Enabling MLOps in the SPIRA Training Pipeline

Page for the discipline MAC0499 - Capstone Project - at IME-USP.

Members: Daniel Angelo Esteves Lawand (10297693)

Supervisors:

• Prof. Dr. Alfredo Goldman
• M.Sc. Renato Cordeiro Ferreira

Summary:

SPIRA is a ML-enabled system to pre-diagnosis insufficiency respiratory based on speech analysis. The training pipeline created during its initial modelling led to an accidental architecture, which lacked good software quality attributes. The goal of this research is to design a continuous training pipeline to enable MLOps for SPIRA. Based on the issues found in the experimental implementation, this research proposes a new architecture for the system. It combines design principles and design patterns with the hexagonal architecture pattern, supporting several quality attributes: maintainability, extensibility, reusability, robustness, efficiency, understandability, and readability. With them, the training pipeline can be easily modified and evolved alongside its three axes of change: code, model, and data. Furthermore, this research also improves the infrastructure of the system, to ensure it is easy to run. This is essential to continuously deliver new versions of the SPIRA model. Together, these changes enable MLOps practices in the project.

Proposal (in Portuguese)

Download

Online

Monograph

Download

Online

Repository

Training Pipeline

Contact Information

Any questions? Send an e-mail to daniel.lawand at usp.br

_{Page inspired by the Bachelor’s Monograph of Vitor Daisuke Tamae}