Installing nevergrad lists bayes_opt and om.algos.nevergrad_bo as available but they do not work without extra dependencies

[lesshaste]

Bug description

If you pip install nevergrad and look at the list of available optimizers given by optimagic, it includes bayes_opt and om.algos.nevergrad_bo. But these are not in fact usable without extra dependencies.

To Reproduce

I made a new clean python virtual environment and did pip install nevergrad optimagic

I then ran:

import optimagic as om
import numpy as np

def three_minima_function(params):
    x, y = params
    res =  -2 * np.exp(-((x - 1) ** 2 + (y - 1) ** 2) / 0.2) \
        - 1.5 * np.exp(-((x + 1) ** 2 + (y + 1) ** 2) / 0.5) \
        - 1 * np.exp(-((x - 1) ** 2 + (y + 1) ** 2) / 1)
    res = np.ceil(res * 100) / 100
    return res


list_of_algos = om.algos.Scalar.Global.GradientFree.Bounded.Available


bounds = om.Bounds(lower=np.array([-5, -5]), upper=np.array([5, 5]))

algo = om.algos.nevergrad_bo  # Also fails with om.algos.bayes_opt
res=om.minimize(
    fun=three_minima_function,
    params=np.arange(2),
    algorithm=algo,
    bounds=bounds,
    algo_options={"stopping_maxfun": 50, "stopping_maxiter": 50},
)
print(algo, three_minima_function(res.x), res.nfev)

This code fails with:

from bayes_optim import RealSpace
ModuleNotFoundError: No module named 'bayes_optim'

If you try:

om.algos.bayes_opt

instead it fails with:

from bayes_opt import acquisition
ImportError: cannot import name 'acquisition' from 'bayes_opt' (/home/raph/python/myenv-opti/lib/python3.12/site-packages/bayes_opt/__init__.py)

Expected behavior

bayes_opt should be not offered when nevergrad is installed but a recent bayesian-optimization version is not. om.algos.nevergrad_bo should not be offered when bayes_optim is not installed.

You can get bayes_opt to work by pip install bayesian-optimization -U but that uninstalls nevergrad due to a version conflict. You can get om.algos.nevergrad_bo to work by pip install bayes_optim (this installs a number of huge torch/cuda related modules even if you don't have a GPU. There is a cpu only version of torch and the install is much faster if you do pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu first).

System

Ubuntu 24.04.2 LTS

[timmens]

Hey! Thanks for raising the issue.

This is definitely a bug. We are not entirely happy with the dependency chaos regarding nevergrad and bayesian-optimization ourselves. We have opened an issue for nevergrad, hoping that we can relax their bayesian-optimization version restriction in the future.

For the current situation:

We set an algorithm (e.g., "nevergrad_pso") as "available" if the corresponding library is installed, e.g., IS_NEVERGRAD_INSTALLED = _is_installed("nevergrad"), where

import importlib.util
def _is_installed(module_name: str) -> bool:
    """Return True if the given module is installed, otherwise False."""
    return importlib.util.find_spec(module_name) is not None

If I understand the problem correctly, we must

1. Update IS_BAYESOPT_INSTALLED to reflect that we require version >=2
2. Create IS_BAYESOPTIM_INSTALLED to check whether bayes-optim is installed. Then update the is_available field of "nevergrad_bo" to IS_NEVERGRAD_INSTALLED and IS_BAYESOPTIM_INSTALLED.
3. Update the documentation of these libraries to (1) mention the version incompatibility and (2) mention for "nevergrad_bo" that bayes-optim needs to be installed as well.

Please let me know if I am missing something.

---

@lesshaste do you think this is something you could fix in a PR?

[gauravmanmode]

Hi @lesshaste, are you up for this?
Otherwise, I would be happy to take this and will the update nevergrad docs.

[lesshaste]

I would love to but I don't realistically have the time at the moment.

[janosg]

Thanks for taking over @gauravmanmode !