Continuous training pipeline with KFP and Cloud AI Platform

Learning Objectives:

Learn how to use KF pre-build components (BiqQuery, CAIP training and predictions)
Learn how to use KF lightweight python components
Learn how to build a KF pipeline with these components
Learn how to compile, upload, and run a KF pipeline with the command line

In this lab, you will build, deploy, and run a KFP pipeline that orchestrates BigQuery and Cloud AI Platform services to train, tune, and deploy a scikit-learn model.

Understanding the pipeline design

The workflow implemented by the pipeline is defined using a Python based Domain Specific Language (DSL). The pipeline's DSL is in the covertype_training_pipeline.py file that we will generate below.

The pipeline's DSL has been designed to avoid hardcoding any environment specific settings like file paths or connection strings. These settings are provided to the pipeline code through a set of environment variables.



In [ ]:

    
!grep 'BASE_IMAGE =' -A 5 pipeline/covertype_training_pipeline.py

The pipeline uses a mix of custom and pre-build components.

Pre-build components. The pipeline uses the following pre-build components that are included with the KFP distribution:
Custom components. The pipeline uses two custom helper components that encapsulate functionality not available in any of the pre-build components. The components are implemented using the KFP SDK's Lightweight Python Components mechanism. The code for the components is in the helper_components.py file:
- Retrieve Best Run. This component retrieves a tuning metric and hyperparameter values for the best run of a AI Platform Training hyperparameter tuning job.
- Evaluate Model. This component evaluates a sklearn trained model using a provided metric and a testing dataset.



In [ ]:

    
%%writefile ./pipeline/covertype_training_pipeline.py
# Copyright 2019 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""KFP pipeline orchestrating BigQuery and Cloud AI Platform services."""

import os

from helper_components import evaluate_model
from helper_components import retrieve_best_run
from jinja2 import Template
import kfp
from kfp.components import func_to_container_op
from kfp.dsl.types import Dict
from kfp.dsl.types import GCPProjectID
from kfp.dsl.types import GCPRegion
from kfp.dsl.types import GCSPath
from kfp.dsl.types import String
from kfp.gcp import use_gcp_secret

# Defaults and environment settings
BASE_IMAGE = os.getenv('BASE_IMAGE')
TRAINER_IMAGE = os.getenv('TRAINER_IMAGE')
RUNTIME_VERSION = os.getenv('RUNTIME_VERSION')
PYTHON_VERSION = os.getenv('PYTHON_VERSION')
COMPONENT_URL_SEARCH_PREFIX = os.getenv('COMPONENT_URL_SEARCH_PREFIX')
USE_KFP_SA = os.getenv('USE_KFP_SA')

TRAINING_FILE_PATH = 'datasets/training/data.csv'
VALIDATION_FILE_PATH = 'datasets/validation/data.csv'
TESTING_FILE_PATH = 'datasets/testing/data.csv'

# Parameter defaults
SPLITS_DATASET_ID = 'splits'
HYPERTUNE_SETTINGS = """
{
    "hyperparameters":  {
        "goal": "MAXIMIZE",
        "maxTrials": 6,
        "maxParallelTrials": 3,
        "hyperparameterMetricTag": "accuracy",
        "enableTrialEarlyStopping": True,
        "params": [
            {
                "parameterName": "max_iter",
                "type": "DISCRETE",
                "discreteValues": [500, 1000]
            },
            {
                "parameterName": "alpha",
                "type": "DOUBLE",
                "minValue": 0.0001,
                "maxValue": 0.001,
                "scaleType": "UNIT_LINEAR_SCALE"
            }
        ]
    }
}
"""


# Helper functions
def generate_sampling_query(source_table_name, num_lots, lots):
    """Prepares the data sampling query."""

