A 3D model is a virtual representation of an object that can be explored from any angle in a virtual environment. It can be used for medical imaging, military training, architecture, film production and even to visualize a spacecraft launch.
A typical Explore model consists of 5 regions to help navigate, modify and display the case information displayed in a single view. The first region is the Recent tab, which displays a list of cases that have been recently opened. The second region is the Views tab, which contains a list of views that are available for use in an Explore model. Views can be defined by reference to a database table or by aggregation. The third and fourth regions contain additional parameters that affect how a view is filtered and displayed. Finally, the fifth region allows you to add an Explore to a query.
The MLflow model visualization tool, Explore models, makes it easy to understand and modify complex models. Visualize a model by comparing its original and converted graph side-by-side, or dive into a graph layer to inspect it at the granularity you need. For example, you can easily see how a MobileBert model is broken down into a self-attention mask and embedding layer with Model Explorer’s hierarchical view.
Model Explorer allows you to quickly identify and fix problems with your model. For example, a sudden change in the expected reward distribution of an arm can indicate that the model is not performing well. In this situation, Plan2Explore will optimize the model’s arms by maximizing their expected rewards using a novelty metric, which measures the additional information that would be gained by collecting new data points.
Once the model is optimized, it is pushed into the serving layer to be used in real-time ranking along with customer-centric click prediction models to capture the subtle nuances of content preferences at the individual customer level. This approach reduces presentation bias faster compared to a traditional approach, which requires the model to ingest new customer interactions and apply an offline model inference on its own.
The Model Explorer also enables you to see the complete lineage of a model version in a single place. You can select a model and view its history in the Model Overview page or in the History dialog.
The History dialog lists changes made to a model. If a model changes multiple dimensions, each dimension appears as its own entry in the list. Each entry details the change and includes a link to the MLflow run that produced the model version. The Description column in the History dialog provides more detail about the change, such as a description of how the change was made or whether it is a regression or non-regression change. In addition, the History dialog shows the number of times that the model has been modified.