Documentation Experience Platform Data Science Workspace Guide

Train and evaluate a model using the Sensei Machine Learning API

Last update: Wed Apr 12 2023 00:00:00 GMT+0000 (Coordinated Universal Time)

Topics:
Data Science Workspace

CREATED FOR:

Developer
User
Admin
Leader

This tutorial will show you how to create, train, and evaluate a Model using API calls. Refer to this document for a detailed list of API documentation.

Prerequisites

Follow the Import a packaged Recipe using the API for creating an Engine, which is required to train and evaluate a Model using the API.

Follow the Experience Platform API authentication tutorial to start making API calls.

From the tutorial you should now have the following values:

{ACCESS_TOKEN}: Your specific bearer token value provided after authentication.
{ORG_ID}: Your organization credentials found in your unique Adobe Experience Platform integration.
{API_KEY}: Your specific API key value found in your unique Adobe Experience Platform integration.
Link to a Docker image of an intelligent service

API workflow

We will be consuming the APIs to create an Experiment Run for training. For this tutorial, we will be focused on the Engines, MLInstances, and Experiments endpoints. The following chart outlines the relationship between the three and also introduces the idea of a Run and a Model.

NOTE

The terms “Engine”, “MLInstance”, “MLService”, “Experiment”, and “Model” are referred to as different terms in the UI. If you’re coming from the UI, the following table maps the differences.

UI Term

API Term

Recipe

Engine

Model

MLInstance

Training runs

Experiment

Service

MLService

Create an MLInstance

Creating a MLInstance can be done using the following request. You will be using the {ENGINE_ID} that was returned when creating an Engine from the Import a packaged Recipe using the API tutorial.

Request

curl -X POST \
  https://platform.adobe.io/data/sensei/mlInstances \
  -H 'Authorization: Bearer {ACCESS_TOKEN}' \
  -H 'Content-Type: application/vnd.adobe.platform.sensei+json;profile=mlInstance.v1.json' \
  -H 'x-api-key: {API_KEY}' \
  -H 'x-gw-ims-org-id: {ORG_ID}' \
  -d `{JSON_PAYLOAD}`

{ACCESS_TOKEN}: Your specific bearer token value provided after authentication.
{ORG_ID}: Your organization credentials found in your unique Adobe Experience Platform integration.
{API_KEY}: Your specific API key value found in your unique Adobe Experience Platform integration.
{JSON_PAYLOAD}: The configuration of our MLInstance. The example we use in our tutorial is shown here:

{
    "name": "Retail - Instance",
    "description": "Instance for ML Instance",
    "engineId": "{ENGINE_ID}",
    "createdBy": {
        "displayName": "John Doe",
        "userId": "johnd"
    },
    "tags": {
        "purpose": "tutorial"
    },
    "tasks": [
        {
            "name": "train",
            "parameters": [
                {
                    "key": "numFeatures",
                    "value": "10"
                },
                {
                    "key": "maxIter",
                    "value": "2"
                },
                {
                    "key": "regParam",
                    "value": "0.15"
                },
                {
                    "key": "trainingDataLocation",
                    "value": "sample_training_data.csv"
                }
            ]
        },
        {
            "name": "score",
            "parameters": [
                {
                    "key": "scoringDataLocation",
                    "value": "sample_scoring_data.csv"
                },
                {
                    "key": "scoringResultsLocation",
                    "value": "scoring_results.net"
                }
            ]
        }
    ]
}

NOTE

In the {JSON_PAYLOAD}, we define parameters used for training and scoring in the tasks array. The {ENGINE_ID} is the ID of the Engine you want to use and the tag field is an optional parameter used to identify the Instance.

The response contains the {INSTANCE_ID} which represents the MLInstance that is created. Multiple model MLInstances with different configurations can be created.

