Custom Models

Custom models provide users the ability to run arbitrary modeling code in an environment defined by the user.

Manage Execution Environments

Execution Environment defines the runtime environment for custom models. Execution Environment Version is a revision of Execution Environment with an actual runtime definition. Please refer to DataRobot User Models (https://github.com/datarobot/datarobot-user-models) for sample environments.

Create Execution Environment

To create an Execution Environment run:

import datarobot as dr

execution_environment = dr.ExecutionEnvironment.create(
    name="Python3 PyTorch Environment",
    description="This environment contains Python3 pytorch library.",
)

execution_environment.id
>>> '5b6b2315ca36c0108fc5d41b'

There are 2 ways to create an Execution Environment Version: synchronous and asynchronous.

Synchronous way means that program execution will be blocked until an Execution Environment Version creation process is finished with either success or failure:

import datarobot as dr

# use execution_environment created earlier

environment_version = dr.ExecutionEnvironmentVersion.create(
    execution_environment.id,
    docker_context_path="datarobot-user-models/public_dropin_environments/python3_pytorch",
    max_wait=3600,  # 1 hour timeout
)

environment_version.id
>>> '5eb538959bc057003b487b2d'
environment_version.build_status
>>> 'success'

Asynchronous way means that program execution will be not blocked, but an Execution Environment Version created will not be ready to be used for some time, until its creation process is finished. In such case, it will be required to manually call refresh() for the Execution Environment Version and check if its build_status is “success”. To create an Execution Environment Version without blocking a program, set max_wait to None:

import datarobot as dr

# use execution_environment created earlier

environment_version = dr.ExecutionEnvironmentVersion.create(
    execution_environment.id,
    docker_context_path="datarobot-user-models/public_dropin_environments/python3_pytorch",
    max_wait=None,  # set None to not block execution on this method
)

environment_version.id
>>> '5eb538959bc057003b487b2d'
environment_version.build_status
>>> 'processing'

# after some time
environment_version.refresh()
environment_version.build_status
>>> 'success'

If your environment requires additional metadata to be supplied for models using it, you can create an environment with additional metadata keys. Custom model versions that use this environment must specify values for these keys before they can be used to run tests or make deployments. The values will be baked in as environment variables with field_name as the environment variable name.

import datarobot as dr
from datarobot.models.execution_environment import RequiredMetadataKey

execution_environment = dr.ExecutionEnvironment.create(
    name="Python3 PyTorch Environment",
    description="This environment contains Python3 pytorch library.",
    required_metadata_keys=[
        RequiredMetadataKey(field_name="MY_VAR", display_name="A value needed by hte environment")
    ],
)

model_version = dr.CustomModelVersion.create_clean(
    custom_model_id=custom_model.id,
    base_environment_id=execution_environment.id,
    folder_path=custom_model_folder,
    required_metadata={"MY_VAR": "a value"}
)

List Execution Environments

Use the following command to list execution environments available to the user.

import datarobot as dr

execution_environments = dr.ExecutionEnvironment.list()
execution_environments
>>> [ExecutionEnvironment('[DataRobot] Python 3 PyTorch Drop-In'), ExecutionEnvironment('[DataRobot] Java Drop-In')]

environment_versions = dr.ExecutionEnvironmentVersion.list(execution_environment.id)
environment_versions
>>> [ExecutionEnvironmentVersion('v1')]

Refer to ExecutionEnvironment for properties of the execution environment object and ExecutionEnvironmentVersion for properties of the execution environment object version.

