model_garden/notebooks/official/ml_metadata/vertex-pipelines-ml-metadata.ipynb

In [ ]:

# Copyright 2021 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
#
#     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.

Vertex AI: Track artifacts and metrics across Vertex AI Pipelines runs using Vertex ML Metadata

Run in Colab

View on GitHub

Open in Vertex AI Workbench

Overview

This notebook demonstrates how to track metrics and artifacts across Vertex AI Pipelines runs, and analyze this metadata using the Vertex AI SDK. If you'd prefer to follow a step-by-step tutorial, check out the codelab version of this notebook.

Learn more about Vertex ML Metadata and Vertex AI Pipelines.

Objective

In this notebook, you learn how to track artifacts and metrics with Vertex ML Metadata in Vertex AI Pipeline runs.

This tutorial uses the following Google Cloud ML services and resources:

• Vertex AI Pipelines
• Vertex ML Metadata

The steps performed include:

• Use the Kubeflow Pipelines SDK to build an ML pipeline that runs on Vertex AI
• The pipeline will create a dataset, train a scikit-learn model, and deploy the model to an endpoint
• Write custom pipeline components that generate artifacts and metadata
• Compare Vertex Pipelines runs, both in the Cloud console and programmatically
• Trace the lineage for pipeline-generated artifacts
• Query your pipeline run metadata

Dataset

In this notebook, we will train a model using scikit-learn to classify bean types using the Dry Beans Dataset from UCI Machine Learning. This is a tabular dataset that includes measurements and characteristics of seven different types of beans taken from images.

Costs

This tutorial uses billable components of Google Cloud:

• Vertex AI
• Cloud Storage

Learn about Vertex AI pricing and Cloud Storage pricing, and use the Pricing Calculator to generate a cost estimate based on your projected usage.

Install additional packages

Run the following commands to install the Vertex AI SDK and packages used in this notebook.

In [ ]:

! pip3 install --upgrade --quiet google-cloud-aiplatform==1.7.0 \
                                 kfp==1.8.9

Colab only: Uncomment the following cell to restart the kernel

In [ ]:

# Automatically restart kernel after installs so that your environment can access the new packages
# import IPython

# app = IPython.Application.instance()
# app.kernel.do_shutdown(True)

Before you begin

Set your project ID

If you don't know your project ID, try the following:

• Run gcloud config list.
• Run gcloud projects list.
• See the support page: Locate the project ID

In [ ]:

PROJECT_ID = "[your-project-id]"  # @param {type:"string"}

# Set the project id
! gcloud config set project {PROJECT_ID}

Region

You can also change the REGION variable used by Vertex AI. Learn more about Vertex AI regions.

In [ ]:

REGION = "us-central1"  # @param {type: "string"}

Authenticate your Google Cloud account

Depending on your Jupyter environment, you may have to manually authenticate. Follow the relevant instructions below.

1. Vertex AI Workbench

• Do nothing as you are already authenticated.

2. Local JupyterLab instance, uncomment and run:

In [ ]:

# ! gcloud auth login

3. Colab, uncomment and run:

In [ ]:

# from google.colab import auth
# auth.authenticate_user()

4. Service account or other

• See how to grant Cloud Storage permissions to your service account at https://cloud.google.com/storage/docs/gsutil/commands/iam#ch-examples.

Enable Cloud services used throughout this notebook.

Run the cell below to the enable Compute Engine, Container Registry, and Vertex AI services.

In [ ]:

!gcloud services enable compute.googleapis.com         \
                       containerregistry.googleapis.com  \
                       aiplatform.googleapis.com

Create a Cloud Storage bucket

Create a storage bucket to store intermediate artifacts such as datasets.

In [ ]:

BUCKET_URI = f"gs://your-bucket-name-{PROJECT_ID}-unique"  # @param {type:"string"}

Only if your bucket doesn't already exist: Run the following cell to create your Cloud Storage bucket.

In [ ]:

! gsutil mb -l $REGION $BUCKET_URI

Service Account

You use a service account to create Vertex AI Pipeline jobs.

If you do not want to use your project's Compute Engine service account, set SERVICE_ACCOUNT to another service account ID.