Response

{
    "id": "{INSTANCE_ID}",
    "name": "Retail - Instance",
    "description": "Instance for ML Instance",
    "engineId": "{ENGINE_ID}",
    "created": "2018-21-21T11:11:11.111Z",
    "createdBy": {
        "displayName": "John Doe",
        "userId": "johnd"
    },
    "updated": "2018-21-01T11:11:11.111Z",
    "deleted": false,
    "tags": {
        "purpose": "tutorial"
    },
    "tasks": [
        {
            "name": "train",
            "parameters": [...]
        },
        {
            "name": "score",
            "parameters": [...]
        }
    ]
}

{ENGINE_ID}: This ID representing the Engine the MLInstance is created under.
{INSTANCE_ID}: The ID that represents the MLInstance.

Create an Experiment

An Experiment is used by a data scientist to arrive at a high performing model while training. Multiple Experiments include changing datasets, features, learning parameters, and hardware. The following is an example of creating an Experiment.

Request

curl -X POST \
  https://platform.adobe.io/data/sensei/experiments \
  -H 'Authorization: Bearer {ACCESS_TOKEN}' \
  -H 'Content-Type: application/vnd.adobe.platform.sensei+json;profile=experiment.v1.json' \
  -H 'x-gw-ims-org-id: {ORG_ID}' \
  -H 'x-api-key: {API_KEY' \
  -d `{JSON PAYLOAD}`

{ORG_ID}: Your organization credentials found in your unique Adobe Experience Platform integration.
{ACCESS_TOKEN}: Your specific bearer token value provided after authentication.
{API_KEY}: Your specific API key value found in your unique Adobe Experience Platform integration.
{JSON_PAYLOAD}: Experiment object that is created. The example we use in our tutorial is shown here:

{
    "name": "Experiment for Retail ",
    "mlInstanceId": "{INSTANCE_ID}",
    "tags": {
        "test": "guide"
    }
}

{INSTANCE_ID}: The ID that represents the MLInstance.

The response from the Experiment creation looks like this.

Response

{
    "id": "{EXPERIMENT_ID}",
    "name": "Experiment for Retail",
    "mlInstanceId": "{INSTANCE_ID}",
    "created": "2018-01-01T11:11:11.111Z",
    "updated": "2018-01-01T11:11:11.111Z",
    "deleted": false,
    "tags": {
        "test": "guide"
    }
}

{EXPERIMENT_ID}: The ID that represents the Experiment you have just created.
{INSTANCE_ID}: The ID that represents the MLInstance.

Create a scheduled Experiment for training

Scheduled Experiments are used so that we do not need to create each single Experiment Runs via an API call. Instead, we provide all necessary parameters during Experiment creation and each run will be created periodically.

To indicate the creation of a scheduled Experiment, we must add a template section in the body of the request. In template, all necessary parameters for scheduling runs are included such as tasks, which indicate what action, and schedule, which indicates the timing of the scheduled runs.

Request

curl -X POST \
  https://platform.adobe.io/data/sensei/experiments \
  -H 'Authorization: Bearer {ACCESS_TOKEN}' \
  -H 'Content-Type: application/vnd.adobe.platform.sensei+json;profile=experiment.v1.json' \
  -H 'x-gw-ims-org-id: {ORG_ID}' \
  -H 'x-api-key: {API_KEY}' \
  -d '{JSON_PAYLOAD}`

{ORG_ID}: Your organization credentials found in your unique Adobe Experience Platform integration.
{ACCESS_TOKEN}: Your specific bearer token value provided after authentication.
{API_KEY}: Your specific API key value found in your unique Adobe Experience Platform integration.
{JSON_PAYLOAD}: Data set to be posted. The example we use in our tutorial is shown here:

{
    "name": "Experiment for Retail",
    "mlInstanceId": "{INSTANCE_ID}",
    "template": {
        "tasks": [{
            "name": "train",
            "parameters": [
                   {
                        "value": "1000",
                        "key": "numFeatures"
                    }
            ],
            "specification": {
                "type": "SparkTaskSpec",
                "executorCores": 5,
                "numExecutors": 5
            }
        }],
        "schedule": {
            "cron": "*/20 * * * *",
            "startTime": "2018-11-11",
            "endTime": "2019-11-11"
        }
    }
}

When we create an Experiment, the body, {JSON_PAYLOAD}, should contain either the mlInstanceId or the mlInstanceQuery parameter. In this example, a scheduled Experiment will invoke a run every 20 minutes, set in the cron parameter, starting from the startTime until the endTime.

