3 min read

Last week, researchers from LG’s Advanced AI team open-sourced Auptimizer. It is a general hyperparameter optimization (HPO) framework to help data scientists speed up machine learning model tuning.

What challenges Auptimizer aims to address

Hyperparameters are adjustable parameters that govern the training process of a machine learning model. These represent important properties of a model, for instance, a penalty in Logistic Regression Classifier or learning rate for training a neural network.

Tuning hyperparameters can often be a very tedious task, especially when the model training is computationally intensive. There are currently both open source and commercial automated HPO solutions like Google AutoML, Amazon SageMaker, and Optunity. However, using them at scale still poses some challenges. In a paper explaining the motivation and system design behind Auptimizer, the team wrote, “But, in all cases, adopting new algorithms or accommodating new computing resources is still challenging.”

To address these challenges and more, the team has come up with Auptimizer with which they aim to automate all the tedious tasks you do when building a machine learning model. This initial open-sourced version provides the following advantages:

Easily switch among different Hyperparameter Optimization algorithms without rewriting the training script

Getting started with Auptimizer only requires adding a few lines of code and it will then guide you to setup all other experiment-related configurations. This enables users to switch among different HPO algorithms and computing resources without rewriting their training script, which is one of the key hurdles in the HPO adoption process. Once set up, it will run and record sophisticated Hyperparameter Optimization experiments for you.

Orchestrating compute resources for faster hyperparameter tuning

Users can specify the resources they want to be used in experiment configurations including processors, graphics chips, nodes, and public cloud instances like Amazon Web Services EC2. Auptimizer keeps track of the resources in a persistent database and queries it to check if the resources specified by the user are available. If the resource is available, it will be taken by Auptimizer for job execution, and if not the system will wait until it is free. Auptimizer is also compatible with existing resource management tools such as Boto 3.

A single interface to various sophisticated HPO algorithms

The current Auptimizer implementation provides a “single seamless access point to top-notch HPO algorithms” such as Spearmint, Hyperopt, Hyperband, BOHB and also supports the simple random search and grid search. Users can also integrate their own proprietary solution and switch between different HPO algorithms with minimal changes to their existing code.

The following table shows the currently supported techniques by Auptimizer:

Source: LG

How Auptimizer works

The following figure shows the system design for Auptimizer:

Auptimizer System Design

Source: LG

The key components of Auptimizer are the Proposer and Resource Manager. The Proposer interface defines two functions: ‘get_param()’ to return the new hyperparameter values and ‘update()’ to update the history. Resource Manager is responsible for automatically connecting compute resources to model training when they are available. Its ‘get_available()’ function acts as the interface between Auptimizer and typical resource management and job scheduling tools. The ‘run()’ function, as the name suggests, executes the provided code.

To enable reproducibility, Auptimizer has a provision for tracking all the experiment history in the user-specified database. Users can also visualize the results from history with a basic visualization tool that comes integrated with Auptimizer. For further analysis, users can directly access the results stored in the database.

Sharing the future vision for Auptimizer the team wrote, “As development progress, Auptimizer will support the end-to-end development cycle for building models for edge devices including robust support for model compression and neural architecture search.”

This article gave you a basic introduction to Auptimizer. Check out the paper, Auptimizer – an Extensible, Open-Source Framework for Hyperparameter Tuning and GitHub repository to know more in detail.

Read Next

Facebook open-sources Hyperparameter autotuning for fastText to automatically find best hyperparameters for your dataset

PyTorch announces the availability of PyTorch Hub for improving machine learning research reproducibility

Training Deep Convolutional GANs to generate Anime Characters [Tutorial]