__ == "__main__":
    bb.run()

to the top of your file, then call a random method like

x = bb.uniform("x", 0, 1)

for each of the tunable parameters in your model, and finally add

bb.maximize(y)      or      bb.minimize(y)

to set the value being optimized. Then, run

bbopt <your file here> -n <number of trials> -j <number of processes>

to train your model, and just

import <your module here>

to serve it!

Note: Neither __file__ nor __name__ are available in Jupyter notebooks. In that case, just setup BBopt with:

import os

# BBopt setup:
from bbopt import BlackBoxOptimizer
bb = BlackBoxOptimizer(data_dir=os.getcwd(), data_name="my_project_name")

Examples

Some examples of BBopt in action:

random_example.py: Extremely basic example using the random backend.
skopt_example.py: Slightly more complex example making use of the gaussian_process algorithm from the scikit-optimize backend.
hyperopt_example.py: Example showcasing the tree_structured_parzen_estimator algorithm from the hyperopt backend.
meta_example.py: Example of using run_meta to dynamically choose an algorithm.
numpy_example.py: Example which showcases how to have numpy array parameters.
conditional_skopt_example.py: Example of having black box parameters that are dependent on other black box parameters using the gaussian_process algorithm from the scikit-optimize backend.
conditional_hyperopt_example.py: Example of doing conditional parameters with the tree_structured_parzen_estimator algorithm from the hyperopt backend.
bask_example.py: Example of using conditional parameters with a semi-random target using the bask_gp algorithm from the bayes-skopt backend.
pysot_example.py: Example of using the full API to implement an optimization loop and avoid the overhead of running the entire file multiple times while making use of the pySOT backend.
keras_example.py: Complete example of using BBopt to optimize a neural network built with Keras. Uses the full API to implement its own optimization loop and thus avoid the overhead of running the entire file multiple times.
any_fast_example.py: Example of using the default algorithm "any_fast" to dynamically select a good backend.
mixture_example.py: Example of using the mixture backend to randomly switch between different algorithms.
json_example.py: Example of using json instead of pickle to save parameters.

Full API

BBopt

Command-Line Interface

The bbopt command is extremely simple in terms of what it actually does. For the command bbopt <file> -n <trials> -j <processes>, BBopt simply runs python <file> a number of times equal to <trials>, split across <processes> different processes.

Why does this work? If you're using the basic boilerplate, then running python <file> will trigger the if __name__ == "__main__": clause, which will run a training episode. But when you go to import your file, the if __name__ == "__main__": clause won't get triggered, and you'll just get served the best parameters found so far. Since the command-line interface is so simple, advanced users who want to use the full API instead of the boilerplate need not use the bbopt command at all. If you want more information on the bbopt command, just run bbopt -h.

Black Box Optimization Methods

Constructor

BlackBoxOptimizer(file, *, tag=None, protocol=None)

BlackBoxOptimizer(data_dir, data_name, *, tag=None, protocol=None)

Create a new bb object; this should be done at the beginning of your program as all the other functions are methods of this object.

file is used by BBopt to figure out where to load and save data to, and should usually just be set to __file__. tag allows additional customization of the BBopt data file for when multiple BBopt instances might be desired for the same file. Specifically, BBopt will save data to os.path.splitext(file)[0] + "_" + tag + extension.

Alternatively, data_dir and data_name can be used to specify where to save and load data to. In that case, BBopt will save data to os.path.join(data_dir, data_name + extension) if no tag is passed, or os.path.join(data_dir, data_name + "_" + tag + extension) if a tag is given.

protocol determines how BBopt serializes data. If None (the default), BBopt will use pickle protocol 2, which is the highest version that works on both Python 2 and Python 3 (unless a json file is present, in which case BBopt will use json). To use the newest protocol instead, pass protocol=-1. If protocol="json", BBopt will use json instead of pickle, which is occasionally useful if you want to access your data outside of Python.