You can also filter the execution environments that are returned by passing a string as search_for parameter - only the execution environments that contain the passed string in name or description will be returned.

import datarobot as dr

execution_environments = dr.ExecutionEnvironment.list(search_for='java')
execution_environments
>>> [ExecutionEnvironment('[DataRobot] Java Drop-In')]

Execution environment versions can be filtered by build status.

import datarobot as dr

environment_versions = dr.ExecutionEnvironmentVersion.list(
    execution_environment.id, dr.EXECUTION_ENVIRONMENT_VERSION_BUILD_STATUS.PROCESSING
)
environment_versions
>>> [ExecutionEnvironmentVersion('v1')]

Retrieve Execution Environment

To retrieve an execution environment and an execution environment version by identifier, rather than list all available ones, do the following:

import datarobot as dr

execution_environment = dr.ExecutionEnvironment.get(execution_environment_id='5506fcd38bd88f5953219da0')
execution_environment
>>> ExecutionEnvironment('[DataRobot] Python 3 PyTorch Drop-In')

environment_version = dr.ExecutionEnvironmentVersion.get(
    execution_environment_id=execution_environment.id, version_id='5eb538959bc057003b487b2d')
environment_version
>>> ExecutionEnvironmentVersion('v1')

Update Execution Environment

To update name and/or description of the execution environment run:

import datarobot as dr

execution_environment = dr.ExecutionEnvironment.get(execution_environment_id='5506fcd38bd88f5953219da0')
execution_environment.update(name='new name', description='new description')

Delete Execution Environment

To delete the execution environment and execution environment version, use the following commands.

import datarobot as dr

execution_environment = dr.ExecutionEnvironment.get(execution_environment_id='5506fcd38bd88f5953219da0')
execution_environment.delete()

Get Execution Environment build log

To get execution environment version build log run:

import datarobot as dr

environment_version = dr.ExecutionEnvironmentVersion.get(
    execution_environment_id='5506fcd38bd88f5953219da0', version_id='5eb538959bc057003b487b2d')
log, error = environment_version.get_build_log()

Manage Custom Models

Custom Inference Model is user-defined modeling code that supports making predictions against it. Custom Inference Model supports regression and binary classification target types.

To upload actual modeling code Custom Model Version must be created for a custom model. Please see Custom Model Version documentation.

Create Custom Inference Model

To create a regression Custom Inference Model run:

import datarobot as dr

custom_model = dr.CustomInferenceModel.create(
    name='Python 3 PyTorch Custom Model',
    target_type=dr.TARGET_TYPE.REGRESSION,
    target_name='MEDV',
    description='This is a Python3-based custom model. It has a simple PyTorch model built on boston housing',
    language='python'
)

custom_model.id
>>> '5b6b2315ca36c0108fc5d41b'

When creating a binary classification Custom Inference Model, positive_class_label and negative_class_label must be set:

import datarobot as dr

custom_model = dr.CustomInferenceModel.create(
    name='Python 3 PyTorch Custom Model',
    target_type=dr.TARGET_TYPE.BINARY,
    target_name='readmitted',
    positive_class_label='False',
    negative_class_label='True',
    description='This is a Python3-based custom model. It has a simple PyTorch model built on 10k_diabetes dataset',
    language='Python 3'
)

custom_model.id
>>> '5b6b2315ca36c0108fc5d41b'

When creating a multiclass classification Custom Inference Model, class_labels must be provided:

import datarobot as dr

custom_model = dr.CustomInferenceModel.create(
    name='Python 3 PyTorch Custom Model',
    target_type=dr.TARGET_TYPE.MULTICLASS,
    target_name='readmitted',
    class_labels=['hot dog', 'burrito', 'hoagie', 'reuben'],
    description='This is a Python3-based custom model. It has a simple PyTorch model built on sandwich dataset',
    language='Python 3'
)

custom_model.id
>>> '5b6b2315ca36c0108fc5d41b'

For convenience when there are many class labels, multiclass labels can also be provided as a file. The file should have all the class labels separated by newline:

import datarobot as dr

custom_model = dr.CustomInferenceModel.create(
    name='Python 3 PyTorch Custom Model',
    target_type=dr.TARGET_TYPE.MULTICLASS,
    target_name='readmitted',
    class_labels_file='/path/to/classlabels.txt',
    description='This is a Python3-based custom model. It has a simple PyTorch model built on sandwich dataset',
    language='Python 3'
)

custom_model.id
>>> '5b6b2315ca36c0108fc5d41b'