In [ ]:

SERVICE_ACCOUNT = "[your-service-account]"  # @param {type:"string"}

In [ ]:

if (
    SERVICE_ACCOUNT == ""
    or SERVICE_ACCOUNT is None
    or SERVICE_ACCOUNT == "[your-service-account]"
):
    # Get your service account from gcloud
    if not IS_COLAB:
        shell_output = !gcloud auth list 2>/dev/null
        SERVICE_ACCOUNT = shell_output[2].replace("*", "").strip()

    else:  # IS_COLAB:
        shell_output = ! gcloud projects describe  $PROJECT_ID
        project_number = shell_output[-1].split(":")[1].strip().replace("'", "")
        SERVICE_ACCOUNT = f"{project_number}-compute@developer.gserviceaccount.com"

    print("Service Account:", SERVICE_ACCOUNT)

Set service account access for Vertex AI Pipelines

Run the following commands to grant your service account access to read and write pipeline artifacts in the bucket that you created in the previous step -- you only need to run these once per service account.

In [ ]:

! gsutil iam ch serviceAccount:{SERVICE_ACCOUNT}:roles/storage.objectCreator $BUCKET_URI

! gsutil iam ch serviceAccount:{SERVICE_ACCOUNT}:roles/storage.objectViewer $BUCKET_URI

Import libraries and define constants

Import required libraries.

In [ ]:

import matplotlib.pyplot as plt
import pandas as pd
# We'll use this beta library for metadata querying
from google.cloud import aiplatform, aiplatform_v1beta1
from google.cloud.aiplatform import pipeline_jobs
from kfp.v2 import compiler, dsl
from kfp.v2.dsl import (Artifact, Dataset, Input, Metrics, Model, Output,
                        OutputPath, component)

Define some constants

In [ ]:

PATH = get_ipython().run_line_magic("env", "PATH")
%env PATH={PATH}:/home/jupyter/.local/bin
REGION = "us-central1"

PIPELINE_ROOT = f"{BUCKET_URI}/pipeline_root/"
PIPELINE_ROOT

Initialize the Vertex AI SDK

In [ ]:

aiplatform.init(project=PROJECT_ID, location=REGION)

Concepts

To better understand Vertex Pipelines and ML Metadata, we'd like to introduce the following concepts:

Pipeline Run

When we use the term run, we're referring to a single execution of your pipeline in Vertex Pipelines. Each run generates artifacts, metrics, and associated metadata.

Artifact

An artifact is a resource generated by your pipeline. Artifacts could include datasets, models, endpoints, or custom resources defined in your pipeline.

Metric

A metric is a way to measure the performance of your pipeline runs and artifacts. For example, a metric could be the accuracy of a classification model artifact created in your pipeline, or the size of the dataset used to train your model.

Metadata

Metadata describes the artifacts and metrics generated by your pipeline runs. Metadata on a model, for example, could include the URL of the model artifacts, its name, and the time it was created.

Creating a 3-step pipeline with custom components

The focus of this lab is on understanding metadata from pipeline runs. In order to do that, we'll need a pipeline to run on Vertex Pipelines, which is where we'll start. Here we'll define a 3-step pipeline with the following custom components:

• get_dataframe: Retrieve data from a BigQuery table and convert it into a pandas DataFrame
• train_sklearn_model: Use the pandas DataFrame to train and export a scikit-learn model, along with some metrics
• deploy_model: Deploy the exported scikit-learn model to an endpoint in Vertex AI

Create and define Python function based components

First, define the get_dataframe component with the code below. This component does the following:

• Creates a reference to a BigQuery table using the BigQuery client library
• Downloads the BigQuery table and converts it to a shuffled pandas DataFrame
• Exports the DataFrame to a CSV file

In [ ]:

@component(
    packages_to_install=["google-cloud-bigquery", "pandas", "pyarrow"],
    base_image="python:3.9",
    output_component_file="create_dataset.yaml",
)
def get_dataframe(bq_table: str, output_data_path: OutputPath("Dataset")):
    from google.cloud import bigquery

    bqclient = bigquery.Client(project=PROJECT_ID)
    table = bigquery.TableReference.from_string(bq_table)
    rows = bqclient.list_rows(table)
    dataframe = rows.to_dataframe(
        create_bqstorage_client=True,
    )
    dataframe = dataframe.sample(frac=1, random_state=2)
    dataframe.to_csv(output_data_path)

