XGBoost -> Deployer Pipeline

xgboost_deploy_pipeline.py

from kfp import dsl, gcp, compiler, components

deployer_component = components.load_component_from_url('https://raw.githubusercontent.com/kubeflow/pipelines/b850360cd6388cce127b8d079c2c6e7d54519ef2/components/gcp/ml_engine/deploy/component.yaml')

def xgboost_component(
  output_location,
  training_data,
  validation_data,
  data_type,
  target_column,
  objective,
  num_round=10):

  return dsl.ContainerOp(
    name='Train XGBoost model',
    image='gcr.io/aihub-c2t-containers/kfp-components/trainer/dist_xgboost:latest',
    arguments=[
      '--output-location', output_location,
      '--training-data', training_data,
      '--validation-data', validation_data,
      '--data-type', data_type,
      '--target-column', target_column,
      '--objective', objective,
      '--target-column', target_column,
      '--num-round', num_round,
    ],
    file_outputs={'model_uri': '/tmp/model_uri.txt'})

@dsl.pipeline(name='XGBoost training', description='XGBoost training pipeline')
def pipeline(
  model_name,
  model_version,
  project_id,
  output_location,
  training_data='gs://aihub-c2t-containers-public/release-0.2.0/kfp-components/trainer/dist_xgboost/data/csv/iris/train.csv',
  validation_data='gs://aihub-c2t-containers-public/release-0.2.0/kfp-components/trainer/dist_xgboost/data/csv/iris/valid.csv',
  data_type='csv',
  target_column='target',
  objective='multi:softprob',
  num_round=10):
  
  xgboost = xgboost_component(
    output_location=output_location,
    training_data=training_data,
    validation_data=validation_data,
    data_type=data_type,
    target_column=target_column,
    objective=objective,
    num_round=num_round,
  ).apply(gcp.use_gcp_secret('user-gcp-sa'))
  
  deployer = deployer_component(
    model_uri=xgboost.outputs['model_uri'],
    project_id=project_id,
    model_id=model_name,
    version_id=model_version,
    runtime_version='1.4',
    replace_existing_version='True',
    set_default='True',
    wait_interval='30'
  ).apply(gcp.use_gcp_secret('user-gcp-sa'))

compiler.Compiler().compile(pipeline, 'pipeline.tar.gz')