For unstructured model target_name parameter is optional and is ignored if provided. To create an unstructured Custom Inference Model run:

import datarobot as dr

custom_model = dr.CustomInferenceModel.create(
    name='Python 3 Unstructured Custom Model',
    target_type=dr.TARGET_TYPE.UNSTRUCTURED,
    description='This is a Python3-based unstructured model',
    language='python'
)

custom_model.id
>>> '5b6b2315ca36c0108fc5d41b'

For anomaly detection models, the target_name parameter is also optional and is ignored if provided. To create an anomaly Custom Inference Model run:

import datarobot as dr

custom_model = dr.CustomInferenceModel.create(
    name='Python 3 Unstructured Custom Model',
    target_type=dr.TARGET_TYPE.ANOMALY,
    description='This is a Python3-based anomaly detection model',
    language='python'
)

custom_model.id
>>> '5b6b2315ca36c0108fc5d41b'

To create a Custom Inference Model with specific k8s resources:

import datarobot as dr

custom_model = dr.CustomInferenceModel.create(
    name='Python 3 PyTorch Custom Model',
    target_type=dr.TARGET_TYPE.BINARY,
    target_name='readmitted',
    positive_class_label='False',
    negative_class_label='True',
    description='This is a Python3-based custom model. It has a simple PyTorch model built on 10k_diabetes dataset',
    language='Python 3',
    maximum_memory=512*1024*1024,
)

Custom Inference Model k8s resources are optional and unless specifically provided, the configured defaults will be used.

To create a Custom Inference Model enabling training data assignment on the model version level, provide the is_training_data_for_versions_permanently_enabled=True parameter. For more information, refer to the Custom model version creation with training data documentation.

import datarobot as dr

custom_model = dr.CustomInferenceModel.create(
    name='Python 3 PyTorch Custom Model',
    target_type=dr.TARGET_TYPE.REGRESSION,
    target_name='MEDV',
    description='This is a Python3-based custom model. It has a simple PyTorch model built on boston housing',
    language='python',
    is_training_data_for_versions_permanently_enabled=True
)

custom_model.id
>>> '5b6b2315ca36c0108fc5d41b'

List Custom Inference Models

Use the following command to list Custom Inference Models available to the user:

import datarobot as dr

dr.CustomInferenceModel.list()
>>> [CustomInferenceModel('my model 2'), CustomInferenceModel('my model 1')]

# use these parameters to filter results:
dr.CustomInferenceModel.list(
    is_deployed=True,  # set to return only deployed models
    order_by='-updated',  # set to define order of returned results
    search_for='model 1',  # return only models containing 'model 1' in name or description
)
>>> CustomInferenceModel('my model 1')

Please refer to list() for detailed parameter description.

Retrieve Custom Inference Model

To retrieve a specific Custom Inference Model, run:

import datarobot as dr

dr.CustomInferenceModel.get('5ebe95044024035cc6a65602')
>>> CustomInferenceModel('my model 1')

Update Custom Model

To update Custom Inference Model properties execute the following:

import datarobot as dr

custom_model = dr.CustomInferenceModel.get('5ebe95044024035cc6a65602')

custom_model.update(
    name='new name',
    description='new description',
)

Please, refer to update() for the full list of properties that can be updated.

Download latest revision of Custom Inference Model

To download content of the latest Custom Model Version of CustomInferenceModel as a ZIP archive:

import datarobot as dr

path_to_download = '/home/user/Documents/myModel.zip'

custom_model = dr.CustomInferenceModel.get('5ebe96b84024035cc6a6560b')

custom_model.download_latest_version(path_to_download)

Assign training data to a custom inference model

This example assigns training data on the model level. To assign training data on the model version level, see the Custom model version creation with training data documentation.

