Nevergrad - A gradient-free optimization platform

Meta Research

Last update: Jan 8, 2023

Related tags

Machine Learning nevergrad

Overview

Nevergrad - A gradient-free optimization platform

nevergrad is 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 the documentation.

You can join Nevergrad users Facebook group here.

Minimizing a function using an optimizer (here NGOpt) is straightforward:

import nevergrad as ng

def square(x):
    return sum((x - .5)**2)

optimizer = ng.optimizers.NGOpt(parametrization=2, budget=100)
recommendation = optimizer.minimize(square)
print(recommendation.value)  # recommended value
>>> [0.49971112 0.5002944]

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

>> {'learning_rate': 0.1998, 'batch_size': 4, 'architecture': 'conv'} ">

import nevergrad as ng

def fake_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 + (0 if architecture == "conv" else 10)

# 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.NGOpt(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 the documentation!

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

Documentation

Check out our documentation! 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

nevergrad is released under the MIT license. See LICENSE for additional details about it. See also our Terms of Use and Privacy Policy.

Comments

Adding IOHexperimenter functions
Types of changes

[ ] Docs change / refactoring / dependency upgrade

[ ] Bug fix (non-breaking change which fixes an issue)

[ x] New feature (non-breaking change which adds functionality)

[ ] Breaking change (fix or feature that would cause existing functionality to change)

Motivation and Context / Related issue

Adds the functions from the PBO-suite of the IOHexperimenter into nevergrad (#338)

How Has This Been Tested (if it applies)

Still in progress

Checklist

[ ] The documentation is up-to-date with the changes I made.

[ x] I have read the CONTRIBUTING document and completed the CLA (see CLA).

[ ] All tests passed, and additional code has been covered with new tests.

CLA Signed
opened by Dvermetten 22
add support for alternative cmaes implementation
Types of changes

[ ] Docs change / refactoring / dependency upgrade

[ ] Bug fix (non-breaking change which fixes an issue)

[x ] New feature (non-breaking change which adds functionality)

[ ] Breaking change (fix or feature that would cause existing functionality to change)

Motivation and Context / Related issue

Bad performance of cmaes for higher dimensions

How Has This Been Tested (if it applies)

Tested by including the new optimization algorithm in several benchmarks and executed them successfully

Checklist

[ ] The documentation is up-to-date with the changes I made.

[x ] I have read the CONTRIBUTING document and completed the CLA (see CONTRIBUTING).

[ ] All tests passed, and additional code has been covered with new tests.

Purpose of this PR is:

a) Enable experiments with an alternative CMA implementation - install with "pip install fcmaes".

b) Exposing popsize as an important CMA parameter

c) Reading _es.stop() / es.stop to check whether CMA is terminated and kill the rest of the assigned budget in this case. Otherwise time comparison with other algos not able to see if they are stuck is unfair. May be some of them can, then this idea should be generalized.

General observations:

Benchmarks where the optimization algos are configured with workers=1 run fastest if

a) Benchmark multiprocessing is configured using workers=multiprocessing.cpu_count(). I checked this with a variety of CPUs from Intel + AMD.

b) The algos use only one thread, which means you have to set
export "MKL_NUM_THREADS=1" or export "OPENBLAS_NUM_THREADS=1" if numpy is configured to use OPENBLAS. Without doing this CMA, which is heavily based on BLAS, suffers badly - sometimes its factor 10 slower.

c) Use "export MKL_DEBUG_CPU_TYPE=5" if you are using an AMD CPU, otherwise Intel MKL "ignores" the advanced SIMD instructions of your CPU.

popsize is an important CMA parameter which should be benchmarked using different settings. Higher popsize means a broader search which can slow down the optimization for lower budgets but often pays off when the budget is higher.

Scaling should not only be viewed as a scalar, often the dimensions should be scaled separately. FCMA does this automatically if bounds are defined, (lb, ub) is mapped on ([-1]*dim, [1]*dim). Nevergrad doesn't provide the optimization algorithms with information about bounds, which may be a bad idea.

The CMA implementation https://github.com/dietmarwo/fast-cma-es offered as an alternative is much faster with higher dimensions because it better utilizes the underlying BLAS library and avoids loops when possible. There is no "DiagonalOnly" option supported yet, but test have to show if this option makes still sense.

With all the benchmarks / tests implemented in Nevergrad it should be possible to check:

If there are cases were FCMA performs worse than CMA.

How the execution times compare.

