Deployment Predictions

Deployment predictions uses the DataRobot Python Client: datarobot to handle authentication and making request to the DataRobot API. The package is automatically installed as a dependency of datarobot-predict.

The following example shows how to authenticate using an API token. For other authentication examples and usage info, take a look at the DataRobot API quickstart. More information about the DataRobot Prediction API can be found in the DataRobot docs.

To compute deployment predictions, instantiate a datarobot.Deployment and pass it to the predict function

import datarobot as dr
from datarobot_predict.deployment import predict

dr.Client(
    endpoint='https://app.datarobot.com/api/v2',
    token='NjE3ZjA3Mzk0MmY0MDFmZGFiYjQ0MztergsgsQwOk9G'
)

deployment = dr.Deployment.get(deployment_id="635f85940ca2497cb57a54b8")
result_df, response_headers = predict(deployment, input_df)

The return type of the predict function is PredictionResult which is a named tuple containing properties dataframe and response_headers. The dataframe property contains prediction results while response_headers makes it possible to retrieve various metadata like the model id associated with the deployment:

result = predict(deployment, input_df)
predictions = result.dataframe
model_id = result.response_headers["X-DataRobot-Model-Id"]

Prediction Explanations

To compute explanations, set max_explanations to a positive value

df_with_explanations, _ = predict(
    deployment,
    df,
    max_explanations=3
)

If max_explanations is 0 and explanation_algorithm is set to shap, all explanations will be computed

df_with_explanations, _ = predict(
    deployment,
    df,
    explanation_algorithm="shap"
)

Time Series

Forecast point predictions are returned by default if no other arguments are provided for a Time Series Model. The forecast point can be specified using the forecast_point parameter or auto-detected.

result_df, _ = predict(
    deployment,
    df,
    forecast_point=datetime.datetime(1958, 6, 1)
)

To do historical predictions, set time_series_type accordingly

from datarobot_predict import TimeSeriesType

result_df, _ = predict(
    deployment,
    df,
    time_series_type=TimeSeriesType.HISTORICAL,
    predictions_start_date=datetime.datetime(2020, 1, 1),
    predictions_end_date=datetime.datetime(2022, 6, 1),
)

Unstructured models

Predictions can be computed for unstructured custom models using the function predict_unstructured. The following example shows how to make a request to a unstructured deployment with a binary file as the argument.

with open("image.jpg", "rb") as file:
    content, response_headers = predict_unstructured(
        deployment,
        data=file,
        content_type="image/jpeg"
    )

It is also possible to use DataFrame input/output for custom models which use csv input/output. Simply, pass a DataFrame as the data argument:

result_df, response_headers = predict_unstructured(
        deployment,
        data=input_df
  )

Using a custom prediction endpoint

It is possible to force a specific prediction endpoint for the predict function using the prediction_endpoint parameter. This can be used in an enterprise installation that has multiple prediction servers or together with the Portable Prediction Server(PPS).

This example shows how to make a request to a PPS using single-model mode

result_df, _ = predict(
    deployment=None,
    data_frame=df,
    prediction_endpoint="http://localhost:8080"
)

The deployment parameter is set to None since there is no deployment id associated with PPS single-model mode.

For PPS multi-model mode, deployment should be set to the sub directory where the model package is located

result_df, _ = predict(
    deployment="my_model",
    data_frame=df,
    prediction_endpoint="http://localhost:8080"
)

For more information about PPS, see the DataRobot docs.

Limitations

• Requests will timeout after 600 seconds.
• The input DataFrame converted to csv has be less than 50MB or an exception will be raised.

For large and long-running requests, consider using the Batch Prediction API.