Response

{
    "id": "{EXPERIMENT_ID}",
    "name": "Experiment for Retail",
    "mlInstanceId": "{INSTANCE_ID}",
    "created": "2018-11-11T11:11:11.111Z",
    "updated": "2018-11-11T11:11:11.111Z",
    "deleted": false,
    "workflowId": "endid123_0379bc0b_8f7e_4706_bcd9_1a2s3d4f5g_abcdf",
    "template": {
        "tasks": [
            {
                "name": "train",
                "parameters": [...],
                "specification": {
                    "type": "SparkTaskSpec",
                    "executorCores": 5,
                    "numExecutors": 5
                }
            }
        ],
        "schedule": {
            "cron": "*/20 * * * *",
            "startTime": "2018-07-04",
            "endTime": "2018-07-06"
        }
    }
}

{EXPERIMENT_ID}: The ID that represents the Experiment.
{INSTANCE_ID}: The ID that represents the MLInstance.

Create an Experiment Run for training

With an Experiment entity created, a training run can be created and run using the call below. You will need the {EXPERIMENT_ID} and state what mode you want to trigger in the request body.

Request

curl -X POST \
  https://platform.adobe.io/data/sensei/experiments/{EXPERIMENT_ID}/runs \
  -H 'Authorization: Bearer {ACCESS_TOKEN}' \
  -H 'Content-Type: application/vnd.adobe.platform.sensei+json;profile=experimentRun.v1.json' \
  -H 'x-gw-ims-org-id: {ORG_ID}' \
  -H 'x-api-key: {API_KEY}' \
  -d '{JSON_PAYLOAD}'

{EXPERIMENT_ID}: The ID corresponding to the Experiment you want to target. This can be found in the response when creating your Experiment.
{ORG_ID}: Your organization credentials found in your unique Adobe Experience Platform integration.
{ACCESS_TOKEN}: Your specific bearer token value provided after authentication.
{API_KEY}: Your specific API key value found in your unique Adobe Experience Platform integration.
{JSON_PAYLOAD}: To create a training run, you will have to include the following in the body:

{
    "mode":"Train"
}

You can also override the configuration parameters by including a tasks array:

{
   "mode":"Train",
   "tasks": [
        {
           "name": "train",
           "parameters": [
                {
                   "key": "numFeatures",
                   "value": "2"
                }
            ]
        }
    ]
}

You will get the following response which will let you know the {EXPERIMENT_RUN_ID} and the configuration under tasks.

Response

{
    "id": "{EXPERIMENT_RUN_ID}",
    "mode": "train",
    "experimentId": "{EXPERIMENT_ID}",
    "created": "2018-01-01T11:11:11.903Z",
    "updated": "2018-01-01T11:11:11.903Z",
    "deleted": false,
    "tasks": [
        {
            "name": "Train",
            "parameters": [...]
        }
    ]
}

{EXPERIMENT_RUN_ID}: The ID that represents the Experiment Run.
{EXPERIMENT_ID}: The ID that represents the Experiment which the Experiment Run is under.

Retrieve an Experiment Run status

The status of the Experiment run can be queried with the {EXPERIMENT_RUN_ID}.

Request

curl -X GET \
  https://platform.adobe.io/data/sensei/experiments/{EXPERIMENT_ID}/runs/{EXPERIMENT_RUN_ID}/status \
  -H 'Authorization: Bearer {ACCESS_TOKEN}' \
  -H 'x-gw-ims-org-id: {ORG_ID}' \
  -H 'x-api-key: {API_KEY}'

{EXPERIMENT_ID}: The ID that represents the Experiment.
{EXPERIMENT_RUN_ID}: The ID that represents the Experiment Run.
{ACCESS_TOKEN}: Your specific bearer token value provided after authentication.
{ORG_ID}: Your organization credentials found in your unique Adobe Experience Platform integration.
{API_KEY}: Your specific API key value found in your unique Adobe Experience Platform integration.