To assign training data to custom inference model, run:

import datarobot as dr

path_to_dataset = '/home/user/Documents/trainingDataset.csv'
dataset = dr.Dataset.create_from_file(file_path=path_to_dataset)

custom_model = dr.CustomInferenceModel.get('5ebe96b84024035cc6a6560b')

custom_model.assign_training_data(dataset.id)

To assign training data without blocking a program, set max_wait to None:

import datarobot as dr

path_to_dataset = '/home/user/Documents/trainingDataset.csv'
dataset = dr.Dataset.create_from_file(file_path=path_to_dataset)

custom_model = dr.CustomInferenceModel.get('5ebe96b84024035cc6a6560b')

custom_model.assign_training_data(
    dataset.id,
    max_wait=None
)

custom_model.training_data_assignment_in_progress
>>> True

# after some time
custom_model.refresh()
custom_model.training_data_assignment_in_progress
>>> False

Note: training data must be assigned to retrieve feature impact from a custom model version. See the Custom Model Version documentation.

Manage Custom Model Versions

Modeling code for Custom Inference Models can be uploaded by creating a Custom Model Version. When creating a Custom Model Version, the version must be associated with a base execution environment. If the base environment supports additional model dependencies (R or Python environments) and the Custom Model Version contains a valid requirements.txt file, the model version will run in an environment based on the base environment with the additional dependencies installed.

Create Custom Model Version

Upload actual custom model content by creating a clean Custom Model Version:

import os
import datarobot as dr

custom_model_folder = "datarobot-user-models/model_templates/python3_pytorch"

# add files from the folder to the custom model
model_version = dr.CustomModelVersion.create_clean(
    custom_model_id=custom_model.id,
    base_environment_id=execution_environment.id,
    folder_path=custom_model_folder,
)

custom_model.id
>>> '5b6b2315ca36c0108fc5d41b'

# or add a list of files to the custom model
model_version_2 = dr.CustomModelVersion.create_clean(
    custom_model_id=custom_model.id,
    base_environment_id=execution_environment.id,
    files=[(os.path.join(custom_model_folder, 'custom.py'), 'custom.py')],
)

# and/or set k8s resources to the custom model
model_version_3 = dr.CustomModelVersion.create_clean(
    custom_model_id=custom_model.id,
    base_environment_id=execution_environment.id,
    files=[(os.path.join(custom_model_folder, 'custom.py'), 'custom.py')],
    network_egress_policy=dr.NETWORK_EGRESS_POLICY.PUBLIC,
    maximum_memory=512*1024*1024,
    replicas=1,
)

To create a new Custom Model Version from a previous one, with just some files added or removed, do the following:

import os
import datarobot as dr

custom_model_folder = "datarobot-user-models/model_templates/python3_pytorch"

file_to_delete = model_version_2.items[0].id

model_version_3 = dr.CustomModelVersion.create_from_previous(
    custom_model_id=custom_model.id,
    base_environment_id=execution_environment.id,
    files=[(os.path.join(custom_model_folder, 'custom.py'), 'custom.py')],
    files_to_delete=[file_to_delete],
)

Please refer to CustomModelFileItem for description of custom model file properties.

Specify a custom environment version when creating a custom model version. By default a version of the same environment does not change between consecutive model versions. This behavior can be overridden:

import os
import datarobot as dr

custom_model_folder = "datarobot-user-models/model_templates/python3_pytorch"

# create a clean version and specify an explicit environment version.
model_version = dr.CustomModelVersion.create_clean(
    custom_model_id=custom_model.id,
    base_environment_id=execution_environment.id,
    base_environment_version_id="642209acc5638929a9b8dc3d",
    folder_path=custom_model_folder,
)

# create a version from a previous one, specify an explicit environment version.
model_version_2 = dr.CustomModelVersion.create_from_previous(
    custom_model_id=custom_model.id,
    base_environment_id=execution_environment.id,
    base_environment_version_id="660186775d016eabb290aee9",
)

