This project provides a benchmark framework to easily compare Bayesian optimization methods on real machine learning tasks.
This project is experimental and the APIs are not considered stable.
This Bayesian optimization (BO) benchmark framework requires a few easy steps for setup. It can be run either on a local machine (in serial) or prepare a commands file to run on a cluster as parallel experiments (dry run mode).
Only Python>=3.6 is officially supported, but older versions of Python likely work as well.
The core package itself can be installed with:
However, to also require installation of all the "built in" optimizers for evaluation, run:
It is also possible to use the same pinned dependencies we used in testing by installing from the repo.
Building an environment to run the included notebooks can be done with:
Or, bayesmark[optimizers,notebooks] can be used.
A quick example of running the benchmark is here. The instructions are used to generate results as below:
To be able to install opentuner some system level (non-pip) dependencies must be installed. This can be done with:
On Ubuntu, this results in:
> dpkg -l | grep libsqlite
ii libsqlite3-0:amd64 3.11.0-1ubuntu1 amd64 SQLite 3 shared library
ii libsqlite3-dev:amd64 3.11.0-1ubuntu1 amd64 SQLite 3 development filesThe environment should now all be setup to run the BO benchmark.
Now we can run each step of the experiments. First, we run all combinations and then run some quick commands to analyze the output.
The experiments are run using the experiment launcher, which has the following interface:
usage: bayesmark-launch [-h] [-dir DB_ROOT] [-odir OPTIMIZER_ROOT] [-v] [-u UUID]
[-dr DATA_ROOT] [-b DB] [-o OPTIMIZER [OPTIMIZER ...]]
[-d DATA [DATA ...]]
[-c [{DT,MLP-adam,MLP-sgd,RF,SVM,ada,kNN,lasso,linear} ...]]
[-m [{acc,mae,mse,nll} ...]] [-n N_CALLS]
[-p N_SUGGEST] [-r N_REPEAT] [-nj N_JOBS] [-ofile JOBS_FILE]The arguments are:
-h, --help show this help message and exit
-dir DB_ROOT, -db-root DB_ROOT
root directory for all benchmark experiments output
-odir OPTIMIZER_ROOT, --opt-root OPTIMIZER_ROOT
Directory with optimization wrappers
-v, --verbose print the study logs to console
-u UUID, --uuid UUID length 32 hex UUID for this experiment
-dr DATA_ROOT, --data-root DATA_ROOT
root directory for all custom csv files
-b DB, --db DB database ID of this benchmark experiment
-o OPTIMIZER [OPTIMIZER ...], --opt OPTIMIZER [OPTIMIZER ...]
optimizers to use
-d DATA [DATA ...], --data DATA [DATA ...]
data sets to use
-c, --classifier [{DT,MLP-adam,MLP-sgd,RF,SVM,ada,kNN,lasso,linear} ...]
classifiers to use
-m, --metric [{acc,mae,mse,nll} ...]
scoring metric to use
-n N_CALLS, --calls N_CALLS
number of function evaluations
-p N_SUGGEST, --suggestions N_SUGGEST
number of suggestions to provide in parallel
-r N_REPEAT, --repeat N_REPEAT
number of repetitions of each study
-nj N_JOBS, --num-jobs N_JOBS
number of jobs to put in the dry run file, the default
0 value disables dry run (real run)
-ofile JOBS_FILE, --jobs-file JOBS_FILE
a jobs file with all commands to be runThe output files will be placed in [DB_ROOT]/[DBID]. If DBID is not specified, it will be a randomly created subdirectory with a new name to avoid overwriting previous experiments. The path to DBID is shown at the beginning of stdout when running bayesmark-launch. In general, let the launcher create and setup DBID unless you are appending to a previous experiment, in which case, specify the existing DBID.
The launcher's sequence of commands can be accessed programmatically via .experiment_launcher.gen_commands. The individual experiments can be launched programmatically via .experiment.run_sklearn_study.
A list of optimizers, classifiers, data sets, and metrics can be listed using the -o/-c/-d/-m commands, respectively. If not specified, the program launches all possible options.
