deploy#

datarobotx.deploy(model, *args, target=None, classes=None, name=None, description=None, hooks=None, extra_requirements=None, **kwargs)[source]#

Deploy a model to MLOps

If model object is a non-DR model, will build and configure an appropriate supporting custom environment and custom inference model within DataRobot as part of deployment.

Parameters:

  • model (Any) – The model object to deploy. See the drx documentation for additional information on supported model types.

  • *args (Any, optional) – Additional model objects; required for certain model types, e.g. Huggingface tokenizer + pre-trained model

  • target (str, optional) – Name of the target variable; required for supervised model types

  • classes (list of str, optional) – Names of the target variable classes; required for supervised classification problems; for binary classification, the first item should be the positive class

  • name (str, optional) – Name of the ML Ops deployment

  • description (str, optional) – Short description for the ML Ops deployment

  • extra_requirements (list of str, optional) – For custom model deployments: additional python pypi package names to include in the custom environment. Default behavior is to include standard dependencies for the model type

  • hooks (dict of callable, optional) – For custom model deployments: additional hooks to include with the deployment; see the DataRobot User Models documentation for details on supported hooks

  • **kwargs – Additional keyword arguments that may be model specific

Returns:

deployment – Resulting ML Ops deployment; returned immediately and automatically updated asynchronously as the deployment process proceeds

Return type:

Deployment

Examples

scikit-learn pipeline

>>> import sklearn
>>>
>>> pipe : sklearn.pipeline # assumes pipe has been defined & fit elsewhere
>>> deployment_1 = deploy(pipe,
...                       target='my_target',
...                       classes=['my_pos_class', 'my_neg_class'])

scikit-learn pipeline with custom preprocessing hook

>>> import io
>>> import pandas as pd
>>>
>>> df : pd.DataFrame # assumes training data was previously read elsewhere
>>> my_types = df.dtypes
>>> def force_schema(input_binary_data, *args, **kwargs):
...     buffer = io.BytesIO(input_binary_data)
...     return pd.read_csv(buffer, dtype=dict(my_types))
>>>
>>> deployment_2 = deploy(pipe,
...                       target='my_target',
...                       classes=['my_pos_class', 'my_neg_class'],
...                       hooks={'read_input_data': force_schema})