APC 技術ブログ


株式会社 エーピーコミュニケーションズの技術ブログです。

Latest Advancements in MLflow


This session delves deep into the latest features of MLflow and its ongoing evolution, introducing some incredibly compelling developments. This section provides a basic overview of the MLflow lifecycle.

Introduction and Overview of the MLflow Lifecycle

First and foremost, it is necessary to highlight the significant contributions of the team behind MLflow, particularly David Briggs and his colleague Ben. These experts have extensively utilized the "Drain-A-Anim" feature in their development efforts.

The session focused on how MLflow is designed to standardize and simplify the management of models throughout the ML lifecycle. Particularly, the recent advancements in deep learning and the improvements related to the "Drain-A-Anim" feature were discussed. These insights suggest promising possibilities for future improvements to MLflow.

We gained a deeper understanding of MLflow's ability to streamline project management. Through concrete examples, the presenters demonstrated how MLflow's tools and features can be optimized across different projects.

Furthermore, the discussion highlighted the "Drain-A-Anim" feature, discussing its crucial role within MLflow and how it integrates with broader MLflow functionalities through practical deployment examples.

As MLflow continues to evolve, look forward to its next version. We were encouraged to leverage this innovative platform to enhance the effectiveness and success of machine learning efforts. Let's continue to refine our approach to managing the machine learning lifecycle with the help of MLflow!

Latest Features and Enhancements of MLflow

Recent updates to MLflow have incorporated many significant improvements, especially with the release of MLflow 2.11 and 2.12, which introduced standout features particularly tailored for deep learning cases.

Main Updates in MLflow 2.11:

  1. Improved Dashboard and UI: The user interface of MLflow has evolved significantly, enhancing the visibility and operability of projects utilizing deep learning.

  2. Comprehensive Support for Model Provisioning and Auto Checkpointing: Comprehensive support is now provided for various models, with automatic checkpoints during deep learning model training to prevent critical data loss.

  3. Simplified Custom MLflow Model Building: Building, deploying, and comparing custom models has become easier, aiding software engineers in the construction and deployment processes of MLflow metrics.

Additional Features in MLflow 2.12:

  1. Enhanced Support for Tracking and Results Exploration: The ability to track and explore results for any model has been significantly improved, enabling rapid evaluation and comparison of various model performances.

  2. Easier Model Packaging and Deployment: Models can now be conveniently packaged and deployed via any Python script or method, facilitating seamless integration into existing workflows, especially valuable for software engineers tasked with deploying models.

These updates have made MLflow an even more user-friendly and efficient tool for managing and automating machine learning projects. Users leveraging these new features can streamline the progress of their projects and lead to the creation of high-precision models.

New Approaches for Efficient Model Management

The new APIs introduced by MLflow have revolutionized the way models are managed. Traditionally, handling models involved using multiple MLflow APIs, preparing, setting up, and then logging in Python, and exporting to the project. The latest approach offers a more efficient method to identify where the "code is" and establish a "history of how the model was built."

A significant improvement is the introduction of the setModel API. This new feature simplifies the definition and instantiation of models, applicable not only to Gemini but any Python function.

Streamlined Model Definition and Logging

The model setup process now includes creating a file named model_for_code.py. This file serves as the foundation for defining models within Python using the setModel API. Leveraging this API automates model setup, and crucial details are recorded as part of MLflowRFact. This process not only accelerates model management but also enhances the system's robustness, ensuring that necessary information is accurately recorded.

Ensuring High Standards of Code Quality

An important part of model setting and logging involves inspections for code quality. These inspections are designed to ensure that the code adheres to pre-established quality standards. Through these checks, MLflow ensures that each part of the code used in model management meets the highest standards, enhancing the overall reliability and effectiveness of the model.

By leveraging these new features, MLflow continues to refine the landscape of model management and evaluation, providing scientists and developers in data science and machine learning fields with more efficient and reliable solutions.

Recent summits have highlighted several new features of MLflow, with a focus on enhancements to "comprehensive evaluation and tracking capabilities." This section will discuss these enhancements in detail and how they can be utilized.

Enhanced Evaluation Features

MLflow's latest update has significantly strengthened its model evaluation capabilities. In this new system, all elements including input values, plans, and scores are comprehensively tracked, enabling intuitive comparisons between models like never before. For instance, if one wishes to compare multiple models against a single input, simply setting up the input and output lines allows for the quick and easy evaluation of each model's scores.

Enhanced Tracking Features

Additionally, tracking features have also been strengthened. Users can now set parameters for a specific method and create models while specifying the underlying model, enabling more detailed tracking and assessment. This adjustment significantly streamlines the update and refinement of models.

Addition of New Metrics

The update also includes the addition of new metrics, allowing users to create their own customized model evaluation criteria, thus enhancing the flexibility and accuracy of model evaluations.


These advancements continue to provide users with a more sophisticated tool for model management, facilitating deeper insights and more managed experiments in machine learning projects. With these upgraded evaluation and tracking tools, users can optimize models with unprecedented precision and clarity.

The introduction of these features strengthens the entire lifecycle of model development and management within MLflow, demonstrating a commitment to innovation and responsiveness to the needs of the machine learning community. As MLflow evolves, it remains a critical tool for developers looking to streamline ML operations efficiently. Therefore, these improvements represent significant advancements in the field of machine learning technology, promoting more detailed and intuitive management of complex ML processes.

About the special site during DAIS

This year, we have prepared a special site to report on the session contents and the situation from the DAIS site! We plan to update the blog every day during DAIS, so please take a look.