Response

The GET call will provide the status in the state parameter as shown below:

{
    "id": "{EXPERIMENT_ID}",
    "name": "RunStatus for experimentRunId {EXPERIMENT_RUN_ID}",
    "experimentRunId": "{EXPERIMENT_RUN_ID}",
    "deleted": false,
    "status": {
        "tasks": [
            {
                "id": "{MODEL_ID}",
                "state": "DONE",
                "tasklogs": [
                    {
                        "name": "execution",
                        "url": "https://mlbaprod1sapwd7jzid.file.core.windows.net/..."
                    },
                    {
                        "name": "stderr",
                        "url": "https://mlbaprod1sapwd7jzid.file.core.windows.net/..."
                    },
                    {
                        "name": "stdout",
                        "url": "https://mlbaprod1sapwd7jzid.file.core.windows.net/..."
                    }
                ]
            }
        ]
    }
}

{EXPERIMENT_RUN_ID}: The ID that represents the Experiment Run.
{EXPERIMENT_ID}: The ID that represents the Experiment which the Experiment Run is under.

In addition to the DONE state, other states include:

PENDING
RUNNING
FAILED

To get more information, the detailed logs can be found under the tasklogs parameter.

Retrieve the trained Model

In order to get the trained Model created above during training, we make the following request:

Request

curl -X GET \
  'https://platform.adobe.io/data/sensei/models/?property=experimentRunId=={EXPERIMENT_RUN_ID}' \
  -H 'Authorization: Bearer {ACCESS_TOKEN}' \
  -H 'x-gw-ims-org-id: {ORG_ID}'

{EXPERIMENT_RUN_ID}: The ID corresponding to the Experiment Run you want to target. This can be found in the response when creating your Experiment Run.
{ACCESS_TOKEN}: Your specific bearer token value provided after authentication.
{ORG_ID}: Your organization credentials found in your unique Adobe Experience Platform integration.

The response represents the trained Model that was created.

Response

{
    "children": [
        {
            "id": "{MODEL_ID}",
            "name": "Tutorial trained Model",
            "experimentId": "{EXPERIMENT_ID}",
            "experimentRunId": "{EXPERIMENT_RUN_ID}",
            "description": "trained model for ID",
            "modelArtifactUri": "wasb://test-models@mlpreprodstorage.blob.core.windows.net/{MODEL_ID}",
            "created": "2018-01-01T11:11:11.011Z",
            "updated": "2018-01-01T11:11:11.011Z",
            "deleted": false
        }
    ],
    "_page": {
        "property": "ExperimentRunId=={EXPERIMENT_RUN_ID},deleted!=true",
        "count": 1
    }
}

{MODEL_ID}: The ID corresponding to the Model.
{EXPERIMENT_ID}: The ID corresponding to the Experiment the Experiment Run is under.
{EXPERIMENT_RUN_ID}: The ID corresponding to the Experiment Run.

Stop and delete a scheduled Experiment

If you want to stop execution of a scheduled Experiment before its endTime, this can be done by querying a DELETE request to the {EXPERIMENT_ID}

Request

curl -X DELETE \
  'https://platform.adobe.io/data/sensei/experiments/{EXPERIMENT_ID}' \
  -H 'Authorization: Bearer {ACCESS_TOKEN}' \
  -H 'x-gw-ims-org-id: {ORG_ID}'

{EXPERIMENT_ID}: The ID corresponding to the Experiment.
{ACCESS_TOKEN}: Your specific bearer token value provided after authentication.
{ORG_ID}: Your organization credentials found in your unique Adobe Experience Platform integration.

NOTE

The API call will disable creation of new Experiment runs. However, it will not stop execution of already running Experiment Runs.

The following is the Response notifying that the Experiment is successfully deleted.

Response

{
    "title": "Success",
    "status": 200,
    "detail": "Experiment successfully deleted"
}

Next steps

This tutorial went over how to consume the APIs to create an Engine, an Experiment, scheduled Experiment Runs, and trained Models. In the next exercise, you will be making predictions by scoring a new dataset using the top performing trained model.

recommendation-more-help

cc79fe26-64da-411e-a6b9-5b650f53e4e9