Watson Machine Learning v3 / v4 beta to v4 migration details

This is a migration guide for developers to migrate their existing assets from WML v3/v4 beta to WML v4.

Table of contents:

1. Specifying credentials
2. Creating a client object
3. Creating deployment space and setting up default deployement space for your project
4. Specifying metadata for storing the model
5. Storing a model (remains same)
6. Specifying metadata for deployment (remains same)
7. Specifying input payload for prediction (remains same)
8. Getting predictions (remains same)

1. Specifying credentials

OLD: Credentials were generated from IBM Cloud > Resources > WML Instance > Service Credentials

wml_credentials = {
    "apikey":"",
    "url": "",
    "instance_id": ""
}

NEW: Apikey has to be generated from IBM Cloud > Resources > WML Instance > Manage > Access (IAM) > API keys > Create an IBM Cloud API key

NOTE:

1. instance_id is no longer required.
2. url will change based on the region where the WML Instance is deployed in your IBM Cloud, Example:
   • London: https://eu-gb.ml.cloud.ibm.com
   • Dallas: https://us-south.ml.cloud.ibm.com

apikey = "<APIKEY_FROM_IAM>"

wml_credentials = {
    "apikey": apikey,
    "url": "https://eu-gb.ml.cloud.ibm.com"
}

2. Creating a client object

OLD: The watson_machine_learning_client and WatsonMachineLearningAPIClient libraries were used in v3 / v4 beta.

from watson_machine_learning_client import WatsonMachineLearningAPIClient

client = WatsonMachineLearningAPIClient(wml_credentials)

NEW: The ibm_watson_machine_learning and APIClient libraries have to be used in v4.

!pip install -U ibm-watson-machine-learning

from ibm_watson_machine_learning import APIClient

client = APIClient(wml_credentials)

3. Creating deployment space and setting up default deployement space for your project

NEW: A deployment space has to be created in Watson Studio / Cloud Pak for Data.

• In IBM Cloud, goto resources and select Watson Studio instance, click on Get Started to launch Watson Studio / Cloud Pak for Data.

• In Watson Studio / Cloud Pak for Data, click on the hamburger menu on the top left corner and select Deployment spaces > View all spaces.

• In deployment spaces, click on New deployment space +.

• Select Create an empty space when prompted.

• Make sure you select the appropriate Cloud object storage service as well as Machine learning service.

NOTE: In v4 Machine learning assets are stored in Cloud Object Storage rather than in the Watson Machine Learning repository.

• Once the deployment space is created, click on View Space to view the details.

• Click on Settings and copy the space ID.

Learn more about deployment space here.

space_id = "<DEPLOYMENT_SPACE_ID_FROM_WATSON_STUDIO>"

NEW: The deployment space id that you specified in the above step has to be set as the default deployment space for the project.

client.spaces.list(limit=10)

client.set.default_space(space_id)

4. Specifying metadata for storing the model

OLD: RUNTIME_UID was being used to specify the runtime of a model in v3 / v4 beta.

metadata = {
    client.repository.ModelMetaNames.NAME: 'Personal Loan Prediction Model',
    client.repository.ModelMetaNames.TYPE: "scikit-learn_0.20",
    client.repository.ModelMetaNames.RUNTIME_UID: "scikit-learn_0.20-py3.6"  
}

NEW: SOFTWARE_SPEC_UID has to be specified in v4.

sofware_spec_uid = client.software_specifications.get_id_by_name("default_py3.7")

metadata = {
    client.repository.ModelMetaNames.NAME: 'Personal Loan Prediction model',
    client.repository.ModelMetaNames.TYPE: 'scikit-learn_0.23',
    client.repository.ModelMetaNames.SOFTWARE_SPEC_UID: sofware_spec_uid
}

Optionally, to get a list of software specifications run the following:

client.software_specifications.list()

5. Storing a model (remains same)

published_model = client.repository.store_model(model=naive_model, meta_props=metadata, \
                                                training_data=train_set, training_target=train_labels)

import json
published_model_uid = client.repository.get_model_uid(published_model)
model_details = client.repository.get_details(published_model_uid)

print(json.dumps(model_details, indent=2))

6. Specifying metadata for deploying a model (remains same)

deploy_meta = {
     client.deployments.ConfigurationMetaNames.NAME: 'Deployment of Personal Loan Prediction',
     client.deployments.ConfigurationMetaNames.ONLINE: {}
}

created_deployment = client.deployments.create(published_model_uid, meta_props=deploy_meta)

deployments = client.deployments.get_details()

7. Specifying input payload for prediction (remains same)

job_payload = {
client.deployments.ScoringMetaNames.INPUT_DATA: [{
 'values': [[39,139,95616,3,3.4,1,483,0,0,1,0]]
}]
}

8. Getting predictions (remains same)

did = client.deployments.get_uid(created_deployment)

predictions = client.deployments.score(did, job_payload)

print(json.dumps(predictions, indent=2))

print(predictions["predictions"][0]['values'][0][0])

References

• Watson Machine Learning Migrating assets.

• Watson Machine Learning Python Client Documentation.

• Watson Machine Learning authentication.