A few different open source optimizers have been included as an example and are considered the "built-in" optimizers. The original repos are shown in the Links.
The data argument -o allows a list containing the "built-in" optimizers:
"HyperOpt", "Nevergrad-OnePlusOne", "OpenTuner-BanditA", "OpenTuner-GA", "OpenTuner-GA-DE", "PySOT", "RandomSearch", "Scikit-GBRT-Hedge", "Scikit-GP-Hedge", "Scikit-GP-LCB"or, one can specify a user-defined optimizer. The class containing an optimizer conforming to the API must be found in in the folder specified by --opt-root. Additionally, a configuration defining each optimizer must be defined in [OPT_ROOT]/config.json. The --opt-root and config.json may be omitted if only built-in optimizers are used.
Additional details for providing a new optimizer are found in adding a new optimizer.
By default, this benchmark uses the sklearn example data sets as the "built-in" data sets for use in ML model tuning problems.
The data argument -d allows a list containing the "built-in" data sets:
"breast", "digits", "iris", "wine", "boston", "diabetes"or, it can refer to a custom csv file, which is the name of file in the folder specified by --data-root. It also follows the convention that regression data sets start with reg- and classification data sets start with clf-. For example, the classification data set in [DATA_ROOT]/clf-foo.csv is specified with -d clf-foo.
The csv file can be anything readable by pandas, but we assume the final column is the target and all other columns are features. The target column should be integer for classification data and float for regression. The features should float (or str for categorical variable columns). See bayesmark.data.load_data for more information.
It is also possible to do a "dry run" of the launcher by specifying a value for --num-jobs greater than zero. For example, if --num-jobs 50 is provided, a text file listing 50 commands to run is produced, with one command (job) per line. This is useful when preparing a list of commands to run later on a cluster.
A dry run will generate a command file (e.g., jobs.txt) like the following (with a meta-data header). Each line corresponds to a command that can be used as a job on a different worker:
# running: {'--uuid': None, '-db-root': '/foo', '--opt-root': '/example_opt_root', '--data-root': None, '--db': 'bo_example_folder', '--opt': ['RandomSearch', 'PySOT'], '--data': None, '--classifier': ['SVM', 'DT'], '--metric': None, '--calls': 15, '--suggestions': 1, '--repeat': 3, '--num-jobs': 50, '--jobs-file': '/jobs.txt', '--verbose': False, 'dry_run': True, 'rev': '9a14ef2', 'opt_rev': None}
# cmd: python bayesmark-launch -n 15 -r 3 -dir foo -o RandomSearch PySOT -c SVM DT -nj 50 -b bo_example_folder
job_e2b63a9_00 bayesmark-exp -c SVM -d diabetes -o PySOT -u 079a155f03095d2ba414a5d2cedde08c -m mse -n 15 -p 1 -dir foo -b bo_example_folder && bayesmark-exp -c SVM -d boston -o RandomSearch -u 400e4c0be8295ad59db22d9b5f31d153 -m mse -n 15 -p 1 -dir foo -b bo_example_folder && bayesmark-exp -c SVM -d digits -o RandomSearch -u fe73a2aa960a5e3f8d78bfc4bcf51428 -m acc -n 15 -p 1 -dir foo -b bo_example_folder
job_e2b63a9_01 bayesmark-exp -c DT -d diabetes -o PySOT -u db1d9297948554e096006c172a0486fb -m mse -n 15 -p 1 -dir foo -b bo_example_folder && bayesmark-exp -c SVM -d boston -o RandomSearch -u 7148f690ed6a543890639cc59db8320b -m mse -n 15 -p 1 -dir foo -b bo_example_folder && bayesmark-exp -c SVM -d breast -o PySOT -u 72c104ba1b6d5bb8a546b0064a7c52b1 -m nll -n 15 -p 1 -dir foo -b bo_example_folder
job_e2b63a9_02 bayesmark-exp -c SVM -d iris -o PySOT -u cc63b2c1e4315a9aac0f5f7b496bfb0f -m nll -n 15 -p 1 -dir foo -b bo_example_folder && bayesmark-exp -c DT -d breast -o RandomSearch -u aec62e1c8b5552e6b12836f0c59c1681 -m nll -n 15 -p 1 -dir foo -b bo_example_folder && bayesmark-exp -c DT -d digits -o RandomSearch -u 4d0a175d56105b6bb3055c3b62937b2d -m acc -n 15 -p 1 -dir foo -b bo_example_folder
...This package does not have built in support for deploying these jobs on a cluster or cloud environment (.e.g., AWS).
