Contents

Tune: Scalable Hyperparameter Tuning

../_images/tune_overview.png

Tune is a Python library for experiment execution and hyperparameter tuning at any scale. You can tune your favorite machine learning framework (PyTorch, XGBoost, Scikit-Learn, TensorFlow and Keras, and more) by running state of the art algorithms such as Population Based Training (PBT) and HyperBand/ASHA. Tune further integrates with a wide range of additional hyperparameter optimization tools, including Ax, BayesOpt, BOHB, Dragonfly, FLAML, HEBO, Hyperopt, Nevergrad, Optuna, SigOpt, skopt, and ZOOpt.

Click on the following tabs to see code examples for various machine learning frameworks:

To run this example, install the following: pip install "ray[tune]".

In this quick-start example you minimize a simple function of the form f(x) = a**2 + b, our objective function. The closer a is to zero and the smaller b is, the smaller the total value of f(x). We will define a so-called search space for a and b and let Ray Tune explore the space for good values.

from ray import tune

# 1. Define an objective function.
def objective(config):
    score = config["a"] ** 2 + config["b"]
    return {"score": score}


# 2. Define a search space.
search_space = {
    "a": tune.grid_search([0.001, 0.01, 0.1, 1.0]),
    "b": tune.choice([1, 2, 3]),
}

# 3. Start a Tune run and print the best result.
analysis = tune.run(objective, config=search_space)
print(analysis.get_best_config(metric="score", mode="min"))

To tune your Keras models with Hyperopt, you wrap your model in an objective function whose config you can access for selecting hyperparameters. In the example below we only tune the activation parameter of the first layer of the model, but you can tune any parameter of the model you want. After defining the search space, you can simply initialize the HyperOptSearch object and pass it to run. It’s important to tell Ray Tune which metric you want to optimize and whether you want to maximize or minimize it.

from ray import tune
from ray.tune.search.hyperopt import HyperOptSearch
import keras


# 1. Wrap a Keras model in an objective function.
def objective(config):
    model = keras.models.Sequential()
    model.add(keras.layers.Dense(784, activation=config["activation"]))
    model.add(keras.layers.Dense(10, activation="softmax"))

    model.compile(loss="binary_crossentropy", optimizer="adam", metrics=["accuracy"])
    # model.fit(...)
    # loss, accuracy = model.evaluate(...)
    return {"accuracy": accuracy}


# 2. Define a search space and initialize the search algorithm.
search_space = {"activation": tune.choice(["relu", "tanh"])}
algo = HyperOptSearch()

# 3. Start a Tune run that maximizes accuracy.
analysis = tune.run(
    objective, search_alg=algo, config=search_space, metric="accuracy", mode="max"
)

To tune your PyTorch models with Optuna, you wrap your model in an objective function whose config you can access for selecting hyperparameters. In the example below we only tune the momentum and learning rate (lr) parameters of the model’s optimizer, but you can tune any other model parameter you want. After defining the search space, you can simply initialize the OptunaSearch object and pass it to run. It’s important to tell Ray Tune which metric you want to optimize and whether you want to maximize or minimize it. We stop tuning this training run after 5 iterations, but you can easily define other stopping rules as well.

# 1. Wrap your PyTorch model in an objective function.
import torch
from ray import tune
from ray.air import session
from ray.tune.search.optuna import OptunaSearch


# 1. Wrap a PyTorch model in an objective function.
def objective(config):
    train_loader, test_loader = load_data()  # Load some data
    model = ConvNet().to("cpu")  # Create a PyTorch conv net
    optimizer = torch.optim.SGD(  # Tune the optimizer
        model.parameters(), lr=config["lr"], momentum=config["momentum"]
    )

    while True:
        train(model, optimizer, train_loader)  # Train the model
        acc = test(model, test_loader)  # Compute test accuracy
        session.report({"mean_accuracy": acc})  # Report to Tune


# 2. Define a search space and initialize the search algorithm.
search_space = {"lr": tune.loguniform(1e-4, 1e-2), "momentum": tune.uniform(0.1, 0.9)}
algo = OptunaSearch()

# 3. Start a Tune run that maximizes mean accuracy and stops after 5 iterations.
analysis = tune.run(
    objective,
    metric="mean_accuracy",
    mode="max",
    search_alg=algo,
    stop={"training_iteration": 5},
    config=search_space,
)
print("Best config is:", analysis.best_config)

With Tune you can also launch a multi-node distributed hyperparameter sweep in less than 10 lines of code. It automatically manages checkpoints and logging to TensorBoard. And you can move your models from training to serving on the same infrastructure with Ray Serve.

