A comprehensive ML Metadata walkthrough for Tensorflow Extended

Accessing Data

There’s a tremendous amount of information available, just for a simple 3-step pipeline that runs locally. This pipeline can run in the cloud on a Dataflow runner, for example, with minimum changes in configuration.

In this scenario, it’s much easier to use data that’s stored in the database, instead of browsing cloud storage buckets and VMs on a server farm.

From this point on, you can connect to a ML Metadata store either from a direct SQL connection, or by gRPC (via stub or plain old calls). Then, it’s a matter of selecting the kinds of data you want to inspect manually. This could be the schema or the statistics protobuf, for example.

Typically, you only need to access the resource identifiers of the resources. You should be able to access them via only the URI if you’re in the same environment (ex. a notebook inside a GCP Project VM).

Example Use Case

Assume that you’ve got a pipeline running in some interval (or event-based triggering) and, sometimes, you want to view the data statistics of the latest pipeline run in comparison to the previous run.

• You need the StatisticsGen/statistics artifacts of 2 different pipeline runs (these are the ExampleStatistics type, with type_id 8). These can be found on the Artifact table.
• You also need access to the artifact from the correct pipeline runs. The Attribution table associates context_id with artifact_id . The only thing missing is to pinpoint the 2 context_id s you need in order to make a simple select query.
• The Context table also contains timestamp information. For example, the row Pipeline .2020–07–14T23:45:00.508181.StatisticsGen has got a context_id 5.

Context Id 5, corresponds to Artifact Id 3 from the Attribution table. Artifact Id 3 is indeed the Statistics artifact we need.

Fortunately, kubeflow pipelines already do this visualisation automatically

Related Articles

Search on MC.AI