`run`

BlackBoxOptimizer.run(alg="any_fast")

Start optimizing using the given black box optimization algorithm. Use algs to get the valid values for alg.

If this method is never called, or called with alg="serving", BBopt will just serve the best parameters found so far, which is how the basic boilerplate works. Note that, if no saved parameter data is found, and a guess is present, BBopt will use that, which is a good way of distributing your parameter values without including all your saved parameter data.

`algs`

BlackBoxOptimizer.algs

A dictionary mapping the valid algorithms for use in run to the pair (backend, kwargs) of the backend and arguments to that backend that the algorithm corresponds to.

Supported algorithms are:

"serving" (serving backend) (used if run is never called)
"random" (random backend)
"tree_structured_parzen_estimator" (hyperopt backend)
"adaptive_tpe" (hyperopt backend; but only Python 3+)
"annealing" (hyperopt backend)
"gaussian_process" (scikit-optimize backend)
"random_forest" (scikit-optimize backend)
"extra_trees" (scikit-optimize backend)
"gradient_boosted_regression_trees" (scikit-optimize backend)
"bask_gaussian_process" (bayes-skopt backend)
"stochastic_radial_basis_function" (pySOT backend)
"expected_improvement" (pySOT backend)
"DYCORS" (pySOT backend)
"lower_confidence_bound" (pySOT backend)
"latin_hypercube" (pySOT backend)
"symmetric_latin_hypercube" (pySOT backend)
"two_factorial" (pySOT backend)
"epsilon_max_greedy" (mixture backend)
"epsilon_greedy" (bandit backend)
"boltzmann_exploration" (bandit backend)
"boltzmann_gumbel_exploration" (bandit backend) (the default meta_alg in run_meta)
"openai" (openai backend)

Additionally, there are also some algorithms of the form safe_<other_alg> which use mixture to defer to <other_alg> if <other_alg> supports the parameter definition functions you're using, otherwise default to a suitable replacement.

algs also includes the following pseudo-algorithms which defer to run_meta:

"any_fast" (same as calling run_meta with a suite of algorithms selected for their speed except that some algorithms are ignored if unsupported parameter definition functions are used, e.g. normalvariate for scikit-optimize) (used if run is called with no args)
"any_hyperopt" (equivalent to calling run_meta with all hyperopt algorithms)
"any_skopt" (equivalent to calling run_meta with all scikit-optimize algorithms)
"any_pysot" (equivalent to calling run_meta with all pySOT algorithms)

Note: The bayes-skopt backend is only available on Python 3.7+ and the pySOT and openai backends are only available on Python 3+.

`run_meta`

BlackBoxOptimizer.run_meta(algs, meta_alg="boltzmann_gumbel_exploration")

run_meta is a special version of run that uses the meta_alg algorithm to dynamically pick an algorithm from among the given algs. Both algs and meta_alg can use any algorithms in algs.

`run_backend`

BlackBoxOptimizer.run_backend(backend, *args, **kwargs)

The base function behind run. Instead of specifying an algorithm, run_backend lets you specify the specific backend you want to call and the parameters you want to call it with. Different backends do different things with the remaining arguments:

scikit-optimize passes the arguments to skopt.Optimizer,
hyperopt passes the arguments to fmin,
mixture expects a distribution argument to specify the mixture of different algorithms to use, specifically a list of (alg, weight) tuples (and also admits a remove_erroring_algs bool to automatically remove erroring algorithms),
bayes-skopt passes the arguments to bask.Optimizer,
pySOT expects a strategy (either a strategy class or one of "SRBF", "EI", "DYCORS", "LCB"), a surrogate (either a surrogate class or one of "RBF", "GP"), and a design (either an experimental design class or one of None, "latin_hypercube", "symmetric_latin_hypercube", "two_factorial"), and
openai expects an engine (the name of the model to use), temperature, max_retries, and api_key (otherwise uses OPENAI_API_KEY env var).

Note: The bayes-skopt backend is only available on Python 3.7+ and the pySOT and openai backends are only available on Python 3+.

