• Category
  • >Machine Learning

What is Optuna? Hyperparameters, Approach, and Features

  • Neelam Tyagi
  • Jun 09, 2020
  • Updated on: Jun 03, 2020
What is Optuna? Hyperparameters, Approach, and Features title banner

“The future is ours to shape. I feel we are in a race that we need to win. It’s a race between the growing power of the technology and the growing wisdom we need to manage it.”  -Max Tegmark


This blog introduces a unique devised approach, i.e. Optuna, for next-generation hyperparameter optimization software in machine learning. 


For a brief, Optuna is a framework of hyperparameter optimization, produced for machine learning and scripted in Python. For the purpose to make hyperparameter optimization handier and extensible for the newbies and experienced, Optuna is mainly framed for.





In machine learning, hyperparameters are the fixed parameters that have assigned before training, unidentical to other parameters of a model.


In the deep learning algorithms, the learning rate, batch size, and a number of training iterations are addressed as hyperparameters. They have consisted of multiple neural networks and channels, and not only numerical values but also deploying Momentum SGD or Adam while training is also referred to as hyperparameters. (Related blog: Introduction to Model Hyperparameter and Tuning in Machine Learning)


Hyperparameters are comprised of parts of the training algorithm and would characterize the architecture of a model. It consists of an enhanced pipeline that can impact hugely model efficiency.


Addressing the necessity in this attribute and in terms of analyzing such specific algorithms and models that need tuning precisely, Optuna has characteristics that facilitate this issue by regulating the hyperparameters configuration.


While dealing with enormous machine learning applications or projects, searching for hyperparameter is one of the most ponderous functions. The ramifications of deep learning algorithms are expanding along with its fame, and the self-regulating structure of hyperparameter tuning is high in demand than before. (You can learn more about what are model parameters from our previous blog)


Multiple hyperparameter optimization software such as Hyperopt, Spearmint, SMAC, and Vizier was designed to meet all the requirements.     


Discussing Optuna: A Hyperparameter Optimization Framework


Optuna is an automated hyperparameter optimization software framework that is knowingly invented for the machine learning-based tasks. It emphasizes an authoritative, define-by-run approach user API. 


Due to the define-by-run API, the code script written with Optuna retains extreme modularity, and the Optuna users could actively compose “search spaces” for the hyperparameters.


The software is available at GitHub repository under MIT license that includes a Python code for installation study and trial; this GitHub project has developed to entertain a group of developers via giving state-of-the-art algorithms, and the define-by-run API is suitable for several real-world applications.


If listening to the essential characteristics of an Optuna;


  1. Comparable distributed optimization and scalable computing,

  2. Resourceful, easy-to-setup and adaptable architecture that can be deployed for various purposes,

  3. Efficient implementation of searching and pruning strategies that enable some sort of user-customization,

  4. Eliminating unfavorable experiments, and

  5. Handle light-weighted experiments through an interactive interface.

  6. Conduct heavy-weight distributed computations. 

  7. Personalize visualization of optimization along with a particular function call.  

  8. Integrated modules for various machine learning libraries like TensorFlow, sci-kit learn, XGboost, Keras, etc. 

In the following video, you find Optuna- a next-generation hyperparameter optimization framework with the latest design-criteria: 

What are the key features of Optuna?


Optuna is a software framework for the automated optimization procedure for hyperparameters. For the known fact, it inspects and identifies optimal values of hyperparameter via trial and error method for efficient performance and high efficiency. 

Let’s find how, following the description of key features of Optuna that defines its working;


  1. Define-by-run API


Define-by-run, in deep learning framework, enables the user to execute dynamically deep networks, similar to this definition, in Optuna, define-by-run implies a framework that permits users to build the search space actively in the particular of optimization framework.


In define-by-run API, users don’t need to define everything earlier about an optimization approach that can be understood with actual code easily. Optuna codifies hyperparameter optimization in the context of a method of maximizing or minimizing objective functions that consider a group of hyperparameters as input and respond back validation score as an output.


An objective function formulates the search space of architecture of neural network in the term of the number of layers and the number of hidden units without depending on defined static variables externally. 


Optuna describes each process as;

  • Study: optimization based on an objective function, and
  • Trial: a single evaluation or execution of the objective function.

Displaying key features of Optuna that are Define-by-run API, efficient sampling and pruning mechanism, easy-to-setup, scalable, and versatile system.

Key features of Optuna

  1. Efficient Sampling and Pruning Mechanism


Normally the cost-effectiveness of the hyperparameter optimization framework is measured via the expertise of searching approach to finding the parameters to be evaluated and the efficiency calculation approach to calculate the evaluated parameters from learning curves and identify the parameters to be eliminated.


The eliminating of unfavorable trails is expressed as pruning or automated early stopping.


  1. The sampling method is of two types; (1) the Relational sampling method that handles the interrelationships amid parameters and (2) Independent sampling that samples every parameter individually where the Optuna is efficient for both sampling method. Also, Optuna has a feature that enables users to deploy their own customized sampling method.
  2. Pruning algorithm is crucial to ensure the “cost” factor of the cost-effectiveness, it is operated in two parts;(1) Regular monitoring of the middle objective values, and (2)  Adjorunes the experiments that don’t meet the predefined possibilities.  
  • In Optuna, ‘report API’ is accountable to monitor functionality, and ‘should prune API’ is subjected in order to eliminate unfavorable trials.
  • While implementing real pruning algorithms in Optuna, inputs that are provided to the algorithm are the trial subjected to pruning, the number of steps, reducing components, least support of resources used for pruning, and required earlier minimum stopping rate.



3. Scalable and versatile System that is easy to setup


The final criteria for the next generation optimization software is a scalable system to control numerous tasks from the conduction of heavy experiments that needs many workers for a trial-level to light-weight experiments through interactive interfaces such as Jupyter Notebook.    


The below diagram shows how the storage is organized into the Optuna system where;

The image depicts the Optuna storage system.

Optuna storage system

  • The trail object shares the history of the evaluation of objective functions through the database.  

  • Optuna also offers users to alter the backend storage in order to meet requirements according to function, for instance, users can conduct experiments on Jupyter notebook in a local device without consuming the time for obtaining a huge-occupant system implemented by some organization.

  • Optuna also gives a web-dashboard for visualization and evaluation of studies in a real-time.

  • Optuna deploys automated in-built, in-memory data-based structure for the backend storage when no specification is provided.




Optuna significantly supports the new-designed approach for next-generation optimization framework that will be embraced for designing frameworks in the future.  As an open-sourced software, Optune will be developed ahead as a posterity software after having interaction with open-source society.     


We have learned major key features of Optuna as the define-by-run API that allows users to construct search spaces, a mixture of efficient searching, and pruning algorithm to enhance the cost-effectiveness of optimization and last the scalable and versatile structure that permit users to use frameworks for many applications. For more learning, follow us at  Facebook, Twitter, and LinkedIn

Latest Comments