Which are the optimal parameters for FCMA

How to reconfigure NGO - if and when to use FCMA with which parameters - to achieve optimal results.

My tests have shown that the advantage of FCMA is higher if both CMAs are compared standalone or with an efficient parallel retry mechanism. But nevertheless FCMA is significantly faster, specialy with higher dimensions outperforming most of the other algorithms regarding execution time.
CLA Signed
opened by dietmarwo 18
NSGA-II implementation
Types of changes

[ ] Docs change / refactoring / dependency upgrade

[ ] Bug fix (non-breaking change which fixes an issue)

[x] New feature (non-breaking change which adds functionality)

[ ] Breaking change (fix or feature that would cause existing functionality to change)

Motivation and Context / Related issue

Related issue: https://github.com/facebookresearch/nevergrad/issues/820 Add the NSGA-II algorithm. The key procedures are the computation of the nondomination rank and crowding distance for the candidates. The initialization, crossover, and mutation procedures are borrowed from DE and ES in the current library.

How Has This Been Tested

The nondomination rank and crowding distance are tested by applying them on manually created candidates (i.e. p.Parameter variables), whose losses are preset.

The NSGA-II algorithm is tested on a multiobjective function, which is originally used to test CMA.

Test for correct population size as defined in test_optimizerlib.py

Things to Review

In the current implementation, I'm not quite sure if I use "ask" and "tell" methods correctly.

Not sure why "_internal_tell_not_asked" will happen as in DE.

Is the custom "recommend" method in DE necessary? I found that ES does not implement "recommend".

Do I need to update the "heritage" variable for a candidate?

Checklist

[ ] The documentation is up-to-date with the changes I made.

[x] I have read the CONTRIBUTING document and completed the CLA (see CLA).

[ ] All tests passed, and additional code has been covered with new tests.

CLA Signed
opened by pacowong 17
ElectricalMix Simulator using nevergrad
pull request to apply for the Nevergrad competition

Types of changes

[ ] Docs change / refactoring / dependency upgrade

[ ] Bug fix (non-breaking change which fixes an issue)

[x] New feature (non-breaking change which adds functionality)

[ ] Breaking change (fix or feature that would cause existing functionality to change)

Motivation and Context / Related issue

How Has This Been Tested (if it applies)

Checklist

[x] The documentation is up-to-date with the changes I made.

[x] I have read the CONTRIBUTING document and completed the CLA (see CLA).

[ ] All tests passed, and additional code has been covered with new tests.

CLA Signed
opened by Foloso 17
How to handle randomness in optimizer
Hi, I have prevously worked¹ on a gradient-free optimization algorithm called SPSA²-³, and I have Matlab/mex code⁴ that I can port to python easily. I am interested in benchmarking SPSA against other zero-order optimization algorithms using nevergrad.

I am following the instructions for benchmarking a new optimizer given in adding-your-own-experiments-andor-optimizers-andor-function. My understanding is that I can just add a new SPSA class in nevergrad/optimization/optimizerlib.py and implement

__init

_internal_ask

_internal_provide_recommendation

_internal_tell

functions and then add SPSA to the optims variable in the right experiment function in the nevergrad/benchmark/experiments.py module and then I should be able to generate graphs like docs/resources/noise_r400s12_xpresults_namecigar,rotationTrue.png

However, SPSA itself uses an rng in the _internal_ask function. But the optimizer base class does not take any seed in the __init__ function. What will be a good way to make the experiments reproducible in such situation?

[1] Pushpendre Rastogi, Jingyi Zhu, James C. Spall (2016). Efficient implementation of Enhanced Adaptive Simultaneous Perturbation Algorithms. CISS 2016, pdf [2] https://en.wikipedia.org/wiki/Simultaneous_perturbation_stochastic_approximation [3] https://www.chessprogramming.org/SPSA [4] https://github.com/se4u/FASPSA/blob/master/src/SPSA.m
opened by se4u 16
Fix random state for test_noisy_artificial_function_loss
Types of changes

[ ] Docs change / refactoring / dependency upgrade

[x] Bug fix (non-breaking change which fixes an issue)

[ ] New feature (non-breaking change which adds functionality)

[ ] Breaking change (fix or feature that would cause existing functionality to change)

Motivation and Context / Related issue

Related issue: https://github.com/facebookresearch/nevergrad/issues/966 The goal is to resolve the bugs in random_state management for test_noisy_artificial_function_loss which causes reproducibility issues. If you want to patch the current system immediately, I can add back np.random.seed(seed) to the test case.

