Help protect the Great Barrier Reef with TensorFlow on Kaggle Join Challenge

Create a TFX pipeline using templates

Introduction

This document will provide instructions to create a TensorFlow Extended (TFX) pipeline using templates which are provided with TFX Python package. Many of the instructions are Linux shell commands, which will run on an AI Platform Notebooks instance. Corresponding Jupyter Notebook code cells which invoke those commands using ! are provided.

You will build a pipeline using Taxi Trips dataset released by the City of Chicago. We strongly encourage you to try building your own pipeline using your dataset by utilizing this pipeline as a baseline.

Step 1. Set up your environment.

AI Platform Pipelines will prepare a development environment to build a pipeline, and a Kubeflow Pipeline cluster to run the newly built pipeline.

Install tfx python package with kfp extra requirement.

import sys
# Use the latest version of pip.
!pip install --upgrade pip
# Install tfx and kfp Python packages.
!pip install --upgrade "tfx[kfp]<2"

Let's check the versions of TFX.

python3 -c "from tfx import version ; print('TFX version: {}'.format(version.__version__))"
TFX version: 0.29.0

In AI Platform Pipelines, TFX is running in a hosted Kubernetes environment using Kubeflow Pipelines.

Let's set some environment variables to use Kubeflow Pipelines.

First, get your GCP project ID.

# Read GCP project id from env.
shell_output=!gcloud config list --format 'value(core.project)' 2>/dev/null
GOOGLE_CLOUD_PROJECT=shell_output[0]
%env GOOGLE_CLOUD_PROJECT={GOOGLE_CLOUD_PROJECT}
print("GCP project ID:" + GOOGLE_CLOUD_PROJECT)
env: GOOGLE_CLOUD_PROJECT=tf-benchmark-dashboard
GCP project ID:tf-benchmark-dashboard

We also need to access your KFP cluster. You can access it in your Google Cloud Console under "AI Platform > Pipeline" menu. The "endpoint" of the KFP cluster can be found from the URL of the Pipelines dashboard, or you can get it from the URL of the Getting Started page where you launched this notebook. Let's create an ENDPOINT environment variable and set it to the KFP cluster endpoint. ENDPOINT should contain only the hostname part of the URL. For example, if the URL of the KFP dashboard is <a href="https://1e9deb537390ca22-dot-asia-east1.pipelines.googleusercontent.com/#/start">https://1e9deb537390ca22-dot-asia-east1.pipelines.googleusercontent.com/#/start</a>, ENDPOINT value becomes 1e9deb537390ca22-dot-asia-east1.pipelines.googleusercontent.com.

# This refers to the KFP cluster endpoint
ENDPOINT='' # Enter your ENDPOINT here.
if not ENDPOINT:
    from absl import logging
    logging.error('Set your ENDPOINT in this cell.')
ERROR:absl:Set your ENDPOINT in this cell.

Set the image name as tfx-pipeline under the current GCP project.

# Docker image name for the pipeline image.
CUSTOM_TFX_IMAGE='gcr.io/' + GOOGLE_CLOUD_PROJECT + '/tfx-pipeline'

And, it's done. We are ready to create a pipeline.

Step 2. Copy the predefined template to your project directory.

In this step, we will create a working pipeline project directory and files by copying additional files from a predefined template.

You may give your pipeline a different name by changing the PIPELINE_NAME below. This will also become the name of the project directory where your files will be put.

PIPELINE_NAME="my_pipeline"
import os
PROJECT_DIR=os.path.join(os.path.expanduser("~"),"imported",PIPELINE_NAME)

TFX includes the taxi template with the TFX python package. If you are planning to solve a point-wise prediction problem, including classification and regresssion, this template could be used as a starting point.

The tfx template copy CLI command copies predefined template files into your project directory.