    sampling_query_template = """
         SELECT *
         FROM 
             `{{ source_table }}` AS cover
         WHERE 
         MOD(ABS(FARM_FINGERPRINT(TO_JSON_STRING(cover))), {{ num_lots }}) IN ({{ lots }})
         """
    query = Template(sampling_query_template).render(
        source_table=source_table_name, num_lots=num_lots, lots=str(lots)[1:-1])

    return query


# Create component factories
component_store = kfp.components.ComponentStore(
    local_search_paths=None, url_search_prefixes=[COMPONENT_URL_SEARCH_PREFIX])

bigquery_query_op = component_store.load_component('bigquery/query')
mlengine_train_op = component_store.load_component('ml_engine/train')
mlengine_deploy_op = component_store.load_component('ml_engine/deploy')
retrieve_best_run_op = func_to_container_op(
    retrieve_best_run, base_image=BASE_IMAGE)
evaluate_model_op = func_to_container_op(evaluate_model, base_image=BASE_IMAGE)


@kfp.dsl.pipeline(
    name='Covertype Classifier Training',
    description='The pipeline training and deploying the Covertype classifierpipeline_yaml'
)
def covertype_train(project_id,
                    region,
                    source_table_name,
                    gcs_root,
                    dataset_id,
                    evaluation_metric_name,
                    evaluation_metric_threshold,
                    model_id,
                    version_id,
                    replace_existing_version,
                    hypertune_settings=HYPERTUNE_SETTINGS,
                    dataset_location='US'):
    """Orchestrates training and deployment of an sklearn model."""

    # Create the training split
    query = generate_sampling_query(
        source_table_name=source_table_name, num_lots=10, lots=[1, 2, 3, 4])

    training_file_path = '{}/{}'.format(gcs_root, TRAINING_FILE_PATH)

    create_training_split = bigquery_query_op(
        query=query,
        project_id=project_id,
        dataset_id=dataset_id,
        table_id='',
        output_gcs_path=training_file_path,
        dataset_location=dataset_location)

    # Create the validation split
    query = generate_sampling_query(
        source_table_name=source_table_name, num_lots=10, lots=[8])

    validation_file_path = '{}/{}'.format(gcs_root, VALIDATION_FILE_PATH)

    create_validation_split = bigquery_query_op(
        query=query,
        project_id=project_id,
        dataset_id=dataset_id,
        table_id='',
        output_gcs_path=validation_file_path,
        dataset_location=dataset_location)

    # Create the testing split
    query = generate_sampling_query(
        source_table_name=source_table_name, num_lots=10, lots=[9])

    testing_file_path = '{}/{}'.format(gcs_root, TESTING_FILE_PATH)

    create_testing_split = bigquery_query_op(
        query=query,
        project_id=project_id,
        dataset_id=dataset_id,
        table_id='',
        output_gcs_path=testing_file_path,
        dataset_location=dataset_location)

    # Tune hyperparameters
    tune_args = [
        '--training_dataset_path',
        create_training_split.outputs['output_gcs_path'],
        '--validation_dataset_path',
        create_validation_split.outputs['output_gcs_path'], '--hptune', 'True'
    ]

    job_dir = '{}/{}/{}'.format(gcs_root, 'jobdir/hypertune',
                                kfp.dsl.RUN_ID_PLACEHOLDER)

    hypertune = mlengine_train_op(
        project_id=project_id,
        region=region,
        master_image_uri=TRAINER_IMAGE,
        job_dir=job_dir,
        args=tune_args,
        training_input=hypertune_settings)

    # Retrieve the best trial
    get_best_trial = retrieve_best_run_op(
            project_id, hypertune.outputs['job_id'])

    # Train the model on a combined training and validation datasets
    job_dir = '{}/{}/{}'.format(gcs_root, 'jobdir', kfp.dsl.RUN_ID_PLACEHOLDER)

    train_args = [
        '--training_dataset_path',
        create_training_split.outputs['output_gcs_path'],
        '--validation_dataset_path',
        create_validation_split.outputs['output_gcs_path'], '--alpha',
        get_best_trial.outputs['alpha'], '--max_iter',
        get_best_trial.outputs['max_iter'], '--hptune', 'False'
    ]

    train_model = mlengine_train_op(
        project_id=project_id,
        region=region,
        master_image_uri=TRAINER_IMAGE,
        job_dir=job_dir,
        args=train_args)