`minimize`

BlackBoxOptimizer.minimize(value)

Finish optimizing and set the loss for this run to value. To start another run, call run again.

`maximize`

BlackBoxOptimizer.maximize(value)

Same as minimize but sets the gain instead of the loss.

`remember`

BlackBoxOptimizer.remember(info)

Update the current run's "memo" field with the given info dictionary. Useful for saving information about a run that shouldn't actually impact optimization but that you would like to have access to later (using get_best_run, for example).

`plot_convergence`

BlackBoxOptimizer.plot_convergence(ax=None, yscale=None)

Plot the running best gain/loss over the course of all previous trials. If passed, ax should be the matplotlib axis to plot on and yscale should be the scale for the y axis.

Run BBopt's keras example to generate an example plot.

`plot_history`

BlackBoxOptimizer.plot_history(ax=None, yscale=None)

Plot the gain/loss at each point over the course of all previous trials. If passed, ax should be the matplotlib axis to plot on and yscale should be the scale for the y axis.

Run BBopt's keras example to generate an example plot.

`partial_dependence`

BlackBoxOptimizer.partial_dependence(i_name, j_name=None, sample_points=None, n_samples=250, n_points=40)

Calls skopt.plots.partial_dependence using previous trial data. The parameters i_name and j_name should be set to names of the parameters you want for the i and j arguments to skopt.plots.partial_dependence.

`plot_partial_dependence_1D`

BlackBoxOptimizer.plot_partial_dependence_1D(i_name, ax=None, yscale=None, sample_points=None, n_samples=250, n_points=40)

Plot the partial dependence of i_name on the given matplotlib axis ax and with the given y axis scale yscale. See partial_dependence for the meaning of the other parameters.

Run BBopt's keras example to generate an example plot.

`plot_evaluations`

BlackBoxOptimizer.plot_evaluations(bins=20)

Calls skopt.plots.plot_evaluations using previous trial data.

Run BBopt's keras example to generate an example plot.

`plot_objective`

BlackBoxOptimizer.plot_objective(levels=10, n_points=40, n_samples=250, size=2, zscale="linear")

Calls skopt.plots.plot_objective using previous trial data.

Run BBopt's keras example to generate an example plot.

`plot_regret`

BlackBoxOptimizer.plot_regret([ax, [true_minimum, [yscale]]])

Calls skopt.plots.plot_regret using previous trial data.

Run BBopt's keras example to generate an example plot.

`get_skopt_result`

BlackBoxOptimizer.get_skopt_result()

Gets an OptimizeResult object usable by skopt.plots functions. Allows for arbitrary manipulation of BBopt optimization results in scikit-optimize including any plotting functions not natively supported by BBopt.

`get_current_run`

BlackBoxOptimizer.get_current_run()

Get information on the current run, including the values of all parameters encountered so far and the loss/gain of the run if specified yet.

`get_best_run`

BlackBoxOptimizer.get_best_run()

Get information on the best run so far. These are the parameters that will be used if run is not called.

`get_data`

BlackBoxOptimizer.get_data(print_data=False)

Dump a dictionary containing "params"—the parameters BBopt knows about and what random function and arguments they were initialized with—and "examples"—all the previous data BBopt has collected. If print_data, pretty prints the data in addition to returning it.

`data_file`

BlackBoxOptimizer.data_file

The path of the file where BBopt is saving data to.

`is_serving`

BlackBoxOptimizer.is_serving

Whether BBopt is currently using the "serving" algorithm.

`tell_examples`

BlackBoxOptimizer.tell_examples(examples)

Add the given examples as in get_data to memory, writing the new data to data_file. Must come before run if you want the new data to be included in the model for that run.

`backend`

BlackBoxOptimizer.backend

The backend object being used by the current BlackBoxOptimizer instance.

`run_id`

BlackBoxOptimizer.run_id

The id of the current run if started by the BBopt command-line interface.