Getting Started

In our getting started tutorial you will learn how to tune a PyTorch model effectively with Tune.

Get Started with Tune

Key Concepts

Understand the key concepts behind Ray Tune. Learn about tune runs, search algorithms, schedulers and other features.

Tune’s Key Concepts

User Guides

Our guides teach you about key features of Tune, such as distributed training or early stopping.

Learn How To Use Tune

Examples

In our examples you can find practical tutorials for using frameworks such as scikit-learn, Keras, TensorFlow, PyTorch, and mlflow, and state of the art search algorithm integrations.

Ray Tune Examples

Ray Tune FAQ

Find answers to commonly asked questions in our detailed FAQ.

Ray Tune FAQ

Ray Tune API

Get more in-depth information about the Ray Tune API, including all about search spaces, algorithms and training configurations.

Read the API Reference

Why choose Tune?

There are many other hyperparameter optimization libraries out there. If you’re new to Tune, you’re probably wondering, “what makes Tune different?”

Cutting-Edge Optimization Algorithms

As a user, you’re probably looking into hyperparameter optimization because you want to quickly increase your model performance.

Tune enables you to leverage a variety of these cutting edge optimization algorithms, reducing the cost of tuning by terminating bad runs early, choosing better parameters to evaluate, or even changing the hyperparameters during training to optimize schedules.

First-class Developer Productivity

A key problem with many hyperparameter optimization frameworks is the need to restructure your code to fit the framework. With Tune, you can optimize your model just by adding a few code snippets.

Also, Tune removes boilerplate from your code training workflow, automatically manages checkpoints and logs results to tools such as MLflow and TensorBoard, while also being highly customizable.

Multi-GPU & Distributed Training Out Of The Box

Hyperparameter tuning is known to be highly time-consuming, so it is often necessary to parallelize this process. Most other tuning frameworks require you to implement your own multi-process framework or build your own distributed system to speed up hyperparameter tuning.

However, Tune allows you to transparently parallelize across multiple GPUs and multiple nodes. Tune even has seamless fault tolerance and cloud support, allowing you to scale up your hyperparameter search by 100x while reducing costs by up to 10x by using cheap preemptible instances.

Coming From Another Hyperparameter Optimization Tool?

You might be already using an existing hyperparameter tuning tool such as HyperOpt or Bayesian Optimization.

In this situation, Tune actually allows you to power up your existing workflow. Tune’s Search Algorithms integrate with a variety of popular hyperparameter tuning libraries (see examples) and allow you to seamlessly scale up your optimization process - without sacrificing performance.

Projects using Tune

Here are some of the popular open source repositories and research projects that leverage Tune. Feel free to submit a pull-request adding (or requesting a removal!) of a listed project.

  • Softlearning: Softlearning is a reinforcement learning framework for training maximum entropy policies in continuous domains. Includes the official implementation of the Soft Actor-Critic algorithm.

  • Flambe: An ML framework to accelerate research and its path to production. See flambe.ai.

  • Population Based Augmentation: Population Based Augmentation (PBA) is a algorithm that quickly and efficiently learns data augmentation functions for neural network training. PBA matches state-of-the-art results on CIFAR with one thousand times less compute.

  • Fast AutoAugment by Kakao: Fast AutoAugment (Accepted at NeurIPS 2019) learns augmentation policies using a more efficient search strategy based on density matching.

  • Allentune: Hyperparameter Search for AllenNLP from AllenAI.

  • machinable: A modular configuration system for machine learning research. See machinable.org.

  • NeuroCard: NeuroCard (Accepted at VLDB 2021) is a neural cardinality estimator for multi-table join queries. It uses state of the art deep density models to learn correlations across relational database tables.

Citing Tune

If Tune helps you in your academic research, you are encouraged to cite our paper. Here is an example bibtex:

@article{liaw2018tune,
    title={Tune: A Research Platform for Distributed Model Selection and Training},
    author={Liaw, Richard and Liang, Eric and Nishihara, Robert
            and Moritz, Philipp and Gonzalez, Joseph E and Stoica, Ion},
    journal={arXiv preprint arXiv:1807.05118},
    year={2018}
}

previous

Ray Train API

next

Getting Started