Running Tune experiments with Nevergrad

In this tutorial we introduce Nevergrad, while running a simple Ray Tune experiment. Tune’s Search Algorithms integrate with Nevergrad and, as a result, allow you to seamlessly scale up a Nevergrad optimization process - without sacrificing performance.

Nevergrad provides gradient/derivative-free optimization able to handle noise over the objective landscape, including evolutionary, bandit, and Bayesian optimization algorithms. Nevergrad internally supports search spaces which are continuous, discrete or a mixture of thereof. It also provides a library of functions on which to test the optimization algorithms and compare with other benchmarks.

In this example we minimize a simple objective to briefly demonstrate the usage of Nevergrad with Ray Tune via NevergradSearch. It's useful to keep in mind that despite the emphasis on machine learning experiments, Ray Tune optimizes any implicit or explicit objective. Here we assume nevergrad==0.4.3.post7 library is installed. To learn more, please refer to Nevergrad website.

Necessary requirements:

• pip install ray[tune] nevergrad==0.4.3.post7

# !pip install ray[tune]
!pip install nevergrad==0.4.3.post7 

Click below to see all the imports we need for this example.

import time

import ray
import nevergrad as ng
from ray import tune
from ray.tune.search import ConcurrencyLimiter
from ray.tune.search.nevergrad import NevergradSearch

Let's start by defining a simple evaluation function. We artificially sleep for a bit (0.1 seconds) to simulate a long-running ML experiment. This setup assumes that we're running multiple steps of an experiment and try to tune two hyperparameters, namely width and height, and activation.

def evaluate(step, width, height, activation):
    time.sleep(0.1)
    activation_boost = 10 if activation=="relu" else 1
    return (0.1 + width * step / 100) ** (-1) + height * 0.1 + activation_boost

Next, our objective function takes a Tune config, evaluates the score of your experiment in a training loop, and uses tune.report to report the score back to Tune.

def objective(config):
    for step in range(config["steps"]):
        score = evaluate(step, config["width"], config["height"], config["activation"])
        tune.report({"iterations": step, "mean_loss": score})

ray.init(configure_logging=False)

Now we construct the hyperparameter search space using ConfigSpace

Next we define the search algorithm built from NevergradSearch, constrained to a maximum of 4 concurrent trials with a ConcurrencyLimiter. Here we use ng.optimizers.OnePlusOne, a simple evolutionary algorithm.

algo = NevergradSearch(
    optimizer=ng.optimizers.OnePlusOne,
)
algo = tune.search.ConcurrencyLimiter(algo, max_concurrent=4)

The number of samples is the number of hyperparameter combinations that will be tried out. This Tune run is set to 1000 samples. (you can decrease this if it takes too long on your machine).

num_samples = 1000

# If 1000 samples take too long, you can reduce this number.
# We override this number here for our smoke tests.
num_samples = 10

Finally, all that's left is to define a search space.

search_config = {
    "steps": 100,
    "width": tune.uniform(0, 20),
    "height": tune.uniform(-100, 100),
    "activation": tune.choice(["relu, tanh"])
}

Finally, we run the experiment to "min"imize the "mean_loss" of the objective by searching search_space via algo, num_samples times. This previous sentence is fully characterizes the search problem we aim to solve. With this in mind, observe how efficient it is to execute tuner.fit().

tuner = tune.Tuner(
    objective,
    tune_config=tune.TuneConfig(
        metric="mean_loss",
        mode="min",
        search_alg=algo,
        num_samples=num_samples,
    ),
    param_space=search_config,
)
results = tuner.fit()

Here are the hyperparameters found to minimize the mean loss of the defined objective.

print("Best hyperparameters found were: ", results.get_best_result().config)

Optional: passing the (hyper)parameter space into the search algorithm

We can also pass the search space into NevergradSearch using their designed format.

space = ng.p.Dict(
    width=ng.p.Scalar(lower=0, upper=20),
    height=ng.p.Scalar(lower=-100, upper=100),
    activation=ng.p.Choice(choices=["relu", "tanh"])
)

algo = NevergradSearch(
    optimizer=ng.optimizers.OnePlusOne,
    space=space,
    metric="mean_loss",
    mode="min"
)
algo = tune.search.ConcurrencyLimiter(algo, max_concurrent=4)

Again we run the experiment, this time with a less passed via the config and instead passed through search_alg.

tuner = tune.Tuner(
    objective,
    tune_config=tune.TuneConfig(
#         metric="mean_loss",
#         mode="min",
        search_alg=algo,
        num_samples=num_samples,
    ),
    param_space={"steps": 100},
)
results = tuner.fit()

Here are the hyperparameters found to minimize the mean loss of the defined objective. Note that we have to pass the metric and mode here because we don't set it in the TuneConfig.

print("Best hyperparameters found were: ", results.get_best_result("mean_loss", "min").config)

ray.shutdown()