Parameter Definition Methods

Every BBopt parameter definition method has the form

bb.<random function>(<name>, <args>, **kwargs)

where

the method itself specifies what distribution is being modeled,
the first argument is always name, a unique string identifying that parameter,
following name are whatever arguments are needed to specify the distribution's parameters, and
at the end are keyword arguments, which are the same for all the different methods. The supported kwargs are:
- guess, which specifies the initial value for the parameter, and
- placeholder_when_missing, which specifies what placeholder value a conditional parameter should be given if missing.

Important note: Once you bind a name to a parameter, you cannot change that parameter's options. Thus, if the options defining your parameters can vary from run to run, you must use a different name for each possible combination.

`randrange`

BlackBoxOptimizer.randrange(name, stop, **kwargs)

BlackBoxOptimizer.randrange(name, start, stop, step=1, **kwargs)

Create a new parameter modeled by random.randrange(start, stop, step).

Backends which support randrange: scikit-optimize, hyperopt, bayes-skopt, pySOT, openai, random.

`randint`

BlackBoxOptimizer.randint(name, a, b, **kwargs)

Create a new parameter modeled by random.randint(a, b), which is equivalent to random.randrange(a, b-1).

Backends which support randint: scikit-optimize, hyperopt, bayes-skopt, pySOT, openai, random.

`getrandbits`

BlackBoxOptimizer.getrandbits(name, k, **kwargs)

Create a new parameter modeled by random.getrandbits(k), which is equivalent to random.randrange(0, 2**k).

Backends which support getrandbits: scikit-optimize, hyperopt, bayes-skopt, pySOT, openai, random.

`choice`

BlackBoxOptimizer.choice(name, seq, **kwargs)

Create a new parameter modeled by random.choice(seq), which chooses an element from seq.

Backends which support choice: scikit-optimize, hyperopt, bayes-skopt, pySOT, random.

`randbool`

BlackBoxOptimizer.randbool(name, **kwargs)

Create a new boolean parameter, modeled by the equivalent of random.choice([False, True]).

Backends which support randbool: scikit-optimize, hyperopt, bayes-skopt, pySOT, random.

`sample`

BlackBoxOptimizer.sample(name, population, k, **kwargs)

Create a new parameter modeled by random.sample(population, k), which chooses k elements from population.

By default, the ordering of elements in the result is random. If random ordering is not important and you're happy to have the same ordering as in population, BlackBoxOptimizer.unshuffled_sample is recommended instead.

Backends which support sample: scikit-optimize, hyperopt, bayes-skopt, pySOT, random.

`shuffle`

BlackBoxOptimizer.shuffle(name, population, **kwargs)

Create a new parameter modeled by random.shuffle(population). A version that returns the shuffled list instead of shuffling it in place is also supported as BlackBoxOptimizer.shuffled.

Backends which support shuffle: scikit-optimize, hyperopt, bayes-skopt, pySOT, random.

`random`

BlackBoxOptimizer.random(name, **kwargs)

Create a new parameter modeled by random.random(), which is equivalent to random.uniform(0, 1) except that 1 is disallowed.

Backends which support random: scikit-optimize, hyperopt, bayes-skopt, pySOT, openai, random.

`uniform`

BlackBoxOptimizer.uniform(name, a, b, **kwargs)

Create a new parameter modeled by random.uniform(a, b), which uniformly selects a float in the range [a, b].

Backends which support uniform: scikit-optimize, hyperopt, bayes-skopt, pySOT, openai, random.

`loguniform`

BlackBoxOptimizer.loguniform(name, min_val, max_val, **kwargs)

Create a new parameter modeled by

math.exp(random.uniform(math.log(min_val), math.log(max_val)))

which logarithmically selects a float between min_val and max_val.

Backends which support loguniform: scikit-optimize, hyperopt, bayes-skopt, pySOT, openai, random.

`normalvariate`

BlackBoxOptimizer.normalvariate(name, mu, sigma, **kwargs)

Create a new parameter modeled by random.normalvariate(mu, sigma).