Checklist

[x] The documentation is up-to-date with the changes I made.

[x] I have read the CONTRIBUTING document and completed the CLA (see CLA).

[ ] All tests passed, and additional code has been covered with new tests.

CLA Signed
opened by pacowong 15
Ability to tune the underlying bayes_opt params

Some problems require tuning the underlying bayes_opt params such as the utility function being used or even the underlying gp params ... it seems that there is no way to change them using nevergrad

opened by robert1826 15
Maint: Hypervolume (pyhv.py) module rewrite
Types of changes

[x] Docs change / refactoring / dependency upgrade

[ ] Bug fix (non-breaking change which fixes an issue)

[ ] New feature (non-breaking change which adds functionality)

[ ] Breaking change (fix or feature that would cause existing functionality to change)

Motivation and Context / Related issue

This PR is related to the #366.

This PR rewrites the code that generates the hypervolume indicator, used for multiobjective optimization.

Any suggestions / comments / requests are very welcome.

The purpose is two-fold:

Rewrite / refactor the existing code to improve the general readability and extendability, spot possible bugs, and add documentation

Get rid of the LGPL licence in favour of the standard MIT licence, used by nevergrad.

How Has This Been Tested (if it applies)

We added unit tests for the data structure that are used by the HypervolumeIndicator class.

We test the main functionality of the new hypervolume algorithm against the old one. More integration tests should be included before this PR can be accepted.

Checklist

[x] The documentation is up-to-date with the changes I made.

[x] I have read the CONTRIBUTING document and completed the CLA (see CONTRIBUTING).

[x] All tests passed, and additional code has been covered with new tests.

CLA Signed
opened by Corwinpro 14
Add initial pyomo support
Types of changes

[ ] Docs change / refactoring / dependency upgrade

[ ] Bug fix (non-breaking change which fixes an issue)

[x] New feature (non-breaking change which adds functionality)

[ ] Breaking change (fix or feature that would cause existing functionality to change)

Motivation and Context / Related issue

Pyomo supports the formulation and analysis of mathematical models for complex optimization applications. It provides lots of models for benchmarking. Two Pyomo examples (diet and maxflow) are added in nevergrad/examples/pyomogallery/

https://github.com/facebookresearch/nevergrad/issues/738

How Has This Been Tested (if it applies)

A unit test is prepared in nevergrad/functions/pyomo/test_core.py Pyomo model of a 2-variable minimization problem is defined. There are four test cases currently.

Checklist

[x] The documentation is up-to-date with the changes I made.

[x] I have read the CONTRIBUTING document and completed the CLA (see CLA).

[ ] All tests passed, and additional code has been covered with new tests.

Type: Enhancement CLA Signed Priority: Medium Difficulty: High
opened by pacowong 13

Parallel Ask-and-Tell

Is it possible to use the ask-and-tell framework in parallel? I can't find examples. I have something as follows (not the complete code):

def ask_and_tell(optimizer, func):
    x = optimizer.ask()
    loss = func(*x.args)
    optimizer.tell(x, loss)

    return x, loss


def run_Optimizer(optimizer, func, max_time=None):
    with multiprocessing.Pool(processes=multiprocessing.cpu_count() - 1) as pool:
        all_args = [(optimizer, func) for _ in range(optimizer.budget)]
        results = sorted(pool.starmap_async(ask_and_tell, all_args).get(max_time), key=lambda r: r[1])
        best_x = results[0][0]
        best_loss = results[0][1]

    return {"x": best_x, "fun": best_loss}

But I'm getting an error saying a local object can't be pickled. Am I doing it right? Or is the problem somewhere in my code?

opened by bacalfa 13

Performance issue in optimization.utils.Pruning.call

Steps to reproduce

Benchmark used:

@registry.register
def illcond_prof(seed: tp.Optional[int] = None) -> tp.Iterator[Experiment]:
    """All optimizers on ill cond problems"""
    seedg = create_seed_generator(seed)
    for budget in [200, 30000]:
        for optim in ["CMA"]:            
            for rotation in [True]:
                for name in ["cigar"]:
                    function = ArtificialFunction(name=name, rotation=rotation, block_dimension=50)
                    yield Experiment(function, optim, budget=budget, seed=next(seedg))

Execute the profiler code below
Modify https://github.com/facebookresearch/nevergrad/blob/294aed2253050a3cc4099b70ca0883588a5582d3/nevergrad/optimization/utils.py#L269-L273 as follows:

new_archive.bytesdict = {}
# new_archive.bytesdict = { 
#     b: v 
#     for b, v in archive.bytesdict.items() 
#     if any(v.get_estimation(n) <= quantiles[n] for n in names)

Execute the profiler again and compare the results

Observed Results

There is a severe performance issue related to new_archive.bytesdict. If you increase the second budget in 'for budget in [200, 30000]:' the execution time unrelated to the optimizer grows further.

