Skip to main content

New in Wave ML: Overview

· 4 min read
Peter Szabó
Peter Szabó
Lead Software Engineer @ H2O.ai

Several versions of H2O Wave ML have been released since the last announcement. Wave ML can now run on H2O AI Hybrid Cloud, utilize H2O Driverless AI (DAI) to train the models and push them into H2O MLOps. In addition, new utility functions were introduced to support H2O Enterprise Steam together with MLOps and minor changes to API were made.

Wave ML on Cloud#

You can run both the H2O-3 and DAI engines on Cloud.

For H2O-3 it works the same as running Wave ML locally - the H2O-3 server is initialized and used. No additional steps are necessary (the future work will extend this further to use Cloud resources).

For DAI it's essential to enable OpenID Connect (OIDC) and set up the correct ENV variables. This is done in your app.toml file.

At first, check if the Cloud you are using (we have several instances) has appropriate, shared secrets available:

> h2o secret list

We are looking for MLOps gateway and Steam address:

h2oai-mlops            ALL_USERS  gatewayh2oai-steam            ALL_USERS  api

Set up your app.toml accordingly:

[Runtime]EnableOIDC = true
[[Env]]Name = "H2O_WAVE_ML_STEAM_ADDRESS"Secret = "h2oai-steam"SecretKey = "api"
[[Env]]Name = "H2O_WAVE_ML_MLOPS_GATEWAY"Secret = "h2oai-mlops"SecretKey = "gateway"

You can use your Steam DAI instance with Wave ML in either of two ways:

  1. Specify the instance name in the ENV variable. This option is suitable for a multinode setup:
[[Env]]Name = "H2O_WAVE_ML_STEAM_CLUSTER_NAME"Value = "my-cool-multinode-project-name"
  1. Specify the instance name directly in API:
model = build_model(_steam_dai_instance_name='my-dai')

The rest of the Wave ML workflow continues to be the same.

Utility Functions#

A new module was introduced to Wave ML available as h2o_wave_ml.utils. It contains some useful functions:

  • list_dai_instances(): Gets a list of all available Driverless instances.
  • list_dai_multinodes(): Gets a list of all available Driverless multinode instances.
  • save_autodoc(): Saves the AutoDoc of DAI model.

Use list_dai_instances() to obtain a list of your DAI instances. You can use it later to feed build_model() with particular instance name:

from h2o_wave_ml.utils import list_dai_instances
@app('/')async def serve(q: Q):    dai_instances = list_dai_instances(access_token=q.auth.access_token)

An AutoDoc is available to download if the model was built using DAI backend. Use save_autodoc() to save a .docx file to drive:

from h2o_wave_ml import build_modelfrom h2o_wave_ml.utils import save_autodoc
@app('/')async def serve(q: Q):    model = build_model(...)    file_name = save_autodoc(project_id=model.project_id, access_token=q.auth.access_token)

The .project_id property of the model is always available if the model was obtained by build_model() function. However, if the model is returned using get_model(), it's available only if the model belongs to an authenticated user.

See the API for utils.

OpenID Connect#

OIDC tokens are used to authenticate across the services on Cloud. Tokens are available within a q argument of the Wave query handler, i.e., q.auth.access_token and q.auth.refresh_token. For more information, see the single sign-on section of the Wave page.

Some of the Wave ML functions were equipped with access_token and refresh_token arguments to handle the authentication routines behind the scenes. Use the appropriate token to authenticate.

In general, access_token has a small lifespan, and it's not suitable for a long-running job. Use it when you are sure your routine is handled swiftly, e.g., the listing example:

from h2o_wave_ml.utils import list_dai_instances
@app('/')async def serve(q: Q):    dai_instances = list_dai_instances(access_token=q.auth.access_token)

If your task needs more time (e.g. build_model()), use refresh_token so Wave ML can generate access tokens as needed:

from h2o_wave_ml import build_model
@app('/')async def serve(q: Q):    model = build_model(refresh_token=q.auth.refresh_token, ...)

Notable Features#

  • build_model() was updated with several H2O-3 and DAI parameters. See an example for H2O-3 parameters.

  • Pandas dataframe can be used for training, testing and validating:

from h2o_wave_ml import build_model
from sklearn.datasets import load_winefrom sklearn.model_selection import train_test_split
data = load_wine(as_frame=True)['frame']train_df, test_df = train_test_split(data, train_size=0.8)
model = build_model(train_df=train_df, target_column='target')

  • A validation dataset can be passed to build_model() function. Choose from validation_file_path in case of a file or validation_df in case of a Pandas dataframe.

  • A list of categorical columns can be specified using the categorical_columns argument of a build_model() function.

  • Columns can be dropped or included using drop_columns or feature_columns of a build_model() function.

See the API for Wave ML and examples on official Wave page.

Follow updates to Wave ML on GitHub: https://github.com/h2oai/wave-ml. Let us know what you think and how we can improve it.

We look forward to continuing our collaboration with the community and hearing your feedback as we further improve and expand the H2O Wave platform.