To create a new Custom Model Version from a previous one, with just new k8s resources values, do the following:

import os
import datarobot as dr

custom_model_folder = "datarobot-user-models/model_templates/python3_pytorch"

file_to_delete = model_version_2.items[0].id

model_version_3 = dr.CustomModelVersion.create_from_previous(
    custom_model_id=custom_model.id,
    base_environment_id=execution_environment.id,
    maximum_memory=1024*1024*1024,
)

Create a custom model version with training data

Model version creation allows to provide training (and holdout) data information. Every custom model has to be explicitly switched to allow training data assignment for model versions. Note that the training data assignment differs for structured and unstructured models, and should be handled differently.

Enable training data assignment for custom model versions

By default, custom model training data is assigned on the model level; for more information, see the Custom model training data assignment documentation. When training data is assigned to a model, the same training data is used for every model version. This method of training data assignment is deprecated and scheduled for removal; however, to avoid introducing issues for existing models, you must individually convert existing models to perform training data assignment by model version. This change is permanent and can not be undone. Because the conversion process is irreversible, it is highly recommended that you do not convert critical models to the new training data assignment method. Instead, you should duplicate the existing model and test the new method.

To permanently enable a training data assignment on the model version level for the specified model, do the following: .. code-block:: python

import datarobot as dr

dr.Client(token=my_token, endpoint=endpoint)

custom_model = dr.CustomInferenceModel.get(custom_model_id)

custom_model.update(is_training_data_for_versions_permanently_enabled=True) custom_model.is_training_data_for_versions_permanently_enabled >>> True

Assign training data for structured models

Assign training data for structured models, you can provide the parameters training_dataset_id and partition_column. Training data assignment is performed asynchronously, so you can create a version in a blocking or non-blocking way (see examples).

Create a structured model version with blocking (default max_wait=600) and wait for the training data assignment result.

If the training data assignment fails:
  • a datarobot.errors.TrainingDataAssignmentError exception is raised. The exception contains the custom model ID, the custom model version ID, the failure message.

  • a new custom model version is still created and can be fetched for further processing, but it’s not possible to create a model package from it or deploy it.

import datarobot as dr
from datarobot.errors import TrainingDataAssignmentError

dr.Client(token=my_token, endpoint=endpoint)

try:
    version = dr.CustomModelVersion.create_from_previous(
        custom_model_id="6444482e5583f6ee2e572265",
        base_environment_id="642209acc563893014a41e24",
        training_dataset_id="6421f2149a4f9b1bec6ad6dd",
    )
except TrainingDataAssignmentError as e:
    print(e)

Fetching model version in the case of the assignment error, example 1:

import datarobot as dr
from datarobot.errors import TrainingDataAssignmentError

dr.Client(token=my_token, endpoint=endpoint)

try:
    version = dr.CustomModelVersion.create_from_previous(
        custom_model_id="6444482e5583f6ee2e572265",
        base_environment_id="642209acc563893014a41e24",
        training_dataset_id="6421f2149a4f9b1bec6ad6dd",
    )
except TrainingDataAssignmentError as e:
    version = CustomModelVersion.get(
        custom_model_id="6444482e5583f6ee2e572265",
        custom_model_version_id=e.custom_model_version_id,
    )
    print(version.training_data.dataset_id)
    print(version.training_data.dataset_version_id)
    print(version.training_data.dataset_name)
    print(version.training_data.assignment_error)