Expected Results

Expected is that execution time is dominated by the cost to execute the optimizer.

Profiler Code

import cProfile
import io
import pstats
import nevergrad.benchmark.__main__ as main

if __name__ == '__main__':
    pr = cProfile.Profile()
    pr.enable()
    main.repeated_launch('illcond_prof',num_workers=1)
    pr.disable()
    s = io.StringIO()
    sortby = pstats.SortKey.CUMULATIVE
    ps = pstats.Stats(pr, stream=s).sort_stats(sortby)
    ps.print_stats()
    print(s.getvalue())

opened by dietmarwo 11

Update install to large resource class in config.yml
Types of changes

[ ] Docs change / refactoring / dependency upgrade

[ ] Bug fix (non-breaking change which fixes an issue)

[ ] New feature (non-breaking change which adds functionality)

[ ] Breaking change (fix or feature that would cause existing functionality to change)

Motivation and Context / Related issue

How Has This Been Tested (if it applies)

Checklist

[ ] The documentation is up-to-date with the changes I made.

[ ] I have read the CONTRIBUTING document and completed the CLA (see CLA).

[ ] All tests passed, and additional code has been covered with new tests.

CLA Signed
opened by teytaud 0
yet another issue in our CI, related to caching
Types of changes

[ ] Docs change / refactoring / dependency upgrade

[ ] Bug fix (non-breaking change which fixes an issue)

[ ] New feature (non-breaking change which adds functionality)

[ ] Breaking change (fix or feature that would cause existing functionality to change)

Motivation and Context / Related issue

How Has This Been Tested (if it applies)

Checklist

[ ] The documentation is up-to-date with the changes I made.

[ ] I have read the CONTRIBUTING document and completed the CLA (see CLA).

[ ] All tests passed, and additional code has been covered with new tests.

CLA Signed
opened by teytaud 0
Adding climate change in Nevergrad/PCSE
Types of changes

[ ] Docs change / refactoring / dependency upgrade

[ ] Bug fix (non-breaking change which fixes an issue)

[ ] New feature (non-breaking change which adds functionality)

[ ] Breaking change (fix or feature that would cause existing functionality to change)

Motivation and Context / Related issue

How Has This Been Tested (if it applies)

Checklist

[ ] The documentation is up-to-date with the changes I made.

[ ] I have read the CONTRIBUTING document and completed the CLA (see CLA).

[ ] All tests passed, and additional code has been covered with new tests.

CLA Signed
opened by teytaud 0
Adding relative improvement as stopping criterion
Types of changes

[ ] Docs change / refactoring / dependency upgrade

[ ] Bug fix (non-breaking change which fixes an issue)

[x ] New feature (non-breaking change which adds functionality)

[ ] Breaking change (fix or feature that would cause existing functionality to change)

Motivation and Context / Related issue
Adding the relative improvement as a stopping criterion callback to the ```base.Optimizer.minimize``` method #589
How Has This Been Tested (if it applies)
i took some parts from the script ```test_callbacks.py``` and run some tests with different thresholds, the thing seems to work properly. After i added the test function in an analogous format as the one done for ```test_duration_criterion```
Checklist

[ ] The documentation is up-to-date with the changes I made.

[ ] I have read the CONTRIBUTING document and completed the CLA (see CLA).

[x ] All tests passed, and additional code has been covered with new tests.

CLA Signed
opened by lolloconsoli 7
Bounds check error in log scalar
Hi, I have an instrumentation that contains a log scalar. Some samples that it generates trigger an out of bounds error when trying to spawn a child off its values.

