03 Azure Machine Learning and OSS Model Fine tuning

This post has been republished via RSS; it originally appeared at: New blog articles in Microsoft Community Hub.

Hyperparameter optimization, also known as hyperparameter tuning, is a fundamental challenge in the field of machine learning. It involves the selection of an optimal set of hyperparameters for a given learning algorithm. Hyperparameters are parameters that dictate the behavior of the learning process, while other parameters, such as node weights, are learned from the data.

A machine learning model can often require different constraints, weights, or learning rates to effectively capture diverse data patterns. These adjustable measures, known as hyperparameters, must be carefully tuned to ensure that the model can successfully solve the machine learning problem at hand. Hyperparameter optimization seeks to find a combination of hyperparameters that yields an optimal model, minimizing a predefined loss function on independent data.

To achieve this, an objective function is utilized, which takes a set of hyperparameters as input and returns the corresponding loss. The goal is to find the set of hyperparameters that maximizes the generalization performance of the model. Cross-validation is commonly employed to estimate this performance and aid in the selection of optimal hyperparameter values. By maximizing the generalization performance, hyperparameter optimization plays a crucial role in enhancing the overall effectiveness and accuracy of machine learning models.


In this post we will cover the open-source tools and Azure Machine learning tools around hyperparameter tuning.  There are three main techniques used for hyperparameter, Grid Search, Random Search, Bayesian Search and more commonly used for Neural Networks, gradient based optimization.


Grid Search

In the realm of hyperparameter optimization, the conventional approach has been to employ grid search or parameter sweep. This technique involves exhaustively exploring a predetermined subset of hyperparameters for a learning algorithm. To guide the grid search algorithm, a performance metric is selected, often determined through cross-validation on the training set or evaluation on a dedicated validation set.

Grid search operates by systematically testing different combinations of hyperparameters within the defined subset. This method, however, can become computationally expensive, especially when dealing with a large number of hyperparameters or a wide range of possible values. Despite its limitations, grid search remains widely used due to its simplicity and interpretability.

During the grid search process, various performance metrics are measured for each combination of hyperparameters. These metrics aid in assessing the model's effectiveness and allow for the identification of hyperparameter configurations that lead to optimal performance. By evaluating the model's performance on either the training set or a separate validation set, grid search facilitates the selection of the most appropriate hyperparameter values.

While grid search has proven to be a valuable technique, alternative methods have emerged to address its shortcomings, such as the computationally efficient and automated approaches of Bayesian optimization and random search. These advanced methods provide more sophisticated ways to explore the hyperparameter space and discover optimal configurations, revolutionizing the field of hyperparameter optimization.


sklearn.model_selection.GridSearchCV — scikit-learn 1.4.2 documentation


Random Search

Random Search offers an alternative to the exhaustive enumeration of all possible combinations of hyperparameters by randomly selecting them. This approach can be applied not only to discrete settings but also to continuous and mixed spaces, providing greater flexibility. Random Search has been found to outperform Grid Search, particularly in scenarios where only a small number of hyperparameters significantly impact the final performance of the machine learning algorithm.

In cases where the optimization problem exhibits a low intrinsic dimensionality, Random Search proves to be particularly effective. This refers to situations where the hyperparameters' interdependencies are limited, allowing for a more efficient exploration of the hyperparameter space. Moreover, Random Search lends itself to embarrassingly parallel implementation, meaning that it can be easily distributed across multiple computing resources for faster processing.


One of the advantages of Random Search is its ability to incorporate prior knowledge by specifying the distribution from which to sample hyperparameters. This enables domain experts to guide the search process based on their understanding of the problem at hand. Despite its simplicity, Random Search remains a significant baseline against which new hyperparameter optimization methods can be compared.

