Syne Tune: Large-Scale and Reproducible Hyperparameter Optimization

Documentation | Blackboxes | Benchmarking | API Reference | PyPI | Latest Blog Post | Discord

Syne Tune is a library for large-scale hyperparameter optimization (HPO) with the following key features:

• State-of-the-art HPO methods for multi-fidelity optimization, multi-objective optimization, transfer learning, and population-based training.

• Tooling that lets you run large-scale experimentation either locally or on SLURM clusters.

• Extensive collection of blackboxes including surrogate and tabular benchmarks for efficient HPO simulation.

Installing

To install Syne Tune from pip:

bash pip install 'syne-tune[extra]'

or to install the latest version from source:

bash git clone https://github.com/awslabs/syne-tune.git cd syne-tune pip install -e '.[extra]'

See our change log to see what changed in the latest version.

Getting started

Syne Tune assumes some python script that given hyperparameter as input arguments trains and validates a machine learning model that somewhat follows this pattern:

```python from argparse import ArgumentParser

if name == 'main': parser = ArgumentParser() parser.addargument('--epochs', type=int) parser.addargument('--hyperparameter1', type=float) parser.addargument('--hyperparameter3', type=float) args, _ = parser.parseknown_args() # instantiate your machine learning model for epoch in range(args.epochs): # training loop # train for some steps or epoch ... # validate your model on some hold-out validation data ```

Step 1: Adapt your training script

First, to enable tuning of your training script, you need to report metrics so they can be communicated to Syne Tune. For example, in the script above, we assume you're tuning two hyperparameters — height and width — to minimize a loss function. To report the loss back to Syne Tune after each epoch, simply add report(epoch=epoch, loss=loss) inside your training loop:

```python

trainheightsimple.py

import logging import time

from syne_tune import Reporter from argparse import ArgumentParser

if name == 'main': root = logging.getLogger() root.setLevel(logging.INFO) parser = ArgumentParser() parser.addargument('--epochs', type=int) parser.addargument('--width', type=float) parser.addargument('--height', type=float) args, _ = parser.parseknownargs() report = Reporter() for step in range(args.epochs): time.sleep(0.1) dummyscore = 1.0 / (0.1 + args.width * step / 100) + args.height * 0.1 # Feed the score back to Syne Tune. report(epoch=step + 1, meanloss=dummyscore) ```

Step 2: Define a launching script

Once the training script is prepared, we first define the search space and then start the tuning process. In this example, we launch ASHA for a total of 30 seconds using four workers. Each worker spawns a separate Python process to evaluate a hyperparameter configuration, meaning that four configurations are trained in parallel.

```python

launchheightsimple.py

from synetune import Tuner, StoppingCriterion from synetune.backend import LocalBackend from synetune.configspace import randint from syne_tune.optimizer.baselines import ASHA

hyperparameter search space to consider

config_space = { 'width': randint(1, 20), 'height': randint(1, 20), 'epochs': 100, }

tuner = Tuner( trialbackend=LocalBackend(entrypoint='trainheightsimple.py'), scheduler=ASHA( configspace, metric='meanloss', timeattr='epoch', ), stopcriterion=StoppingCriterion(maxwallclocktime=30), # total runtime in seconds n_workers=4, # how many trials are evaluated in parallel ) tuner.run() ```

Step 3: Plot the results

Next, we can plot the results as follows. Replace TUNER_NAME with the name of the tuning job used earlier — this is shown at the beginning of the logs.

```python import matplotlib.pyplot as plt from synetune.experiments import loadexperiment

e = loadexperiment('TUNERNAME') # name of the tuning run which is printed at the beginning of the run e.plottrialsovertime(metrictoplot='meanloss') plt.show() ```

Benchmarking

Checkout this tutorial to run large-scale benchmarking with Syne Tune.

Blog Posts

• Run distributed hyperparameter and neural architecture tuning jobs with Syne Tune
• Hyperparameter optimization for fine-tuning pre-trained transformer models from Hugging Face (notebook)
• Learn Amazon Simple Storage Service transfer configuration with Syne Tune (code)

Videos

• Martin Wistuba: Hyperparameter Optimization for the Impatient (PyData 2023)
• David Salinas: Syne Tune: A Library for Large-Scale Hyperparameter Tuning and Reproducible Research (AutoML Seminar)

Contributing

See CONTRIBUTING for more information.

Citing Syne Tune

If you use Syne Tune in a scientific publication, please cite the following paper:

"Syne Tune: A Library for Large Scale Hyperparameter Tuning and Reproducible Research" First Conference on Automated Machine Learning, 2022.

bibtex @inproceedings{ salinas2022syne, title={Syne Tune: A Library for Large Scale Hyperparameter Tuning and Reproducible Research}, author={David Salinas and Matthias Seeger and Aaron Klein and Valerio Perrone and Martin Wistuba and Cedric Archambeau}, booktitle={International Conference on Automated Machine Learning, AutoML 2022}, year={2022}, url={https://proceedings.mlr.press/v188/salinas22a.html} }

License

This project is licensed under the Apache-2.0 License.