Steps to reproduce

I've contrived this minimum example to highlight the issue: using: python 3.10.4 on linux, nevergrad 0.5.0

import nevergrad as ng param = ng.p.Log(init=1e-5, lower=1e-5, upper=1e1) # This does not trigger an out-of-bounds error param.spawn_child(1e-5) # neither does this print(f"{param.value=}") param.spawn_child(param.value) # but this does

Observed Results

Here's the output & traceback:

param.value=9.999999999999989e-06 Traceback (most recent call last): File "/home/agimg/projects/fastpbrl/scripts/scratchpads/nevergrad_bounds.py", line 6, in param.spawn_child(param.value) File "/home/agimg/anaconda3/envs/pbrl/lib/python3.10/site-packages/nevergrad/parametrization/core.py", line 348, in spawn_child child.value = new_value File "/home/agimg/anaconda3/envs/pbrl/lib/python3.10/site-packages/nevergrad/parametrization/_layering.py", line 190, in set obj._layers[-1]._layered_set_value(value) File "/home/agimg/anaconda3/envs/pbrl/lib/python3.10/site-packages/nevergrad/parametrization/_layering.py", line 216, in _layered_set_value super()._layered_set_value(np.array([value], dtype=float)) File "/home/agimg/anaconda3/envs/pbrl/lib/python3.10/site-packages/nevergrad/parametrization/_layering.py", line 96, in _layered_set_value return self._call_deeper("_layered_set_value", value) File "/home/agimg/anaconda3/envs/pbrl/lib/python3.10/site-packages/nevergrad/parametrization/_layering.py", line 84, in _call_deeper return func(*args, **kwargs) File "/home/agimg/anaconda3/envs/pbrl/lib/python3.10/site-packages/nevergrad/parametrization/_layering.py", line 227, in _layered_set_value super()._layered_set_value(np.asarray(value)) File "/home/agimg/anaconda3/envs/pbrl/lib/python3.10/site-packages/nevergrad/parametrization/_layering.py", line 96, in _layered_set_value return self._call_deeper("_layered_set_value", value) File "/home/agimg/anaconda3/envs/pbrl/lib/python3.10/site-packages/nevergrad/parametrization/_layering.py", line 84, in _call_deeper return func(*args, **kwargs) File "/home/agimg/anaconda3/envs/pbrl/lib/python3.10/site-packages/nevergrad/parametrization/_datalayers.py", line 172, in _layered_set_value super()._layered_set_value(self.backward(value)) File "/home/agimg/anaconda3/envs/pbrl/lib/python3.10/site-packages/nevergrad/parametrization/_layering.py", line 96, in _layered_set_value return self._call_deeper("_layered_set_value", value) File "/home/agimg/anaconda3/envs/pbrl/lib/python3.10/site-packages/nevergrad/parametrization/_layering.py", line 84, in _call_deeper return func(*args, **kwargs) File "/home/agimg/anaconda3/envs/pbrl/lib/python3.10/site-packages/nevergrad/parametrization/_datalayers.py", line 279, in _layered_set_value super()._layered_set_value(self._transform.backward(value)) File "/home/agimg/anaconda3/envs/pbrl/lib/python3.10/site-packages/nevergrad/parametrization/transforms.py", line 220, in backward raise ValueError( ValueError: Only data between [-5.] and [1.] can be transformed back. Got: [-5.]

Expected Results

This is a contrived example to highlight the issue. In my real code, the 1e-5 isn't an initial parameter set by me, it's randomly produced by a sample() call to the log dist. While this behavior is technically correct because param.value=9.999999999999989e-06 and is indeed out of bounds, it doesn't make sense that the user should have boiler plate code to check that values we pass to initialize a parameter are valid if they're sampled from the same parameter

Relevant Code

See steps to reproduce
opened by llucid-97 0

How to perform an equality constraint

I was wondering how to perform an equality constrain by using nevergrad:

the objective is defined as:

def objective(X):
  t1 = np.sum(X*price_list)
  t2 = reg.predict(np.append(feature_mean, [np.sum(X[:10]), np.sum(X[10:])]).reshape(1,-1))

  return t1 + t2

where reg is a linear regression function, price_list is array of shape (20,) I want to minimize this objective function.

let's say I have a array of shape (20,) to optimize, it has lower bound of 0, and upper bound of 100.

instrum = ng.p.Instrumentation(
    ng.p.Array(shape=(20,)).set_bounds(lower=0, upper=100),
)

The constraint is:

optimizer.parametrization.register_cheap_constraint(lambda X: np.sum(X[:10])*chemi_list[0] + np.sum(X[10:])*chemi_list[1] - chemi_to_achieve)`

which means the first 10 elements of array times the first array of chemi_list plus last 10 elements of array times the second array of chemi_list0 minus chemi_to_achieve should equal to 0.