While Random Search has been instrumental in advancing hyperparameter optimization, it is important to note that other sophisticated techniques, such as Bayesian optimization, have emerged as promising alternatives. These methods leverage probabilistic models to intelligently explore the hyperparameter space and efficiently find optimal configurations. The continuous development of new approaches continues to enhance the field of hyperparameter optimization, offering exciting opportunities for improving the performance and efficiency of machine learning models.


sklearn.model_selection.RandomizedSearchCV — scikit-learn 1.4.2 documentation


Bayesian Search

Bayesian optimization is a powerful method for globally optimizing noisy black-box functions. When applied to hyperparameter optimization, Bayesian optimization constructs a probabilistic model that captures the relationship between hyperparameter values, and the objective function evaluated on a validation set. Through an iterative process, Bayesian optimization intelligently selects hyperparameter configurations based on the current model, evaluates their performance, and updates the model to gather valuable information about the function and, more importantly, the location of the optimum.

The key idea behind Bayesian optimization is to strike a balance between exploration and exploitation. Exploration involves selecting hyperparameters that yield uncertain outcomes, while exploitation focuses on hyperparameters that are expected to be close to the optimum. By carefully navigating this trade-off, Bayesian optimization effectively explores the hyperparameter space, gradually narrowing down the search to regions with higher potential for optimal performance.

In practice, Bayesian optimization has demonstrated superior performance compared to traditional methods such as grid search and random search. This advantage stems from its ability to reason about the quality of experiments before actually running them. By leveraging the probabilistic model, Bayesian optimization can make informed decisions about which hyperparameter configurations are most likely to lead to better results, thereby reducing the number of evaluations required.

The efficiency and effectiveness of Bayesian optimization have been widely observed in various domains. Researchers and practitioners have embraced this approach due to its ability to achieve better outcomes with fewer evaluations, making it an invaluable tool for hyperparameter optimization.


skopt.BayesSearchCV — scikit-optimize 0.8.1 documentation

Hyperparameter optimization - Wikipedia



Having familiarized ourselves with the foundational aspects of hyperparameter tuning, a natural inquiry arises: Does Azure Machine Learning (AML) provide support for hyperparameter tuning? The answer to this question is undoubtedly affirmative, as AML offers comprehensive hyperparameter tuning capabilities. Detailed documentation on this topic can be accessed in the link below, providing users with valuable guidance and information.

Hyperparameter tuning a model (v2) - Azure Machine Learning | Microsoft Learn

AML, through its Python SDK V2, facilitates hyperparameter tuning by offering three distinct algorithms: Grid, Random, and Bayesian. These algorithms empower users to effectively explore the hyperparameter search space and optimize their machine learning models. To leverage the hyperparameter tuning capabilities in AML, the following essential steps can be followed:

  1. Define the parameter search space for your trial: Specify the range and feasible values for each hyperparameter that will undergo tuning.
  2. Specify the sampling algorithm for your sweep job: Select the desired algorithm that will be employed to sample hyperparameter configurations during the tuning process.
  3. Specify the objective to optimize: Define the performance metric or objective function that will be utilized to evaluate and compare the various hyperparameter configurations.
  4. Specify an early termination policy for low-performing jobs: Establish criteria that will automatically terminate underperforming jobs during the hyperparameter tuning process.
  5. Define limits for the sweep job: Set the maximum number of iterations or allocate resources according to your requirements for the hyperparameter tuning experiment.
  6. Launch an experiment with the defined configuration: Initiate the hyperparameter tuning experiment by utilizing the specified settings and parameters.
  7. Visualize the training jobs: Monitor and analyze the progress and outcomes of the hyperparameter tuning experiment, including the performance of individual training jobs.
  8. Select the best configuration for your model: Upon completion of the hyperparameter tuning experiment, identify the hyperparameter configuration that yielded the most favorable performance, and incorporate it into your machine learning model.


Principal author:

  • Shep Sheppard | Senior Customer Engineer, FastTrack for ISV and Startups

Other contributors:

  • Yoav Dobrin Principal Customer Engineer, FastTrack for ISV and Startups
  • Jones Jebaraj | Senior Customer Engineer, FastTrack for ISV and Startups
  • Olga Molocenco-Ciureanu | Customer Engineer, FastTrack for ISV and Startups

Leave a Reply

Your email address will not be published. Required fields are marked *


This site uses Akismet to reduce spam. Learn how your comment data is processed.