Fetching model version in the case of the assignment error, example 2:

import datarobot as dr
from datarobot.errors import TrainingDataAssignmentError

dr.Client(token=my_token, endpoint=endpoint)
custom_model = dr.CustomInferenceModel.get("6444482e5583f6ee2e572265")

try:
    version = dr.CustomModelVersion.create_from_previous(
        custom_model_id="6444482e5583f6ee2e572265",
        base_environment_id="642209acc563893014a41e24",
        training_dataset_id="6421f2149a4f9b1bec6ad6dd",
    )
except TrainingDataAssignmentError as e:
    pass

custom_model.refresh()
version = custom_model.latest_version
print(version.training_data.dataset_id)
print(version.training_data.dataset_version_id)
print(version.training_data.dataset_name)
print(version.training_data.assignment_error)

Create a structured model version with a non-blocking (set max_wat=None) training data assignment. In this case, it is the user’s responsibility to poll for version.training_data.assignment_in_progress. Once the assignment is finished, check for errors if version.training_data.assignment_in_progress==False. If version.training_data.assignment_error is None, then there is no error.

import datarobot as dr

dr.Client(token=my_token, endpoint=endpoint)

version = dr.CustomModelVersion.create_from_previous(
    custom_model_id="6444482e5583f6ee2e572265",
    base_environment_id="642209acc563893014a41e24",
    training_dataset_id="6421f2149a4f9b1bec6ad6dd",
    max_wait=None,
)

while version.training_data.assignment_in_progress:
    time.sleep(10)
    version.refresh()
if version.training_data.assignment_error:
    print(version.training_data.assignment_error["message"])

Assign training data for unstructured models

For unstructured models: you can provide the parameters training_dataset_id and holdout_dataset_id. The training data assignment is performed synchronously and the max_wait parameter is ignored.

The example below shows how to create an unstructured model version with training and holdout data.

import datarobot as dr

dr.Client(token=my_token, endpoint=endpoint)

version = dr.CustomModelVersion.create_from_previous(
    custom_model_id="6444482e5583f6ee2e572265",
    base_environment_id="642209acc563893014a41e24",
    training_dataset_id="6421f2149a4f9b1bec6ad6dd",
    holdout_dataset_id="6421f2149a4f9b1bec6ad6ef",
)
if version.training_data.assignment_error:
    print(version.training_data.assignment_error["message"])

Remove training data

By default, training and holdout data are copied to a new model version from the previous model version. If you don’t want to keep training and holdout data for the new version, set keep_training_holdout_data to False.

import datarobot as dr

dr.Client(token=my_token, endpoint=endpoint)

version = dr.CustomModelVersion.create_from_previous(
    custom_model_id="6444482e5583f6ee2e572265",
    base_environment_id="642209acc563893014a41e24",
    keep_training_holdout_data=False,
)

List Custom Model Versions

Use the following command to list Custom Model Versions available to the user:

import datarobot as dr

dr.CustomModelVersion.list(custom_model.id)

>>> [CustomModelVersion('v2.0'), CustomModelVersion('v1.0')]

Retrieve Custom Model Version

To retrieve a specific Custom Model Version, run:

import datarobot as dr

dr.CustomModelVersion.get(custom_model.id, custom_model_version_id='5ebe96b84024035cc6a6560b')

>>> CustomModelVersion('v2.0')

Update Custom Model Version

To update Custom Model Version description execute the following:

import datarobot as dr

custom_model_version = dr.CustomModelVersion.get(
    custom_model.id,
    custom_model_version_id='5ebe96b84024035cc6a6560b',
)

custom_model_version.update(description='new description')

custom_model_version.description
>>> 'new description'

Download Custom Model Version

Download content of the Custom Model Version as a ZIP archive:

import datarobot as dr

path_to_download = '/home/user/Documents/myModel.zip'

custom_model_version = dr.CustomModelVersion.get(
    custom_model.id,
    custom_model_version_id='5ebe96b84024035cc6a6560b',
)

custom_model_version.download(path_to_download)

Start Custom Model Inference Legacy Conversion

Custom model version may include SAS files, with a main program entrypoint. In order to be able to use this model it is required to run a conversion. The conversion can later be fetched and examined by reading the conversion print-outs. By default, a conversion is initiated in a non-blocking mode. If a max_wait parameter is provided, than the call is blocked until the conversion is completed. The results can than be read by fetching the conversion entity.

