H2O-3 MLFlow Flavor
===================
A tiny library containing a `MLFlow <https://mlflow.org/>`_ flavor for working with H2O-3 MOJO and POJO models.
Logging Models to MLFlow Registry
---------------------------------
The model that was trained with H2O-3 runtime can be exported to MLFlow registry with `log_model` function.:
.. code-block:: Python
import mlflow
import h2o_mlflow_flavor
mlflow.set_tracking_uri("http://127.0.0.1:8080")
h2o_model = ... training phase ...
with mlflow.start_run(run_name="myrun") as run:
h2o_mlflow_flavor.log_model(h2o_model=h2o_model,
artifact_path="folder",
model_type="MOJO",
extra_prediction_args=["--predictCalibrated"])
Compared to `log_model` functions of the other flavors being a part of MLFlow, this function has two extra arguments:
* ``model_type`` - It indicates whether the model should be exported as `MOJO <https://docs.h2o.ai/h2o/latest-stable/h2o-docs/mojo-quickstart.html#what-is-a-mojo>`_ or `POJO <https://docs.h2o.ai/h2o/latest-stable/h2o-docs/pojo-quickstart.html#what-is-a-pojo>`_. The default value is `MOJO`.
* ``extra_prediction_args`` - A list of extra arguments for java scoring process. Possible values:
* ``--setConvertInvalidNum`` - The scoring process will convert invalid numbers to NA.
* ``--predictContributions`` - The scoring process will Return also Shapley values a long with the predictions. Model must support that Shapley values, otherwise scoring process will throw an error.
* ``--predictCalibrated`` - The scoring process will also return calibrated prediction values.
The `save_model` function that persists h2o binary model to MOJO or POJO has the same signature as the `log_model` function.
Extracting Information about Model
----------------------------------
The flavor offers several functions to extract information about the model.
* ``get_metrics(h2o_model, metric_type=None)`` - Extracts metrics from the trained H2O binary model. It returns dictionary and takes following parameters:
* ``h2o_model`` - An H2O binary model.
* ``metric_type`` - The type of metrics. Possible values are "training", "validation", "cross_validation". If parameter is not specified, metrics for all types are returned.
* ``get_params(h2o_model)`` - Extracts training parameters for the H2O binary model. It returns dictionary and expects one parameter:
* ``h2o_model`` - An H2O binary model.
* ``get_input_example(h2o_model, number_of_records=5, relevant_columns_only=True)`` - Creates an example Pandas dataset from the training dataset of H2O binary model. It takes following parameters:
* ``h2o_model`` - An H2O binary model.
* ``number_of_records`` - A number of records that will be extracted from the training dataset.
* ``relevant_columns_only`` - A flag indicating whether the output dataset should contain only columns required by the model. Defaults to ``True``.
The functions can be utilized as follows:
.. code-block:: Python
import mlflow
import h2o_mlflow_flavor
mlflow.set_tracking_uri("http://127.0.0.1:8080")
h2o_model = ... training phase ...
with mlflow.start_run(run_name="myrun") as run:
mlflow.log_params(h2o_mlflow_flavor.get_params(h2o_model))
mlflow.log_metrics(h2o_mlflow_flavor.get_metrics(h2o_model))
input_example = h2o_mlflow_flavor.get_input_example(h2o_model)
h2o_mlflow_flavor.log_model(h2o_model=h2o_model,
input_example=input_example,
artifact_path="folder",
model_type="MOJO",
extra_prediction_args=["--predictCalibrated"])
Model Scoring
-------------
After a model obtained from the model registry, the model doesn't require h2o runtime for ability to score. The only thing
that model requires is a ``h2o-gemodel.jar`` which was persisted with the model during saving procedure.
The model could be loaded by the function ``load_model(model_uri, dst_path=None)``. It returns an objecting making
predictions on Pandas dataframe and takes the following parameters:
* ``model_uri`` - An unique identifier of the model within MLFlow registry.