Next, create a component to train a scikit-learn model. This component does the following:

• Imports a CSV as a pandas DataFrame
• Splits the DataFrame into train and test sets
• Trains a scikit-learn model
• Logs metrics from the model
• Saves the model artifacts as a local model.joblib file

In [ ]:

@component(
    packages_to_install=["scikit-learn", "pandas", "joblib"],
    base_image="python:3.9",
    output_component_file="beans_model_component.yaml",
)
def sklearn_train(
    dataset: Input[Dataset], metrics: Output[Metrics], model: Output[Model]
):
    import pandas as pd
    from joblib import dump
    from sklearn.model_selection import train_test_split
    from sklearn.tree import DecisionTreeClassifier

    df = pd.read_csv(dataset.path)
    labels = df.pop("Class").tolist()
    data = df.values.tolist()
    x_train, x_test, y_train, y_test = train_test_split(data, labels)

    skmodel = DecisionTreeClassifier()
    skmodel.fit(x_train, y_train)
    score = skmodel.score(x_test, y_test)
    print("accuracy is:", score)

    metrics.log_metric("accuracy", (score * 100.0))
    metrics.log_metric("framework", "Scikit Learn")
    metrics.log_metric("dataset_size", len(df))
    dump(skmodel, model.path + ".joblib")

Finally, our last component will take the trained model from the previous step, upload it to Vertex AI, and deploy it to an endpoint:

In [ ]:

@component(
    packages_to_install=["google-cloud-aiplatform"],
    base_image="python:3.9",
    output_component_file="beans_deploy_component.yaml",
)
def deploy_model(
    model: Input[Model],
    project: str,
    region: str,
    vertex_endpoint: Output[Artifact],
    vertex_model: Output[Model],
):
    from google.cloud import aiplatform

    aiplatform.init(project=project, location=region)

    deployed_model = aiplatform.Model.upload(
        display_name="beans-model-pipeline",
        artifact_uri=model.uri.replace("model", ""),
        serving_container_image_uri="us-docker.pkg.dev/vertex-ai/prediction/sklearn-cpu.0-24:latest",
    )
    endpoint = deployed_model.deploy(machine_type="n1-standard-4")

    # Save data to the output params
    vertex_endpoint.uri = endpoint.resource_name
    vertex_model.uri = deployed_model.resource_name

Define and compile the pipeline

In [ ]:

@dsl.pipeline(
    # Default pipeline root. You can override it when submitting the pipeline.
    pipeline_root=PIPELINE_ROOT,
    # A name for the pipeline.
    name="mlmd-pipeline",
)
def pipeline(
    bq_table: str = "",
    output_data_path: str = "data.csv",
    project: str = PROJECT_ID,
    region: str = REGION,
):
    dataset_task = get_dataframe(bq_table)

    model_task = sklearn_train(dataset_task.output)

    deploy_model(model=model_task.outputs["model"], project=project, region=region)

The following will generate a JSON file that you'll use to run the pipeline:

In [ ]:

compiler.Compiler().compile(pipeline_func=pipeline, package_path="mlmd_pipeline.json")

Start two pipeline runs

Next we'll kick off two runs of our pipeline. First let's define a timestamp to use for our pipeline job IDs:

In [ ]:

from datetime import datetime

TIMESTAMP = datetime.now().strftime("%Y%m%d%H%M%S")

Our pipeline takes one parameter when we run it: the bq_table we want to use for training data. This pipeline run will use a smaller version of the beans dataset:

In [ ]:

run1 = pipeline_jobs.PipelineJob(
    display_name="mlmd-pipeline",
    template_path="mlmd_pipeline.json",
    job_id="mlmd-pipeline-small-{}".format(TIMESTAMP),
    parameter_values={"bq_table": "sara-vertex-demos.beans_demo.small_dataset"},
    enable_caching=True,
)

Next, create another pipeline run using a larger version of the same dataset.

