pipelines/components/contrib/google-cloud/ml_engine/train

Name

Component: Submitting an AI Platform training job as a pipeline step

Label

AI Platform, Kubeflow

Summary

A Kubeflow pipeline component to submit an AI Platform training job as a step in a pipeline.

Facets

Use case: Other

Technique: Other

Input data type: Tabular

ML workflow: Training

Details

Intended use

Use this component to submit a training job to AI Platform from a Kubeflow pipeline.

Runtime arguments

Argument	Description	Optional	Data type	Accepted values	Default
project_id	The Google Cloud Platform (GCP) project ID of the job.	No	GCPProjectID	-	-
python_module	The name of the Python module to run after installing the training program.	Yes	String	-	None
package_uris	The Cloud Storage location of the packages that contain the training program and any additional dependencies. The maximum number of package URIs is 100.	Yes	List	-	None
region	The Compute Engine region in which the training job is run.	Yes	GCPRegion	-	us-central1
args	The command line arguments to pass to the training program.	Yes	List	-	None
job_dir	A Cloud Storage path in which to store the training outputs and other data needed for training. This path is passed to your TensorFlow program as the command-line argument, job-dir. The benefit of specifying this field is that Cloud ML validates the path for use in training.	Yes	GCSPath	-	None
python_version	The version of Python used in training. If it is not set, the default version is 2.7. Python 3.5 is available when the runtime version is set to 1.4 and above.	Yes	String	-	None
runtime_version	The runtime version of AI Platform to use for training. If it is not set, AI Platform uses the default.	Yes	String	-	1
master_image_uri	The Docker image to run on the master replica. This image must be in Container Registry.	Yes	GCRPath	-	None
worker_image_uri	The Docker image to run on the worker replica. This image must be in Container Registry.	Yes	GCRPath	-	None
training_input	The input parameters to create a training job.	Yes	Dict	TrainingInput	None
job_id_prefix	The prefix of the job ID that is generated.	Yes	String	-	None
job_id	The ID of the job to create, takes precedence over generated job id if set.	Yes	String	-	None
wait_interval	The number of seconds to wait between API calls to get the status of the job.	Yes	Integer	-	30

Input data schema

The component accepts two types of inputs:

• A list of Python packages from Cloud Storage.
  • You can manually build a Python package and upload it to Cloud Storage by following this guide.
• A Docker container from Container Registry.
  • Follow this guide to publish and use a Docker container with this component.

Output

Name	Description	Type
job_id	The ID of the created job.	String
job_dir	The Cloud Storage path that contains the output files with the trained model.	GCSPath

Cautions & requirements

To use the component, you must:

• Set up a cloud environment by following this guide.
• The component can authenticate to GCP. Refer to Authenticating Pipelines to GCP for details.
• Grant the following access to the Kubeflow user service account:
  • Read access to the Cloud Storage buckets which contain the input data, packages, or Docker images.
  • Write access to the Cloud Storage bucket of the output directory.

Detailed description

The component builds the TrainingInput payload and submits a job via the AI Platform REST API.

The steps to use the component in a pipeline are:

1. Install the Kubeflow pipeline's SDK:

   %%capture --no-stderr
   
   !pip3 install kfp --upgrade
   

2. Load the component using the Kubeflow pipeline's SDK:

   import kfp.components as comp
   
   mlengine_train_op = comp.load_component_from_url('https://raw.githubusercontent.com/kubeflow/pipelines/1.7.0-rc.3/components/gcp/ml_engine/train/component.yaml')
   help(mlengine_train_op)
   

Sample

The following sample code works in an IPython notebook or directly in Python code.

In this sample, you use the code from the census estimator sample to train a model on AI Platform. To upload the code to AI Platform, package the Python code and upload it to a Cloud Storage bucket.

Note: You must have read and write permissions on the bucket that you use as the working directory.

Set sample parameters

# Required parameters
PROJECT_ID = '<Put your project ID here>'
GCS_WORKING_DIR = 'gs://<Put your GCS path here>' # No ending slash

# Optional parameters
EXPERIMENT_NAME = 'CLOUDML - Train'
TRAINER_GCS_PATH = GCS_WORKING_DIR + '/train/trainer.tar.gz'
OUTPUT_GCS_PATH = GCS_WORKING_DIR + '/train/output/'

Clean up the working directory

%%capture --no-stderr
!gsutil rm -r $GCS_WORKING_DIR

Download the sample trainer code to a local directory

%%capture --no-stderr
!wget https://github.com/GoogleCloudPlatform/cloudml-samples/archive/master.zip
!unzip master.zip

Package code and upload the package to Cloud Storage

%%capture --no-stderr
%%bash -s "$TRAINER_GCS_PATH"
pushd ./cloudml-samples-master/census/estimator/
python setup.py sdist
gsutil cp dist/preprocessing-1.0.tar.gz $1
popd
rm -fr ./cloudml-samples-master/ ./master.zip ./dist

Example pipeline that uses the component

import kfp.dsl as dsl
import json
@dsl.pipeline(
    name='CloudML training pipeline',
    description='CloudML training pipeline'
)
def pipeline(
    project_id = PROJECT_ID,
    python_module = 'trainer.task',
    package_uris = json.dumps([TRAINER_GCS_PATH]),
    region = 'us-central1',
    args = json.dumps([
        '--train-files', 'gs://cloud-samples-data/ml-engine/census/data/adult.data.csv',
        '--eval-files', 'gs://cloud-samples-data/ml-engine/census/data/adult.test.csv',
        '--train-steps', '1000',
        '--eval-steps', '100',
        '--verbosity', 'DEBUG'
    ]),
    job_dir = OUTPUT_GCS_PATH,
    python_version = '',
    runtime_version = '1.10',
    master_image_uri = '',
    worker_image_uri = '',
    training_input = '',
    job_id_prefix = '',
    job_id = '',
    wait_interval = '30'):
    task = mlengine_train_op(
        project_id=project_id, 
        python_module=python_module, 
        package_uris=package_uris, 
        region=region, 
        args=args, 
        job_dir=job_dir, 
        python_version=python_version,
        runtime_version=runtime_version, 
        master_image_uri=master_image_uri, 
        worker_image_uri=worker_image_uri, 
        training_input=training_input, 
        job_id_prefix=job_id_prefix, 
        job_id=job_id,
        wait_interval=wait_interval)

Compile the pipeline

pipeline_func = pipeline
pipeline_filename = pipeline_func.__name__ + '.zip'
import kfp.compiler as compiler
compiler.Compiler().compile(pipeline_func, pipeline_filename)

Submit the pipeline for execution

#Specify values for the pipeline's arguments
arguments = {}

#Get or create an experiment 
import kfp
client = kfp.Client()
experiment = client.create_experiment(EXPERIMENT_NAME)

#Submit a pipeline run
run_name = pipeline_func.__name__ + ' run'
run_result = client.run_pipeline(experiment.id, run_name, pipeline_filename, arguments)

Inspect the results

Use the following command to inspect the contents in the output directory:

!gsutil ls $OUTPUT_GCS_PATH

References

• Component Python code
• Component Docker file
• Sample notebook
• AI Platform REST API - Resource: Job

License

By deploying or using this software you agree to comply with the AI Hub Terms of Service and the Google APIs Terms of Service. To the extent of a direct conflict of terms, the AI Hub Terms of Service will control.