The UUID is a 32-char hex string used as a master random seed which we use to draw random seeds for the experiments. If UUID is not specified a version 4 UUID is generated. The used UUID is displayed at the beginning of stdout. In general, the UUID should not specified/re-used except for debugging because it violates the assumption that the experiment UUIDs are unique.
Next to aggregate all the experiment files into combined (json) files we need to run the aggregation command:
usage: bayesmark-agg [-h] [-dir DB_ROOT] [-odir OPTIMIZER_ROOT] [-v] -b DB [-rv]The arguments are:
-h, --help show this help message and exit
-dir DB_ROOT, -db-root DB_ROOT
root directory for all benchmark experiments output
-odir OPTIMIZER_ROOT, --opt-root OPTIMIZER_ROOT
Directory with optimization wrappers
-v, --verbose print the study logs to console
-b DB, --db DB database ID of this benchmark experiment
-rv, --ravel ravel all studies to store batch suggestions as if
they were serialThe DB_ROOT must match the folder from the launcher bayesmark-launch, and DBID must match that displayed from the launcher as well. The aggregate files are found in [DB_ROOT]/[DBID]/derived.
The result aggregation can be done programmatically via .experiment_aggregate.concat_experiments.
Finally, to run a statistical analysis presenting a summary of the experiments we run
usage: bayesmark-anal [-h] [-dir DB_ROOT] [-odir OPTIMIZER_ROOT] [-v] -b DBThe arguments are:
-h, --help show this help message and exit
-dir DB_ROOT, -db-root DB_ROOT
root directory for all benchmark experiments output
-odir OPTIMIZER_ROOT, --opt-root OPTIMIZER_ROOT
Directory with optimization wrappers
-v, --verbose print the study logs to console
-b DB, --db DB database ID of this benchmark experimentThe DB_ROOT must match the folder from the launcher bayesmark-launch, and DBID must match that displayed from the launcher as well. The aggregate files are found in [DB_ROOT]/[DBID]/derived.
The bayesmark-anal command looks for a baseline.json file in [DB_ROOT]/[DBID]/derived, which states the best possible and random search performance. If no such file is present, bayesmark-anal automatically calls bayesmark-baseline to build it. The baselines are inferred from the random search performance in the logs. The baseline values are considered fixed (not random) quantities when bayesmark-anal builds confidence intervals. Therefore, we allow the user to leave them fixed and do not rebuild them when bayesmark-anal is called if a baselines file is already present.
The result analysis can be done programmatically via .experiment_analysis.compute_aggregates, and the baseline computation via .experiment_baseline.compute_baseline.
See how-scoring-works for more information on how the scores are computed and aggregated.
After finishing the setup (environment) a small-scale serial can be run as follows:
> # setup
> DB_ROOT=./notebooks # path/to/where/you/put/results
> DBID=bo_example_folder
> mkdir $DB_ROOT
> # experiments
> bayesmark-launch -n 15 -r 3 -dir $DB_ROOT -b $DBID -o RandomSearch PySOT -c SVM DT -v
Supply --uuid 3adc3182635e44ea96969d267591f034 to reproduce this run.
Supply --dbid bo_example_folder to append to this experiment or reproduce jobs file.
User must ensure equal reps of each optimizer for unbiased results
-c DT -d boston -o PySOT -u a1b287b450385ad09b2abd7582f404a2 -m mae -n 15 -p 1 -dir /notebooks -b bo_example_folder
-c DT -d boston -o PySOT -u 63746599ae3f5111a96942d930ba1898 -m mse -n 15 -p 1 -dir /notebooks -b bo_example_folder
-c DT -d boston -o RandomSearch -u 8ba16c880ef45b27ba0909199ab7aa8a -m mae -n 15 -p 1 -dir /notebooks -b bo_example_folder
...