* ``dst_path`` - (Optional) A local filesystem path for downloading the persisted form of the model.
The object for scoring could be obtained also via the `pyfunc` flavor as follows:
.. code-block:: Python
import mlflow
mlflow.set_tracking_uri("http://127.0.0.1:8080")
logged_model = 'runs:/9a42265cf0ef484c905b02afb8fe6246/iris'
loaded_model = mlflow.pyfunc.load_model(logged_model)
import pandas as pd
data = pd.read_csv("http://h2o-public-test-data.s3.amazonaws.com/smalldata/iris/iris_wheader.csv")
loaded_model.predict(data)
Raw data
{
"_id": null,
"home_page": "https://github.com/h2oai/h2o-3.git",
"name": "h2o-mlflow-flavor",
"maintainer": "",
"docs_url": null,
"requires_python": "",
"maintainer_email": "",
"keywords": "ML Flow,H2O-3",
"author": "H2O.ai",
"author_email": "support@h2o.ai",
"download_url": "",
"platform": null,
"description": "H2O-3 MLFlow Flavor\n===================\n\nA tiny library containing a `MLFlow <https://mlflow.org/>`_ flavor for working with H2O-3 MOJO and POJO models.\n\nLogging Models to MLFlow Registry\n---------------------------------\n\nThe model that was trained with H2O-3 runtime can be exported to MLFlow registry with `log_model` function.:\n\n.. code-block:: Python\n\n import mlflow\n import h2o_mlflow_flavor\n\n mlflow.set_tracking_uri(\"http://127.0.0.1:8080\")\n \n h2o_model = ... training phase ...\n \n with mlflow.start_run(run_name=\"myrun\") as run:\n\th2o_mlflow_flavor.log_model(h2o_model=h2o_model,\n artifact_path=\"folder\",\n model_type=\"MOJO\",\n extra_prediction_args=[\"--predictCalibrated\"])\n\n\nCompared to `log_model` functions of the other flavors being a part of MLFlow, this function has two extra arguments:\n\t\n* ``model_type`` - It indicates whether the model should be exported as `MOJO <https://docs.h2o.ai/h2o/latest-stable/h2o-docs/mojo-quickstart.html#what-is-a-mojo>`_ or `POJO <https://docs.h2o.ai/h2o/latest-stable/h2o-docs/pojo-quickstart.html#what-is-a-pojo>`_. The default value is `MOJO`.\n\n* ``extra_prediction_args`` - A list of extra arguments for java scoring process. Possible values:\n\n * ``--setConvertInvalidNum`` - The scoring process will convert invalid numbers to NA.\n\n * ``--predictContributions`` - The scoring process will Return also Shapley values a long with the predictions. Model must support that Shapley values, otherwise scoring process will throw an error.\n\n * ``--predictCalibrated`` - The scoring process will also return calibrated prediction values.\n \nThe `save_model` function that persists h2o binary model to MOJO or POJO has the same signature as the `log_model` function.\n\nExtracting Information about Model\n----------------------------------\n\nThe flavor offers several functions to extract information about the model.\n\n* ``get_metrics(h2o_model, metric_type=None)`` - Extracts metrics from the trained H2O binary model. It returns dictionary and takes following parameters:\n\n * ``h2o_model`` - An H2O binary model.\n\n * ``metric_type`` - The type of metrics. Possible values are \"training\", \"validation\", \"cross_validation\". If parameter is not specified, metrics for all types are returned.\n\n* ``get_params(h2o_model)`` - Extracts training parameters for the H2O binary model. It returns dictionary and expects one parameter:\n\n * ``h2o_model`` - An H2O binary model.