TrainingJob.deploy method | Google Vertex AI teradataml Extension Library - TrainingJob.deploy Method - Teradata Vantage

Teradata® VantageCloud Lake
Deployment
VantageCloud
Edition
Lake
Product
Teradata Vantage
Published
January 2023
ft:locale
en-US
ft:lastEdition
2024-12-11
dita:mapPath
phg1621910019905.ditamap
dita:ditavalPath
pny1626732985837.ditaval
dita:id
phg1621910019905
Use the TrainingJob.deploy method to deploy a trained model in Google Vertex AI environment or in Vantage system. A TDPredictor object is returned which can be used to run prediction.
  • For Vertex AI environment, it requires argument format the same as the Model.deploy function in Vertex AI Python SDK.
  • For Vantage system, it requires arguments that are required by BYOM and Predictor functions in teradataml.
  • You should have registered the model files in Vertex AI before running this method.
  • This method should be called after the fit method, that is, model training has completed.
Required Arguments:
  • model: Specifies a Vertex AI model object.
  • platform: Specifies the platform to deploy the given model:
    • 'vantage'
    • 'vx-endpoint'
  • model_type: Required if platform is 'vantage'. Specifies the type of the model:
    • 'pmml'
    • 'onnx'
    • 'h2o'
  • model_filename: Required if platform is 'vantage'. Specifies the Google Cloud Storage file name of the model artifact.
Optional Arguments:
  • save_byom_kwargs: Specifies the keyword arguments of teradataml save_byom() function.

    If neither table_name is provided nor BYOM catalog set using set_byom_catalog, a table with name 'tdapiclient_byom_models' will be created in the current schema.

  • retrieve_byom_kwargs: Specifies the keyword arguments of teradataml retrieve_byom() function.
  • vertex_kwargs: Specifies any additional arguments required by TrainingJob.run method.

    These parameters are directly supplied to TrainingJob.run method.

Example 1: Use Vertex AI CustomTrainingJob with Vantage DataFrame

  • Import required packages.
    from tdapiclient import create_tdapi_context, TDApiClient
from teradataml import DataFrame
  • Create TDApiClient class.
    tdapi_context = create_tdapi_context('gcp', gcp_bucket_name="tdapiclient", gcp_bucket_path="/tmp/")
tdapi_client = TDApiClient(tdapi_context)
  • Create teradataml DataFrame for model training.
    input_df = DataFrame(table_name="inputTable")
  • Create training job with TDApiClient.
    job = tdapi_client.CustomTrainingJob(
    display_name="custom-train",
    script_path="train.py",
    container_uri="<TRAINING_IMAGE_URI>",
    model_serving_container_image_uri="<PREDICTION_IMAGE_URI>",
)
  • Submit training job.
    model = job.fit(input_df, replica_count=1, model_display_name="custom-model")
  • Call model.deploy method to deploy the model in a Vertex AI endpoint.
    td_predictor = job.deploy(model, platform='vertex-endpoint')
  • Run predict using the deployed model in UDF mode.
    df =  DataFrame(query="select feature1, feature2 from tbl")
output = td_predictor.predict(df, mode='udf')

Example 2: Use Vertex AI without TDApiClient

  • Import required packages.
    import os
from google.cloud import aiplatform
  • Define Vertex AI constants.
    PROJECT_ID = 'gcp-project'
LOCATION = 'us-central1'
TRAINING_IMAGE = "us-docker.pkg.dev/vertex-ai/training/tf-cpu.2-9:latest"
PREDICTION_IMAGE = "us-docker.pkg.dev/vertex-ai/prediction/tf2-cpu.2-9:latest"
BUCKET_ROOT = "gs://tdapiclient-vertex-test"
  • Define the credentials to access Google Cloud - the service account JSON file.
    os.environ["GOOGLE_APPLICATION_CREDENTIALS"] = "local/path/to/credentials.json"
  • Initialize Vertex AI client.
    aiplatform.init(
    project=PROJECT_ID,
    location=LOCATION,
    staging_bucket=BUCKET_ROOT)
  • Create a Managed Tabular Dataset from CSV file.
    ds = aiplatform.TabularDataset.create(
    display_name="abalone",
    gcs_source=[f"{BUCKET_ROOT}/data/abalone_train.csv"])
  • Launch a training job to create a Model.
    job = aiplatform.CustomTrainingJob(
    display_name="custom-train",
    script_path="train.py",
    container_uri=TRAINING_IMAGE,
    requirements=["gcsfs==0.7.1", "tf2onnx"],
    model_serving_container_image_uri=PREDICTION_IMAGE,
)
    model = job.run(ds, replica_count=1, model_display_name="abalone-model")
  • Deploy the Model.
    endpoint = model.deploy(machine_type="n1-standard-4")
    endpoint.predict(
    [
        [0.435, 0.335, 0.11, 0.33399999999999996, 0.1355, 0.0775, 0.0965],
        [0.585, 0.45, 0.125, 0.874, 0.3545, 0.2075, 0.225],
    ]
)

Example 3: Deploy Vertex AI model in Vantage through BYOM mode

  • Import required packages.
    from tdapiclient import create_tdapi_context, TDApiClient
from teradataml import DataFrame
  • Create TDApiClient class.
    tdapi_context = create_tdapi_context('gcp', gcp_bucket_name="tdapiclient", gcp_bucket_path="/tmp/")
tdapi_client = TDApiClient(tdapi_context)
  • Create teradataml DataFrames for model training and testing.
    train_df = DataFrame(table_name="input_train_table")
test_df = DataFrame(table_name="input_test_table")
  • Create training job with TDApiClient.
    job = tdapi_client.CustomTrainingJob(
    display_name="custom-train",
    script_path="train.py",
    container_uri="<TRAINING_IMAGE_URI>",
    model_serving_container_image_uri="<PREDICTION_IMAGE_URI>",
)
  • Submit training job.
    model = job.fit(train_df, replica_count=1, model_display_name="custom-model")
  • Model.deploy method uses model artifact found in Cloud Storage after training.
    td_predictor = job.deploy(model, "vantage", model_type="pmml", model_filename="model.pmml")
  • Access DataFrame of model deployed in Vantage.
    print(td_predictor.model_df)
  • Score model.
    output = td_predictor.predict(test_df, mode='udf')