import datarobot as dr

    # Read a custom model version
    custom_model_version = dr.CustomModelVersion.get(model_id, model_version_id)

    # Find the main program item ID
    main_program_item_id = None
    for item in cm_ver.items:
            if item.file_name.lower().endswith('.sas'):
                    main_program_item_id = item.id

    # Execute the conversion
    if async:
            # This is a non-blocking call
            conversion_id = dr.models.CustomModelVersionConversion.run_conversion(
                    custom_model_version.custom_model_id,
                    custom_model_version.id,
                    main_program_item_id,
            )
    else:
            # This call is blocked until a completion or a timeout
            conversion_id = dr.models.CustomModelVersionConversion.run_conversion(
                    custom_model_version.custom_model_id,
                    custom_model_version.id,
                    main_program_item_id,
                    max_wait=60,
            )

Monitor Custom Model Inference Legacy Conversion Process

If a custom model version conversion was initiated in a non-blocking mode, it is possible to monitor the progress as follows:

import datarobot as dr

    while True:
            conversion = dr.models.CustomModelVersionConversion.get(
                    custom_model_id, custom_model_version_id, conversion_id,
            )
            if conversion.conversion_in_progress:
                    logging.info('Conversion is in progress...')
                    time.sleep(1)
            else:
                    if conversion.conversion_succeeded:
                            logging.info('Conversion succeeded')
                    else:
                            logging.error(f'Conversion failed!\n{conversion.log_message}')
                    break

Stop a Custom Model Inference Legacy Conversion

It is possible to stop a custom model version conversion that is in progress. The call is non-blocking and you may keep monitoring the conversion progress (see above) until is it completed.

import datarobot as dr

    dr.models.CustomModelVersionConversion.stop_conversion(
            custom_model_id, custom_model_version_id, conversion_id,
    )

Calculate Custom ModelVersion feature impact

To trigger calculation of custom model version Feature Impact, training data must be assigned to a custom inference model. Please refer to the custom inference model documentation. If training data is assigned, run the following to trigger the calculation of the feature impact:

import datarobot as dr

version = dr.CustomModelVersion.get(custom_model.id, custom_model_version_id='5ebe96b84024035cc6a6560b')

version.calculate_feature_impact()

To trigger calculating feature impact without blocking a program, set max_wait to None:

import datarobot as dr

version = dr.CustomModelVersion.get(custom_model.id, custom_model_version_id='5ebe96b84024035cc6a6560b')

version.calculate_feature_impact(max_wait=None)

Retrieve Custom Inference Image feature impact

To retrieve Custom Model Version feature impact, it must be calculated beforehand. Please refer to Custom Inference Image feature impact documentation. Run the following to get feature impact:

import datarobot as dr

version = dr.CustomModelVersion.get(custom_model.id, custom_model_version_id='5ebe96b84024035cc6a6560b')

version.get_feature_impact()
>>> [{'featureName': 'B', 'impactNormalized': 1.0, 'impactUnnormalized': 1.1085356209402688, 'redundantWith': 'B'}...]

Preparing a Custom Model Version for Use

If your custom model version has dependencies, a dependency build must be completed before the model can be used. The dependency build installs your model’s dependencies into the base environment associated with the model version.

Starting the Dependency Build

To start the Custom Model Version Dependency Build, run:

import datarobot as dr

build_info = dr.CustomModelVersionDependencyBuild.start_build(
    custom_model_id=custom_model.id,
    custom_model_version_id=model_version.id,
    max_wait=3600,  # 1 hour timeout
)

build_info.build_status
>>> 'success'

To start Custom Model Version Dependency Build without blocking a program until the test finishes, set max_wait to None:

import datarobot as dr

build_info = dr.CustomModelVersionDependencyBuild.start_build(
    custom_model_id=custom_model.id,
    custom_model_version_id=model_version.id,
    max_wait=None,
)

build_info.build_status
>>> 'submitted'

# after some time
build_info.refresh()
build_info.build_status
>>> 'success'

In case the build fails, or you are just curious, do the following to retrieve the build log once complete:

print(build_info.get_log())

To cancel a Custom Model Version Dependency Build, simply run:

build_info.cancel()

Manage Custom Model Tests

A Custom Model Test represents testing performed on custom models.