    # Evaluate the model on the testing split
    eval_model = evaluate_model_op(
        dataset_path=str(create_testing_split.outputs['output_gcs_path']),
        model_path=str(train_model.outputs['job_dir']),
        metric_name=evaluation_metric_name)

    # Deploy the model if the primary metric is better than threshold
    with kfp.dsl.Condition(eval_model.outputs['metric_value'] > evaluation_metric_threshold):
        deploy_model = mlengine_deploy_op(
        model_uri=train_model.outputs['job_dir'],
        project_id=project_id,
        model_id=model_id,
        version_id=version_id,
        runtime_version=RUNTIME_VERSION,
        python_version=PYTHON_VERSION,
        replace_existing_version=replace_existing_version)

    # Configure the pipeline to run using the service account defined
    # in the user-gcp-sa k8s secret
    if USE_KFP_SA == 'True':
        kfp.dsl.get_pipeline_conf().add_op_transformer(
              use_gcp_secret('user-gcp-sa'))

The custom components execute in a container image defined in base_image/Dockerfile.



In [ ]:

    
!cat base_image/Dockerfile

The training step in the pipeline employes the AI Platform Training component to schedule a AI Platform Training job in a custom training container. The custom training image is defined in trainer_image/Dockerfile.



In [ ]:

    
!cat trainer_image/Dockerfile

Building and deploying the pipeline

Before deploying to AI Platform Pipelines, the pipeline DSL has to be compiled into a pipeline runtime format, also refered to as a pipeline package. The runtime format is based on Argo Workflow, which is expressed in YAML.

Configure environment settings

Update the below constants with the settings reflecting your lab environment.

REGION - the compute region for AI Platform Training and Prediction
ARTIFACT_STORE - the GCS bucket created during installation of AI Platform Pipelines. The bucket name starts with the hostedkfp-default- prefix.
ENDPOINT - set the ENDPOINT constant to the endpoint to your AI Platform Pipelines instance. Then endpoint to the AI Platform Pipelines instance can be found on the AI Platform Pipelines page in the Google Cloud Console.

Open the SETTINGS for your instance
Use the value of the host variable in the Connect to this Kubeflow Pipelines instance from a Python client via Kubeflow Pipelines SKD section of the SETTINGS window.



In [ ]:

    
REGION = 'us-central1'
ENDPOINT = '337dd39580cbcbd2-dot-us-central2.pipelines.googleusercontent.com'
ARTIFACT_STORE_URI = 'gs://hostedkfp-default-e8c59nl4zo'
PROJECT_ID = !(gcloud config get-value core/project)
PROJECT_ID = PROJECT_ID[0]

Build the trainer image



In [ ]:

    
IMAGE_NAME='trainer_image'
TAG='latest'
TRAINER_IMAGE='gcr.io/{}/{}:{}'.format(PROJECT_ID, IMAGE_NAME, TAG)



In [ ]:

    
!gcloud builds submit --timeout 15m --tag $TRAINER_IMAGE trainer_image

Build the base image for custom components



In [ ]:

    
IMAGE_NAME='base_image'
TAG='latest'
BASE_IMAGE='gcr.io/{}/{}:{}'.format(PROJECT_ID, IMAGE_NAME, TAG)



In [ ]:

    
!gcloud builds submit --timeout 15m --tag $BASE_IMAGE base_image

Compile the pipeline

You can compile the DSL using an API from the KFP SDK or using the KFP compiler.

To compile the pipeline DSL using the KFP compiler.

Set the pipeline's compile time settings

The pipeline can run using a security context of the GKE default node pool's service account or the service account defined in the user-gcp-sa secret of the Kubernetes namespace hosting Kubeflow Pipelines. If you want to use the user-gcp-sa service account you change the value of USE_KFP_SA to True.

Note that the default AI Platform Pipelines configuration does not define the user-gcp-sa secret.