the all constants arrays look like below:

price_list = np.array([0.28, 0.27, 0.25, 0.27, 0.28, 0.26, 0.28, 0.28, 0.27, 0.29, 0.32,
       0.28, 0.28, 0.31, 0.32, 0.29, 0.3 , 0.29, 0.33, 0.29])

chemi_list = np.array([[8.200e-02, 5.700e-03, 0.000e+00, 9.000e-05, 3.700e-04, 0.000e+00,
        6.475e-01, 0.000e+00, 0.000e+00, 0.000e+00, 0.000e+00, 0.000e+00,
        0.000e+00, 0.000e+00, 0.000e+00],
       [1.400e-03, 7.520e-01, 0.000e+00, 0.000e+00, 0.000e+00, 0.000e+00,
        0.000e+00, 0.000e+00, 0.000e+00, 0.000e+00, 0.000e+00, 0.000e+00,
        1.400e-02, 0.000e+00, 0.000e+00]])

chemi_to_achieve = np.array([15, 3, 0, 3.5, 0, 0, 120, 8, 8, 0, 0, 0, 0, 0, 0])

But I kept getting error

from concurrent import futures

with futures.ThreadPoolExecutor(max_workers=optimizer.num_workers) as executor:
  recommendation = optimizer.minimize(objective, verbosity=2, executor=executor, batch_mode=False)

TypeError: can only concatenate tuple (not "dict") to tuple

can anyone tell me how to solve this?

opened by redcican 1

Releases(v0.5.0)

v0.5.0(Dec 6, 2022)

Source code(tar.gz)
Source code(zip)
0.5.0(Mar 8, 2022)

See CHANGELOG for details.
Source code(tar.gz)
Source code(zip)
0.4.3(Jan 28, 2021)

This version provides a few fixes and the new multi-objective API of optimizers (you can now provide a list/array of float to tell directly). This allows fore more efficient multi-objective optimization for some optimizers (DE, NGOpt). Future work will continue to improve multi-objective capacities and aim at improving constraints management.

See CHANGELOG for details.
Source code(tar.gz)
Source code(zip)
0.4.2(Aug 4, 2020)

This version should be robust. Following versions may become more unstable as we will add more native multiobjective optimization as an experimental feature. We also are in the process of simplifying the naming pattern for the "NGO/Shiwa" type optimizers which may cause some changes in the future.

See CHANGELOG for details.
Source code(tar.gz)
Source code(zip)
0.4.1(May 7, 2020)

See CHANGELOG for details.
Source code(tar.gz)
Source code(zip)
v0.4.0(Mar 9, 2020)
This is the final step for creating the new instrumentation/parametrization framework and removing the old one. Learn more on the Facebook user group

Important changes:

the old instrumentation system disappears, all deprecation warnings are removed and are now errors.

archive does not store anymore all evaluated points, for memory reasons.

See CHANGELOG for more details.
Source code(tar.gz)
Source code(zip)
v0.3.2(Feb 5, 2020)
This is the second step to propagate the new instrumentation/parametrization framework. Learn more on the Facebook user group

If you are looking for stability, await for version 0.4.0, but the intermediary releases will help by providing deprecation warnings. In particular here are the important changes for this release:

Fist argument of optimizers is renamed to parametrization instead of instrumentation for consistency (deprecation warning)

Old instrumentation classes now raise deprecation warnings.

create_candidate raises deprecation warnings.

Candidate class is completely removed, and is completely replaced by Parameter

See CHANGELOG for more details. All deprecated code will be removed in the following version (v0.4.0)
Source code(tar.gz)
Source code(zip)
v0.3.1(Jan 23, 2020)

This is the first step to propagate the new instrumentation/parametrization framework. Learn more on the Facebook user group and in the CHANGELOG. If you are looking for stability, await for version 0.4.0, but the intermediary releases will help by providing deprecation warnings.
Source code(tar.gz)
Source code(zip)
v0.3.0(Jan 8, 2020)
This release includes new experiment features such as:

constraint management

multiobjective functions

new optimizers This is a stable release before a transition phase in which we will refactor the instrumentation part of the package, allowing a lot more flexibility. These changes will unfortunately probably break some use code, and we expect a few bugs during the transition period.

