Experiment tracking. Commented [M(75]: Reference: MLflow, “MLflow Website,” [Online]. Available: xxxxx://xxxxxx.xxx/. [Accessed 23 11 2022]. The next application topic is the assessment of learning tracking software, more specifically [A21]. MLflow is an open-source platform that claims to streamline the end-to-end machine learning lifecycle. It offers tools for experiment tracking, project packaging, model management, and deployment. Data scientists and machine learning engineers use MLflow to log and compare experiments, package code into reusable projects, manage machine learning models, and transition from development to deployment. Of interest in the use case is the capability of managing experiments (i.e., training runs) and attach the resulting actor and critic models to it. Its server-client-structure enables tracking across machines including remote hosting of databases. Implementation-wise, a new experiment is initialized at the tracking server by posting training parameters during at the start of each training, as shown in Figure 23. Each experiment possesses a unique id which is attached to posts of values of interest (loss, return, etc.) to the tracking server during the training. The tracking server stores the parameter and values in SQL storage as well as files in an artifact store if desired. Examples of files of interest would be the weights of the trained NNs. In the UUV UC, the storage databases are in the cloud for better sharing of experimental results between consortium partners. Version Status Date Page 2.0 Non-Confidential 2024.05.1172022.03.1 48/100 Figure 23 Use-case agnostic Overview diagram of the MLFlow components.