A shortcut for the standard normal distribution is also available via BlackBoxOptimizer.stdnormal.

Backends which support normalvariate: hyperopt, openai, random.

`lognormvariate`

BlackBoxOptimizer.lognormvariate(name, mu, sigma, **kwargs)

Create a new parameter modeled by random.lognormvariate(mu, sigma) such that the natural log is a normal distribution with mean mu and standard deviation sigma.

Backends which support lognormvariate: hyperopt, openai, random.

`rand`

BlackBoxOptimizer.rand(name, *shape, **kwargs)

Create a new parameter modeled by numpy.random.rand(*shape), which creates a numpy array of the given shape with entries generated uniformly in [0, 1).

Backends which support rand: scikit-optimize, hyperopt, bayes-skopt, pySOT, openai, random.

`randn`

BlackBoxOptimizer.randn(name, *shape, **kwargs)

Create a new parameter modeled by numpy.random.randn(*shape), which creates a numpy array of the given shape with entries generated according to a standard normal distribution.

Backends which support randn: hyperopt, openai, random.

`param`

BlackBoxOptimizer.param(name, func, *args, **kwargs)

Create a new parameter modeled by the parameter definition function func with the given arguments. This function is mostly useful if you want to use a custom backend that implements parameter definition functions not included in BBopt by default.

Writing Your Own Backend

BBopt's backend system is built to be extremely extensible, allowing anyone to write and register their own BBopt backends. The basic template for writing a BBopt backend is as follows:

from bbopt.backends.util import StandardBackend

class MyBackend(StandardBackend):
    backend_name = "my-backend"
    implemented_funcs = [
        # list the random functions you support here
        #  (you don't need to include all random functions,
        #  only base random functions, primarily randrange,
        #  choice, uniform, and normalvariate)
        ...,
    ]

    def setup_backend(self, params, **options):
        # initialize your backend; you can use params
        #  to get the args for each param

    def tell_data(self, new_data, new_losses):
        # load new data points into your backend; new_data is
        #  a list of dictionaries containing data and new_losses
        #  is a list of losses for each of those data points

    def get_next_values(self):
        # return the values you want to use for this run as a dict

MyBackend.register()
MyBackend.register_alg("my_alg")

Once you've written a BBopt backend as above, you simply need to import it to trigger the register calls and enable it to be used in BBopt. For some example BBopt backends, see BBopt's default backends (written in Coconut):

random.coco
skopt.coco
bask.coco
hyperopt.coco
pysot.coco
serving.coco
mixture.coco
bandit.coco
openai.coco

Files

README.md

Latest commit

History

README.md

File metadata and controls

BBopt

Installation

Basic Usage

Examples

Full API

Command-Line Interface

Black Box Optimization Methods

Constructor

run

algs

run_meta

run_backend

minimize

maximize

remember

plot_convergence

plot_history

partial_dependence

plot_partial_dependence_1D

plot_evaluations

plot_objective

plot_regret

get_skopt_result

get_current_run

get_best_run

get_data

data_file

is_serving

tell_examples

backend

run_id

Parameter Definition Methods

randrange

randint

getrandbits

choice

randbool

sample

shuffle

random

uniform

loguniform

normalvariate

lognormvariate

rand

randn

param

Writing Your Own Backend

`run`

`algs`

`run_meta`

`run_backend`

`minimize`

`maximize`

`remember`

`plot_convergence`

`plot_history`

`partial_dependence`

`plot_partial_dependence_1D`

`plot_evaluations`

`plot_objective`

`plot_regret`

`get_skopt_result`

`get_current_run`

`get_best_run`

`get_data`

`data_file`

`is_serving`

`tell_examples`

`backend`

`run_id`

`randrange`

`randint`

`getrandbits`

`choice`

`randbool`

`sample`

`shuffle`

`random`

`uniform`

`loguniform`

`normalvariate`

`lognormvariate`

`rand`

`randn`

`param`