AI Artifact Management & Traceability via H2O MLOps | Part 9

Models don't exist in isolation. They're the output of experiments built from data sets trained with specific configurations and evaluated with various metrics. Let me show you how we link all these artifacts together. Every model in our repository is linked back to the experiment that produced it. You can see the details of the exact driverless experiment, the data version, the feature engine configuration, and all the training parameters. Models carry their artifacts with them. For example, the autod do report we generated explaining the model. The mojo pipeline for low latency scoring, Python scoring pipeline for flexibility, even featur analysis. All of these artifacts are attached to the model record in the workspace and saved together. Let me show you the experiment side of the relationship. In driver sei, every experiment is tracked with complete configuration capture. the data set reference, the target variable, validation strategy, and so forth. You can run multiple experiments in parallel. Maybe you're testing different validation strategies or comparing different accuracy levels. All experiments are tracked and you can compare them side by side to see which approach worked best. We provide managed containerized runtimes for model deployment. System administrators define available runtimes via help. Maybe you have a standard Python 313 runtime, GPU enabled runtime for deep learning models, and a lock down runtime for highly regulated workloads. Data scientists deploy these managed runtimes without needing to build the containers themselves.