See CHANGELOG for more about this release, and checkout Nevergrad users Facebook group for more information about upcoming changes.
Source code(tar.gz)
Source code(zip)
v0.2.3(Sep 2, 2019)

Source code(tar.gz)
Source code(zip)
v0.2.2(Jun 20, 2019)

This release improves reproducibility by providing a random state to each instrumentation, which is used by the optimizers. It also introduces some namespace changes to make code clearer. See the CHANGELOG for more details.
Source code(tar.gz)
Source code(zip)
v0.2.1(May 16, 2019)

This release fixes bugs and adds minor improvements. See CHANGELOG for more information.
Source code(tar.gz)
Source code(zip)
v0.2.0(Apr 11, 2019)
This release makes major API changes. Most noticeably:

first parameter of optimizers is now instrumentation instead of dimension. This allows the optimizer to have information on the underlying structure. ints are still allowed as before and will set the instrumentation to the Instrumentation(var.Array(n)) (which is basically the identity).

ask() and provide_recommendation() now return a Candidate with attributes args, kwargs (depending on the instrumentation) and data (the array which was formerly returned). tell must now receive this candidate as well instead of the array.

More details can be found in the CHANGELOG and in the documentation.
Source code(tar.gz)
Source code(zip)
v0.1.6(Mar 15, 2019)

This fixes a bug in PSO introduced by v0.1.5. This is also the last release with the BaseFunction class which will disappear in favor of InstrumentedFunction (breaking change for custom benchmark functions).
Source code(tar.gz)
Source code(zip)
v0.1.5(Mar 7, 2019)

See CHANGELOG for more information
Source code(tar.gz)
Source code(zip)
v0.1.3(Jan 28, 2019)

See CHANGELOGS for an overview of updates since the initial release.
Source code(tar.gz)
Source code(zip)
v0.1.2(Jan 25, 2019)

Updated Pypi with current version (bug fixes, new features). A new version v0.1.3 will follow soon with more information.
Source code(tar.gz)
Source code(zip)
v0.1.1(Jan 8, 2019)

This release contains minor updates and the version matches pypi version.
Source code(tar.gz)
Source code(zip)
v0.1.0(Dec 19, 2018)

Source code(tar.gz)
Source code(zip)

Owner

Meta Research

GitHub https://facebookresearch.github.io/nevergrad/

Scalable, Portable and Distributed Gradient Boosting (GBDT, GBRT or GBM) Library, for Python, R, Java, Scala, C++ and more. Runs on single machine, Hadoop, Spark, Dask, Flink and DataFlow

eXtreme Gradient Boosting Community | Documentation | Resources | Contributors | Release Notes XGBoost is an optimized distributed gradient boosting l

Distributed (Deep) Machine Learning Community

23.6k Jan 3, 2023

A fast, distributed, high performance gradient boosting (GBT, GBDT, GBRT, GBM or MART) framework based on decision tree algorithms, used for ranking, classification and many other machine learning tasks.

Light Gradient Boosting Machine LightGBM is a gradient boosting framework that uses tree based learning algorithms. It is designed to be distributed a

14.5k Jan 7, 2023

A fast, scalable, high performance Gradient Boosting on Decision Trees library, used for ranking, classification, regression and other machine learning tasks for Python, R, Java, C++. Supports computation on CPU and GPU.

Website | Documentation | Tutorials | Installation | Release Notes CatBoost is a machine learning method based on gradient boosting over decision tree

6.9k Jan 5, 2023

Iterative stochastic gradient descent (SGD) linear regressor with regularization

SGD-Linear-Regressor Iterative stochastic gradient descent (SGD) linear regressor with regularization Dataset: Kaggle “Graduate Admission 2” https://w

1 Oct 29, 2021

Houseprices - Predict sales prices and practice feature engineering, RFs, and gradient boosting

House Prices - Advanced Regression Techniques Predicting House Prices with Machine Learning This project is build to enhance my knowledge about machin

1 Jan 1, 2022

DeepSpeed is a deep learning optimization library that makes distributed training easy, efficient, and effective.

DeepSpeed is a deep learning optimization library that makes distributed training easy, efficient, and effective. 10x Larger Models 10x Faster Trainin

8.4k Dec 30, 2022

CS 7301: Spring 2021 Course on Advanced Topics in Optimization in Machine Learning

141 Nov 10, 2022

jaxfg - Factor graph-based nonlinear optimization library for JAX.

Factor graphs + nonlinear optimization in JAX

134 Dec 21, 2022

Pyomo is an object-oriented algebraic modeling language in Python for structured optimization problems.