In [ ]:

    
USE_KFP_SA = False

COMPONENT_URL_SEARCH_PREFIX = 'https://raw.githubusercontent.com/kubeflow/pipelines/0.2.5/components/gcp/'
RUNTIME_VERSION = '1.15'
PYTHON_VERSION = '3.7'

%env USE_KFP_SA={USE_KFP_SA}
%env BASE_IMAGE={BASE_IMAGE}
%env TRAINER_IMAGE={TRAINER_IMAGE}
%env COMPONENT_URL_SEARCH_PREFIX={COMPONENT_URL_SEARCH_PREFIX}
%env RUNTIME_VERSION={RUNTIME_VERSION}
%env PYTHON_VERSION={PYTHON_VERSION}

Use the CLI compiler to compile the pipeline



In [ ]:

    
!dsl-compile --py pipeline/covertype_training_pipeline.py --output covertype_training_pipeline.yaml

The result is the covertype_training_pipeline.yaml file.



In [ ]:

    
!head covertype_training_pipeline.yaml

Deploy the pipeline package



In [ ]:

    
PIPELINE_NAME='covertype_continuous_training'

!kfp --endpoint $ENDPOINT pipeline upload \
-p $PIPELINE_NAME \
covertype_training_pipeline.yaml

Submitting pipeline runs

You can trigger pipeline runs using an API from the KFP SDK or using KFP CLI. To submit the run using KFP CLI, execute the following commands. Notice how the pipeline's parameters are passed to the pipeline run.

List the pipelines in AI Platform Pipelines



In [ ]:

    
!kfp --endpoint $ENDPOINT pipeline list

Submit a run

Find the ID of the covertype_continuous_training pipeline you uploaded in the previous step and update the value of PIPELINE_ID .



In [ ]:

    
PIPELINE_ID='0918568d-758c-46cf-9752-e04a4403cd84'



In [ ]:

    
EXPERIMENT_NAME = 'Covertype_Classifier_Training'
RUN_ID = 'Run_001'
SOURCE_TABLE = 'covertype_dataset.covertype'
DATASET_ID = 'splits'
EVALUATION_METRIC = 'accuracy'
EVALUATION_METRIC_THRESHOLD = '0.69'
MODEL_ID = 'covertype_classifier'
VERSION_ID = 'v01'
REPLACE_EXISTING_VERSION = 'True'

GCS_STAGING_PATH = '{}/staging'.format(ARTIFACT_STORE_URI)



In [ ]:

    
!kfp --endpoint $ENDPOINT run submit \
-e $EXPERIMENT_NAME \
-r $RUN_ID \
-p $PIPELINE_ID \
project_id=$PROJECT_ID \
gcs_root=$GCS_STAGING_PATH \
region=$REGION \
source_table_name=$SOURCE_TABLE \
dataset_id=$DATASET_ID \
evaluation_metric_name=$EVALUATION_METRIC \
evaluation_metric_threshold=$EVALUATION_METRIC_THRESHOLD \
model_id=$MODEL_ID \
version_id=$VERSION_ID \
replace_existing_version=$REPLACE_EXISTING_VERSION

where

EXPERIMENT_NAME is set to the experiment used to run the pipeline. You can choose any name you want. If the experiment does not exist it will be created by the command
RUN_ID is the name of the run. You can use an arbitrary name
PIPELINE_ID is the id of your pipeline. Use the value retrieved by the kfp pipeline list command
GCS_STAGING_PATH is the URI to the GCS location used by the pipeline to store intermediate files. By default, it is set to the staging folder in your artifact store.
REGION is a compute region for AI Platform Training and Prediction.

You should be already familiar with these and other parameters passed to the command. If not go back and review the pipeline code.

Monitoring the run

You can monitor the run using KFP UI. Follow the instructor who will walk you through the KFP UI and monitoring techniques.

To access the KFP UI in your environment use the following URI:

https://[ENDPOINT]

NOTE that your pipeline run may fail due to the bug in a BigQuery component that does not handle certain race conditions. If you observe the pipeline failure, retry the run from the KFP UI

Licensed under the Apache License, Version 2.0 (the \"License\"); you may not use this file except in compliance with the License. You may obtain a copy of the License at https://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an \"AS IS\" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.</font>