In [ ]:

run2 = pipeline_jobs.PipelineJob(
    display_name="mlmd-pipeline",
    template_path="mlmd_pipeline.json",
    job_id="mlmd-pipeline-large-{}".format(TIMESTAMP),
    parameter_values={"bq_table": "sara-vertex-demos.beans_demo.large_dataset"},
    enable_caching=True,
)

Finally, kick off pipeline executions for both runs. It's best to do this in two separate notebook cells so you can see the output for each run.

In [ ]:

run1.submit()

Then, kick off the second run:

In [ ]:

run2.submit()

After running this cell, you'll see a link to view each pipeline in the Vertex AI console. Open that link to see more details on your pipeline.

These pipeline runs will take 10-15 minutes to complete.

Comparing pipeline runs

Now that you have two pipeline completed pipeline runs, we're ready to take a closer look at pipeline metrics using the Vertex AI SDK.

For guidance on inspecting pipeline artifacts and metadata in the Vertex AI Console, see this codelab.

You can use the aiplatform.get_pipeline_df() method to access run metadata. Here, we'll get metadata for the last two runs of the same pipeline and load it into a Pandas DataFrame. The mlmd-pipeline parameter here refers to the name we gave our pipeline in our pipeline definition:

In [ ]:

df = aiplatform.get_pipeline_df(pipeline="mlmd-pipeline")
print(df)

We've only executed our pipeline twice here, but you can imagine how many metrics you'd have with more executions. Next, we'll create a custom visualization with matplotlib to see the relationship between our model's accuracy and the amount of data used for training. Run the following to generate a graph:

In [ ]:

plt.plot(df["metric.dataset_size"], df["metric.accuracy"], label="Accuracy")
plt.title("Accuracy and dataset size")
plt.legend(loc=4)
plt.show()

Querying pipeline metrics

In addition to getting a DataFrame of all pipeline metrics, you may want to programmatically query artifacts created in your ML system. From there you could create a custom dashboard or let others in your organizaiton get details on specific artifacts.

Getting all Model artifacts

To query artifacts in this way, we'll create a MetadataServiceClient:

In [ ]:

API_ENDPOINT = "{}-aiplatform.googleapis.com".format(REGION)
metadata_client = aiplatform_v1beta1.MetadataServiceClient(
    client_options={"api_endpoint": API_ENDPOINT}
)

Next, we'll make a list_artifacts request to that endpoint and pass a filter indicating which artifacts we'd like in our response. First, let's get all the artifacts in our project that are models. To do that, run the following in your notebook:

In [ ]:

MODEL_FILTER = 'schema_title = "system.Model"'
artifact_request = aiplatform_v1beta1.ListArtifactsRequest(
    parent="projects/{}/locations/{}/metadataStores/default".format(PROJECT_ID, REGION),
    filter=MODEL_FILTER,
)
model_artifacts = metadata_client.list_artifacts(artifact_request)

The resulting model_artifacts response contains an iterable object for each model artifact in your project, along with associated metadata for each model.

Filtering objects and displaying in a DataFrame

It would be handy if we could more easily visualize the resulting artifact query. Next, let's get all artifacts created after August 10, 2021 with a LIVE state. After we run this request, we'll display the results in a pandas DataFrame. First, execute the request:

In [ ]:

LIVE_FILTER = 'create_time > "2021-08-10T00:00:00-00:00" AND state = LIVE'
artifact_req = {
    "parent": "projects/{}/locations/{}/metadataStores/default".format(
        PROJECT_ID, REGION
    ),
    "filter": LIVE_FILTER,
}
live_artifacts = metadata_client.list_artifacts(artifact_req)

Then, display the results in a DataFrame:

In [ ]:

data = {"uri": [], "createTime": [], "type": []}

for i in live_artifacts:
    data["uri"].append(i.uri)
    data["createTime"].append(i.create_time)
    data["type"].append(i.schema_title)

df = pd.DataFrame.from_dict(data)
print(df)

Cleaning up

To clean up all Google Cloud resources used in this project, you can delete the Google Cloud project you used for the tutorial.

If you don't want to delete the project, do the following to clean up the resources you used:

• If you used Google Cloud Notebooks to run this, stop or delete the notebook instance

• The pipeline runs we executed deployed endpoints in Vertex AI. Navigate to the Vertex AI console to delete those endpoints

• Delete the Cloud Storage bucket you created