\n\n* ``get_input_example(h2o_model, number_of_records=5, relevant_columns_only=True)`` - Creates an example Pandas dataset from the training dataset of H2O binary model. It takes following parameters:\n\n * ``h2o_model`` - An H2O binary model.\n\n * ``number_of_records`` - A number of records that will be extracted from the training dataset.\n\n * ``relevant_columns_only`` - A flag indicating whether the output dataset should contain only columns required by the model. Defaults to ``True``.\n \nThe functions can be utilized as follows:\n\n.. code-block:: Python\n\n import mlflow\n import h2o_mlflow_flavor\n \n mlflow.set_tracking_uri(\"http://127.0.0.1:8080\")\n\n h2o_model = ... training phase ...\n\n with mlflow.start_run(run_name=\"myrun\") as run:\n\t mlflow.log_params(h2o_mlflow_flavor.get_params(h2o_model))\n\t mlflow.log_metrics(h2o_mlflow_flavor.get_metrics(h2o_model))\n\t input_example = h2o_mlflow_flavor.get_input_example(h2o_model)\n\t h2o_mlflow_flavor.log_model(h2o_model=h2o_model,\n input_example=input_example,\n artifact_path=\"folder\",\n model_type=\"MOJO\",\n extra_prediction_args=[\"--predictCalibrated\"])\n\n\nModel Scoring\n-------------\n\nAfter a model obtained from the model registry, the model doesn't require h2o runtime for ability to score. The only thing\nthat model requires is a ``h2o-gemodel.jar`` which was persisted with the model during saving procedure.\nThe model could be loaded by the function ``load_model(model_uri, dst_path=None)``. It returns an objecting making\npredictions on Pandas dataframe and takes the following parameters:\n\n* ``model_uri`` - An unique identifier of the model within MLFlow registry.\n\n* ``dst_path`` - (Optional) A local filesystem path for downloading the persisted form of the model. \n\nThe object for scoring could be obtained also via the `pyfunc` flavor as follows:\n\n.. code-block:: Python\n\n import mlflow\n mlflow.set_tracking_uri(\"http://127.0.0.1:8080\")\n\n logged_model = 'runs:/9a42265cf0ef484c905b02afb8fe6246/iris'\n loaded_model = mlflow.pyfunc.load_model(logged_model)\n\n import pandas as pd\n data = pd.read_csv(\"http://h2o-public-test-data.s3.amazonaws.com/smalldata/iris/iris_wheader.csv\")\n loaded_model.predict(data)\n",
"bugtrack_url": null,
"license": "Apache v2",
"summary": "A mlflow flavor for working with H2O-3 MOJO and POJO models",
"version": "0.1.0",
"project_urls": {
"Homepage": "https://github.com/h2oai/h2o-3.git"
},
"split_keywords": [
"ml flow",
"h2o-3"
],
"urls": [
{
"comment_text": "",
"digests": {
"blake2b_256": "82158e4b5d385c40f9a811dde97f37b7fd12bef614c03828c7db9b5d8ca3fcee",
"md5": "070433e6ca894c25c2f8a0b241aa8cf7",
"sha256": "2b06523b02d98ef914d2dc76914430a6655e949376ad150fe764b73cf5158777"
},
"downloads": -1,
"filename": "h2o_mlflow_flavor-0.1.0-py2.py3-none-any.whl",
"has_sig": false,
"md5_digest": "070433e6ca894c25c2f8a0b241aa8cf7",
"packagetype": "bdist_wheel",
"python_version": "py2.py3",
"requires_python": null,
"size": 7416,
"upload_time": "2023-11-13T18:13:52",
"upload_time_iso_8601": "2023-11-13T18:13:52.175331Z",
"url": "https://files.pythonhosted.org/packages/82/15/8e4b5d385c40f9a811dde97f37b7fd12bef614c03828c7db9b5d8ca3fcee/h2o_mlflow_flavor-0.1.0-py2.py3-none-any.whl",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2023-11-13 18:13:52",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "h2oai",
"github_project": "h2o-3",
"travis_ci": false,
"coveralls": false,
"github_actions": true,
"lcname": "h2o-mlflow-flavor"
}
JSON