Create Custom Model Test

To create Custom Model Test, run:

import datarobot as dr

path_to_dataset = '/home/user/Documents/testDataset.csv'
dataset = dr.Dataset.create_from_file(file_path=path_to_dataset)

custom_model_test = dr.CustomModelTest.create(
    custom_model_id=custom_model.id,
    custom_model_version_id=model_version.id,
    dataset_id=dataset.id,
    max_wait=3600,  # 1 hour timeout
)

custom_model_test.overall_status
>>> 'succeeded'

or, with k8s resources:

import datarobot as dr

path_to_dataset = '/home/user/Documents/testDataset.csv'
dataset = dr.Dataset.create_from_file(file_path=path_to_dataset)

custom_model_test = dr.CustomModelTest.create(
    custom_model_id=custom_model.id,
    custom_model_version_id=model_version.id,
    dataset_id=dataset.id,
    max_wait=3600,  # 1 hour timeout
    maximum_memory=1024*1024*1024,
)

custom_model_test.overall_status
>>> 'succeeded'

To start Custom Model Test without blocking a program until the test finishes, set max_wait to None:

import datarobot as dr

path_to_dataset = '/home/user/Documents/testDataset.csv'
dataset = dr.Dataset.create_from_file(file_path=path_to_dataset)

custom_model_test = dr.CustomModelTest.create(
    custom_model_id=custom_model.id,
    custom_model_version_id=model_version.id,
    dataset_id=dataset.id,
    max_wait=None,
)

custom_model_test.overall_status
>>> 'in_progress'

# after some time
custom_model_test.refresh()
custom_model_test.overall_status
>>> 'succeeded'

Running a Custom Model Test uses the Custom Model Version’s base image with its dependencies installed as an execution environment. To start Custom Model Test using an execution environment “as-is”, without the model’s dependencies installed, supply an environment ID and (optionally) and environment version ID:

import datarobot as dr

path_to_dataset = '/home/user/Documents/testDataset.csv'
dataset = dr.Dataset.create_from_file(file_path=path_to_dataset)

custom_model_test = dr.CustomModelTest.create(
    custom_model_id=custom_model.id,
    custom_model_version_id=model_version.id,
    dataset_id=dataset.id,
    max_wait=3600,  # 1 hour timeout
)

custom_model_test.overall_status
>>> 'succeeded'

In case a test fails, do the following to examine details of the failure:

for name, test in custom_model_test.detailed_status.items():
    print('Test: {}'.format(name))
    print('Status: {}'.format(test['status']))
    print('Message: {}'.format(test['message']))

print(custom_model_test.get_log())

To cancel a Custom Model Test, simply run:

custom_model_test.cancel()

To start Custom Model Test for an unstructured custom model, dataset details should not be provided:

import datarobot as dr

custom_model_test = dr.CustomModelTest.create(
    custom_model_id=custom_model.id,
    custom_model_version_id=model_version.id,
)

List Custom Model Tests

Use the following command to list Custom Model Tests available to the user:

import datarobot as dr

dr.CustomModelTest.list(custom_model_id=custom_model.id)
>>> [CustomModelTest('5ec262604024031bed5aaa16')]

Retrieve Custom Model Test

To retrieve a specific Custom Model Test, run:

import datarobot as dr

dr.CustomModelTest.get(custom_model_test_id='5ec262604024031bed5aaa16')
>>> CustomModelTest('5ec262604024031bed5aaa16')