An H2O MLOps Python Client

Example

import h2o_mlops
import h2o_mlops.options as options
import h2o_mlops.types as types

First, we need to connect to MLOps. In the default case, the client detects credentials and configuration options from the environment.

mlops = h2o_mlops.Client()

Alternatively, you can initialize the client explicitly by passing the required parameters.

mlops = h2o_mlops.Client(
    h2o_cloud_url=<H2O_CLOUD_URL>,
    refresh_token=<REFRESH_TOKEN>,
    ssl_cacert="/path/to/your/ca_certificate.pem",  # If SSL is not needed, you can omit it.
)

Replace <H2O_CLOUD_URL> and <REFRESH_TOKEN> with your actual values.

Everything Starts with a Workspace

A workspace is the main base of operations for most MLOps activities.

workspace = mlops.workspaces.create(name="demo")

mlops.workspaces.list(name="demo")

    | name   | uid
----+--------+--------------------------------------
  0 | demo   | 45e5a888-ec1f-4f9c-85ca-817465344b1f

You can also do workspace = mlops.workspaces.get(uid=...).

Upload an Experiment

experiment = workspace.experiments.create(
    data="/path/to/your/model.zip",
    name="experiment-from-client"
)

Some experiment attributes of interest.

experiment.uid

'e307aa9f-895f-4b07-9404-b0728d1b7f03'

Existing experiments can be viewed and retrieved.

workspace.experiments.list()

    | name                   | uid                                  | tags
----+------------------------+--------------------------------------+--------
  0 | experiment-from-client | e307aa9f-895f-4b07-9404-b0728d1b7f03 |

You can also do experiment = workspaces.experiments.get(uid=...).

Create a Model

model = workspace.models.create(name="model-from-client")

Existing models can be viewed and retrieved.

workspace.models.list()

    | name              | uid
----+-------------------+--------------------------------------
  0 | model-from-client | d18a677f-b800-4a4b-8642-0f59e202d225

You can also do model = workspaces.models.get(uid=...).

Register an Experiment to a Model

In order to deploy a model, it needs to have experiments registered to it.

model.register(experiment=experiment)

model.versions()

    |   version | experiment_uid
----+-----------+--------------------------------------
  0 |         1 | e307aa9f-895f-4b07-9404-b0728d1b7f03

model.experiment(model_version="latest").name

'experiment-from-client'

Deployment

What is needed for a single model deployment?

• workspace
• model
• scoring runtime
• security options
• name for deployment

We already have a workspace and model. Next we'll get the scoring_runtime for our model type, from the scoring runtime suggestions for the experiment to select the appropriate one.

model.experiment().scoring_runtimes

    | name              | artifact_type   | uid
----+-------------------+-----------------+-------------------
  0 | H2O-3 MOJO scorer | h2o3_mojo       | h2o3_mojo_runtime

scoring_runtime = model.experiment().scoring_runtimes[0]

Now we can create a deployment.

deployment = workspace.deployments.create(
    name="deployment-from-client",
    composition_options=options.CompositionOptions(
        model=model,
        scoring_runtime=scoring_runtime,
    ),
    security_options=options.SecurityOptions(
        security_type=types.SecurityType.DISABLED,
    ),
)

deployment.wait_for_healthy()
    
deployment.state

'HEALTHY'

Score

Once you have a deployment, you can score with it through the HTTP protocol.

scorer = deployment.scorer

scorer.score(
    payload=scorer.sample_request(auth_value=...),
    auth_value=...,
)

{'fields': ['C11.0', 'C11.1'],
 'id': 'e307aa9f-895f-4b07-9404-b0728d1b7f03',
 'score': [['0.49786656666743145', '0.5021334333325685']]}