0 failures of benchmark script after 144 studies.
done
> # aggregate
> bayesmark-agg -dir $DB_ROOT -b $DBID
> # analyze
> bayesmark-anal -dir $DB_ROOT -b $DBID -v
...
median score @ 15:
optimizer
PySOT_0.2.3_9b766b6 0.330404
RandomSearch_0.0.1_9b766b6 0.961829
mean score @ 15:
optimizer
PySOT_0.2.3_9b766b6 0.124262
RandomSearch_0.0.1_9b766b6 0.256422
normed mean score @ 15:
optimizer
PySOT_0.2.3_9b766b6 0.475775
RandomSearch_0.0.1_9b766b6 0.981787
doneThe aggregate result files (i.e., summary.json) will now be available in $DB_ROOT/$DBID/derived. However, this will be high variance since it was from only 3 trials and only to 15 function evaluations.
Plotting the quantitative results found in $DB_ROOT/$DBID/derived can be done using the notebooks found in the notebooks/ folder of the git repository. The notebook plot_mean_score.ipynb generates plots for aggregate scores averaging over all problems. The notebook plot_test_case.ipynb generates plots for each test problem.
To use the notebooks, first copy over the notebooks/ folder from git repository.
To setup the kernel for running the notebooks use:
Now, the notebooks for plotting can be run with the command jupyter notebook and selecting the kernel bobm_ipynb.
It is also possible to convert the notebooks to an HTML report at the command line using nbconvert. For example, use the command:
The output file will be in ./notebooks/plot_mean_score.html. Here is an example export. See the nbconvert documentation page for more output formats. By default, the notebooks look in ./notebooks/bo_example_folder/ for the summary.json from bayesmark-anal.
To run plot_test_case.ipynb use the command:
The --ExecutePreprocessor.timeout=600 timeout increase is needed due to the large number of plots being generated. The output will be in ./notebooks/plot_test_case.html.
All optimizers in this benchmark are required to follow the interface specified of the AbstractOptimizer class in bayesmark.abstract_optimizer. In general, this requires creating a wrapper class around the new optimizer. The wrapper classes must all be placed in a folder referred to by the --opt-root argument. This folder must also contain the config.json folder.
The interface is simple, one must merely implement the suggest and observe functions. The suggest function generates new guesses for evaluating the function. Once evaluated, the function evaluations are passed to the observe function. The objective function is not evaluated by the optimizer class. The objective function is evaluated on outside and results are passed to observe. This is the correct setup for Bayesian optimization because:
- We can observe/try inputs that were never suggested
- We can ignore suggestions
- The objective function may not be something as simple as a Python function
So passing the function as an argument as is done in scipy.optimization is artificially restrictive.
The implementation of the wrapper will look like the following:
from bayesmark.abstract_optimizer import AbstractOptimizer
from bayesmark.experiment import experiment_main
class NewOptimizerName(AbstractOptimizer):
# Used for determining the version number of package used
primary_import = "name of import used e.g, opentuner"
def __init__(self, api_config, optional_arg_foo=None, optional_arg_bar=None):
"""Build wrapper class to use optimizer in benchmark.
Parameters
----------
api_config : dict-like of dict-like
Configuration of the optimization variables. See API description.
"""
AbstractOptimizer.__init__(self, api_config)
# Do whatever other setup is needed
# ...
def suggest(self, n_suggestions=1):
"""Get suggestion from the optimizer.
Parameters
----------
n_suggestions : int
Desired number of parallel suggestions in the output
Returns
-------
next_guess : list of dict
List of `n_suggestions` suggestions to evaluate the objective
function. Each suggestion is a dictionary where each key
corresponds to a parameter being optimized.
"""
# Do whatever is needed to get the parallel guesses
# ...
return x_guess
def observe(self, X, y):
"""Feed an observation back.
