Vertex AI Experiments is a tool that helps you track and analyze different model architectures, hyperparameters, and training environments, letting you track the steps, inputs, and outputs of an experiment run. Vertex AI Experiments can also evaluate how your model performed in aggregate, against test datasets, and during the training run. You can then use this information to select the best model for your particular use case.
Experiment runs don't incur additional charges. You're only charged for resources that you use during your experiment as described in Vertex AI pricing.
| What do you want to do? | Check out notebook sample |
|---|---|
| track metrics and parameters | Compare models |
| track experiment lineage | Model training |
| track pipeline runs | Compare pipeline runs |
Track steps, inputs, and outputs
Vertex AI Experiments lets you track:
- steps of an experiment run, for example, preprocessing, training,
- inputs, for example, algorithm, parameters, datasets,
- outputs of those steps, for example, models, checkpoints, metrics.
You can then figure out what worked and what didn't, and identify further avenues for experimentation.
For user journey examples, check out:
Analyze model performance
Vertex AI Experiments lets you track and evaluate how the model performed in aggregate, against test datasets, and during the training run. This ability helps to understand the performance characteristics of the models -- how well a particular model works overall, where it fails, and where the model excels.
For user journey examples, check out:
Compare model performance
Vertex AI Experiments lets you group and compare multiple models across experiment runs. Each model has its own specified parameters, modeling techniques, architectures, and input. This approach helps select the best model.
For user journey examples, check out:
Search experiments
The Google Cloud console provides a centralized view of experiments, a cross-sectional view of the experiment runs, and the details for each run. The Vertex AI SDK for Python provides APIs to consume experiments, experiment runs, experiment run parameters, metrics, and artifacts.
Vertex AI Experiments, along with Vertex ML Metadata, provides a way to find the artifacts tracked in an experiment. This lets you quickly view the artifact's lineage and the artifacts consumed and produced by steps in a run.
Scope of support
Vertex AI Experiments supports development of models using Vertex AI custom training, Vertex AI Workbench notebooks, Notebooks, and all Python ML Frameworks across most ML Frameworks. For some ML frameworks, such as TensorFlow, Vertex AI Experiments provides deep integrations into the framework that makes the user experience automagical. For other ML frameworks, Vertex AI Experiments provides a framework neutral Vertex AI SDK for Python that you can use. (see: Prebuilt containers for TensorFlow, scikit-learn, PyTorch, XGBoost).
Data models and concepts
Vertex AI Experiments is a context
in Vertex ML Metadata where an experiment
can contain n experiment runs in addition to n pipeline runs. An experiment
run consists of parameters, summary metrics, time series metrics, and
PipelineJob, Artifact,
and Execution Vertex AI resources.
Vertex AI TensorBoard, a
managed version of open source TensorBoard, is used for time-series metrics
storage. Executions and artifacts of a pipeline run are viewable
in the Google Cloud console.
Vertex AI Experiments terms
Experiment, experiment run, and pipeline run
experiment
- An experiment is a context that can contain a set of n experiment runs in addition to pipeline runs where a user can investigate, as a group, different configurations such as input artifacts or hyperparameters.
experiment run
- An experiment run can contain user-defined metrics, parameters, executions, artifacts, and Vertex resources (for example, PipelineJob).
pipeline run
- One or more Vertex PipelineJobs can be associated with an experiment where each PipelineJob is represented as a single run. In this context, the parameters of the run are inferred by the parameters of the PipelineJob. The metrics are inferred from the system.Metric artifacts produced by that PipelineJob. The artifacts of the run are inferred from artifacts produced by that PipelineJob.
PipelineJob resource
can be associated with an ExperimentRun resource.
In this context, the parameters, metrics, and artifacts are not inferred.
See Associate a pipeline with an experiment.
Parameters and metrics
parameters
- Parameters are keyed input values that configure a run, regulate the behavior of the run, and affect the results of the run. Examples include learning rate, dropout rate, and number of training steps.
See Log parameters.
summary metrics
- Summary metrics are a single value for each metric key in an experiment run. For example, the test accuracy of an experiment is the accuracy calculated against a test dataset at the end of training that can be captured as a single value summary metric.
See Log summary metrics.
time series metrics
- Time series metrics are longitudinal metric values where each value represents a step in the training routine portion of a run. Time series metrics are stored in Vertex AI TensorBoard. Vertex AI Experiments stores a reference to the Vertex TensorBoard resource.
Resource types
pipeline job
- A pipeline job or a pipeline run corresponds to the PipelineJob resource in the Vertex AI API. It's an execution instance of your ML pipeline definition, which is defined as a set of ML tasks interconnected by input-output dependencies.
artifact
- An artifact is a discrete entity or piece of data produced and consumed by a machine learning workflow. Examples of artifacts include datasets, models, input files, and training logs.
Vertex AI Experiments lets you use a schema to define the type of
artifact. For example, supported schema types include system.Dataset,
system.Model, and system.Artifact. For more information, see
System schemas.