Pyomo is a Python-based open-source software package that supports a diverse set of optimization capabilities for formulating and analyzing optimization models. Pyomo can be used to define symbolic problems, create concrete problem instances, and solve these instances with standard solvers.

1.4k Dec 28, 2022

This is an implementation of the proximal policy optimization algorithm for the C++ API of Pytorch

This is an implementation of the proximal policy optimization algorithm for the C++ API of Pytorch. It uses a simple TestEnvironment to test the algorithm

59 Dec 9, 2022

Bayesian optimization based on Gaussian processes (BO-GP) for CFD simulations.

BO-GP Bayesian optimization based on Gaussian processes (BO-GP) for CFD simulations. The BO-GP codes are developed using GPy and GPyOpt. The optimizer

8 Mar 31, 2022

A Python step-by-step primer for Machine Learning and Optimization

early-ML Presentation General Machine Learning tutorials A Python step-by-step primer for Machine Learning and Optimization This github repository gat

8 Dec 1, 2022

Tools for mathematical optimization region

15 Nov 30, 2022

Implementation of linesearch Optimization Algorithms in Python

Nonlinear Optimization Algorithms During my time as Scientific Assistant at the Karlsruhe Institute of Technology (Germany) I implemented various Opti

3 Dec 6, 2022

The Ultimate FREE Machine Learning Study Plan

2.5k Jan 5, 2023

a distributed deep learning platform

Apache SINGA Distributed deep learning system http://singa.apache.org Quick Start Installation Examples Issues JIRA tickets Code Analysis: Mailing Lis

2.7k Jan 5, 2023

AutoOED: Automated Optimal Experiment Design Platform

AutoOED is an optimal experiment design platform powered with automated machine learning to accelerate the discovery of optimal solutions. Our platform solves multi-objective optimization problems and automatically guides the design of experiment to be evaluated.

107 Jan 3, 2023

Apache Liminal is an end-to-end platform for data engineers & scientists, allowing them to build, train and deploy machine learning models in a robust and agile way

Apache Liminals goal is to operationalise the machine learning process, allowing data scientists to quickly transition from a successful experiment to an automated pipeline of model training, validation, deployment and inference in production. Liminal provides a Domain Specific Language to build ML workflows on top of Apache Airflow.

121 Dec 28, 2022

Nevergrad - A gradient-free optimization platform

Related tags

Overview

Nevergrad - A gradient-free optimization platform

Documentation

Citing

License

Comments

Types of changes

Motivation and Context / Related issue

How Has This Been Tested (if it applies)

Checklist

Types of changes

Motivation and Context / Related issue

How Has This Been Tested (if it applies)

Checklist

Types of changes

Motivation and Context / Related issue

How Has This Been Tested

Things to Review

Checklist

Types of changes

Motivation and Context / Related issue

How Has This Been Tested (if it applies)

Checklist

Types of changes

Motivation and Context / Related issue

Checklist

Types of changes

Motivation and Context / Related issue

How Has This Been Tested (if it applies)

Checklist

Types of changes

Motivation and Context / Related issue

How Has This Been Tested (if it applies)

Checklist

Steps to reproduce

Observed Results

Expected Results

Profiler Code

Types of changes

Motivation and Context / Related issue

How Has This Been Tested (if it applies)

Checklist

Types of changes

Motivation and Context / Related issue

How Has This Been Tested (if it applies)

Checklist

Types of changes

Motivation and Context / Related issue

How Has This Been Tested (if it applies)

Checklist

Types of changes

Motivation and Context / Related issue

How Has This Been Tested (if it applies)

Checklist

Steps to reproduce

Observed Results

Expected Results

Relevant Code

Releases(v0.5.0)

v0.5.0(Dec 6, 2022)

0.5.0(Mar 8, 2022)

0.4.3(Jan 28, 2021)

0.4.2(Aug 4, 2020)

0.4.1(May 7, 2020)

v0.4.0(Mar 9, 2020)

v0.3.2(Feb 5, 2020)

v0.3.1(Jan 23, 2020)

v0.3.0(Jan 8, 2020)

v0.2.3(Sep 2, 2019)

v0.2.2(Jun 20, 2019)

v0.2.1(May 16, 2019)

v0.2.0(Apr 11, 2019)

v0.1.6(Mar 15, 2019)

v0.1.5(Mar 7, 2019)

v0.1.3(Jan 28, 2019)

v0.1.2(Jan 25, 2019)

v0.1.1(Jan 8, 2019)

v0.1.0(Dec 19, 2018)