Parameters
----------
X : list of dict-like
Places where the objective function has already been evaluated.
Each suggestion is a dictionary where each key corresponds to a
parameter being optimized.
y : array-like, shape (n,)
Corresponding values where objective has been evaluated
"""
# Update the model with new objective function observations
# ...
# No return statement needed
if __name__ == "__main__":
# This is the entry point for experiments, so pass the class to experiment_main to use this optimizer.
# This statement must be included in the wrapper class file:
experiment_main(NewOptimizerName)Depending on the API of the optimizer being wrapped, building this wrapper class may only or require a few lines of code, or be a total pain.
Note: A config file is now optional. If no config.json is provided, the experiment launcher will look for all folders with an optimizer.py in the --opt-root directory.
Each optimizer wrapper can have multiple configurations, which is each referred to as a different optimizer in the benchmark. For example, the JSON config file will have entries as follows:
Basically, the entries are "name_of_strategy": ["file_with_class", {kwargs_for_the_constructor}]. Here, OpenTuner-BanditA, OpenTuner-GA-DE, and OpenTuner-GA are all treated as different optimizers by the benchmark even though the all use the same class from opentuner_optimizer.py.
This config.json must be in the same folder as the optimizer classes (e.g., opentuner_optimizer.py).
To run the benchmarks using a new optimizer, simply provide its name (from config.json) in the -o list. The --opt-root argument must be specified in this case. For example, the launch command from the example becomes:
Here, we are using the example PySOT-New wrapper from the example_opt_root folder in the git repo. It is equivalent to the builtin PySOT, but gives an example of how to provide a new custom optimizer.
The following instructions have been tested with Python 3.6.8 on Ubuntu (16.04.5 LTS).
First, define the variables for the paths we will use:
Then clone the repo in your git directory $GIT:
Inside your virtual environments folder $ENVS, make the environment:
Now we can install the pip dependencies. Move back into your git directory and run
You may want to run pip install -U pip first if you have an old version of pip. The file optimizers.txt contains the dependencies for all the optimizers used in the benchmark. The analysis and aggregation programs can be run using only the requirements in base.txt.
First, we need to setup some needed tools:
To install the pre-commit hooks for contributing run (in the bayesmark_tools environment):
To rebuild the requirements, we can run:
cd $GIT/bayesmark
# Get py files from notebooks to analyze
jupyter nbconvert --to script notebooks/*.ipynb
# Generate the .in files (but pins to latest, which we might not want)
pipreqs bayesmark/ --ignore bayesmark/builtin_opt/ --savepath requirements/base.in
pipreqs test/ --savepath requirements/test.in
pipreqs bayesmark/builtin_opt/ --savepath requirements/optimizers.in
pipreqs notebooks/ --savepath requirements/ipynb.in
pipreqs docs/ --savepath requirements/docs.in
# Regenerate the .txt files from .in files
pip-compile-multi --no-upgradeFirst setup the environment for building with Sphinx:
Then we can do the build:
Documentation will be available in all formats in Makefile. Use make html to only generate the HTML documentation.
The tests for this package can be run with:
The script creates a conda environment using the requirements found in requirements/test.txt.
The test.sh script must be run from a clean git repo.
Or if we only want to run the unit tests and not check the adequacy of the requirements files, one can use
# Setup environment
cd $ENVS
virtualenv bayesmark_test --python=python3.6
source $ENVS/bayesmark_test/bin/activate
pip install -r $GIT/bayesmark/requirements/test.txt
pip install -e $GIT/bayesmark
# Now run tests
cd $GIT/bayesmark/
pytest test/ -s -v --hypothesis-seed=0 --disable-pytest-warnings --cov=bayesmark --cov-report htmlA code coverage report will also be produced in $GIT/bayesmark/htmlcov/index.html.
The wheel (tar ball) for deployment as a pip installable package can be built using the script:
The source is hosted on GitHub.
The documentation is hosted at Read the Docs.
Installable from PyPI.
The builtin optimizers are wrappers on the following projects:
This project is licensed under the Apache 2 License - see the LICENSE file for details.