!tfx template copy \
  --pipeline-name={PIPELINE_NAME} \
  --destination-path={PROJECT_DIR} \
  --model=taxi
CLI
Copying taxi pipeline template
kubeflow_runner.py -> /home/kbuilder/imported/my_pipeline/kubeflow_runner.py
kubeflow_v2_dag_runner.py -> /home/kbuilder/imported/my_pipeline/kubeflow_v2_dag_runner.py
features_test.py -> /home/kbuilder/imported/my_pipeline/models/features_test.py
model_test.py -> /home/kbuilder/imported/my_pipeline/models/estimator/model_test.py
constants.py -> /home/kbuilder/imported/my_pipeline/models/estimator/constants.py
model.py -> /home/kbuilder/imported/my_pipeline/models/estimator/model.py
__init__.py -> /home/kbuilder/imported/my_pipeline/models/estimator/__init__.py
model_test.py -> /home/kbuilder/imported/my_pipeline/models/keras/model_test.py
constants.py -> /home/kbuilder/imported/my_pipeline/models/keras/constants.py
model.py -> /home/kbuilder/imported/my_pipeline/models/keras/model.py
__init__.py -> /home/kbuilder/imported/my_pipeline/models/keras/__init__.py
preprocessing_test.py -> /home/kbuilder/imported/my_pipeline/models/preprocessing_test.py
preprocessing.py -> /home/kbuilder/imported/my_pipeline/models/preprocessing.py
__init__.py -> /home/kbuilder/imported/my_pipeline/models/__init__.py
features.py -> /home/kbuilder/imported/my_pipeline/models/features.py
pipeline.py -> /home/kbuilder/imported/my_pipeline/pipeline/pipeline.py
configs.py -> /home/kbuilder/imported/my_pipeline/pipeline/configs.py
__init__.py -> /home/kbuilder/imported/my_pipeline/pipeline/__init__.py
local_runner.py -> /home/kbuilder/imported/my_pipeline/local_runner.py
model_analysis.ipynb -> /home/kbuilder/imported/my_pipeline/model_analysis.ipynb
__init__.py -> /home/kbuilder/imported/my_pipeline/__init__.py
data_validation.ipynb -> /home/kbuilder/imported/my_pipeline/data_validation.ipynb
.gitignore -> /home/kbuilder/imported/my_pipeline/.gitignore
Traceback (most recent call last):
  File "/tmpfs/src/tf_docs_env/bin/tfx", line 8, in <module>
    sys.exit(cli_group())
  File "/tmpfs/src/tf_docs_env/lib/python3.7/site-packages/click/core.py", line 829, in __call__
    return self.main(*args, **kwargs)
  File "/tmpfs/src/tf_docs_env/lib/python3.7/site-packages/click/core.py", line 782, in main
    rv = self.invoke(ctx)
  File "/tmpfs/src/tf_docs_env/lib/python3.7/site-packages/click/core.py", line 1259, in invoke
    return _process_result(sub_ctx.command.invoke(sub_ctx))
  File "/tmpfs/src/tf_docs_env/lib/python3.7/site-packages/click/core.py", line 1259, in invoke
    return _process_result(sub_ctx.command.invoke(sub_ctx))
  File "/tmpfs/src/tf_docs_env/lib/python3.7/site-packages/click/core.py", line 1066, in invoke
    return ctx.invoke(self.callback, **ctx.params)
  File "/tmpfs/src/tf_docs_env/lib/python3.7/site-packages/click/core.py", line 610, in invoke
    return callback(*args, **kwargs)
  File "/tmpfs/src/tf_docs_env/lib/python3.7/site-packages/click/decorators.py", line 73, in new_func
    return ctx.invoke(f, obj, *args, **kwargs)
  File "/tmpfs/src/tf_docs_env/lib/python3.7/site-packages/click/core.py", line 610, in invoke
    return callback(*args, **kwargs)
  File "/tmpfs/src/tf_docs_env/lib/python3.7/site-packages/tfx/tools/cli/commands/template.py", line 73, in copy
    template_handler.copy_template(ctx.flags_dict)
  File "/tmpfs/src/tf_docs_env/lib/python3.7/site-packages/tfx/tools/cli/handler/template_handler.py", line 185, in copy_template
    fileio.copy(src_path, dst_path)
  File "/tmpfs/src/tf_docs_env/lib/python3.7/site-packages/tfx/dsl/io/fileio.py", line 51, in copy
    src_fs.copy(src, dst, overwrite=overwrite)
  File "/tmpfs/src/tf_docs_env/lib/python3.7/site-packages/tfx/dsl/io/plugins/tensorflow_gfile.py", line 48, in copy
    tf.io.gfile.copy(src, dst, overwrite=overwrite)
  File "/tmpfs/src/tf_docs_env/lib/python3.7/site-packages/tensorflow/python/lib/io/file_io.py", line 516, in copy_v2
    compat.path_to_bytes(src), compat.path_to_bytes(dst), overwrite)
tensorflow.python.framework.errors_impl.AlreadyExistsError: file already exists

Change the working directory context in this notebook to the project directory.

%cd {PROJECT_DIR}
/home/kbuilder/imported/my_pipeline

Step 3. Browse your copied source files

The TFX template provides basic scaffold files to build a pipeline, including Python source code, sample data, and Jupyter Notebooks to analyse the output of the pipeline. The taxi template uses the same Chicago Taxi dataset and ML model as the Airflow Tutorial.

Here is brief introduction to each of the Python files.

  • pipeline - This directory contains the definition of the pipeline
    • configs.py — defines common constants for pipeline runners
    • pipeline.py — defines TFX components and a pipeline
  • models - This directory contains ML model definitions.
    • features.py, features_test.py — defines features for the model
    • preprocessing.py, preprocessing_test.py — defines preprocessing jobs using tf::Transform
    • estimator - This directory contains an Estimator based model.
      • constants.py — defines constants of the model
      • model.py, model_test.py — defines DNN model using TF estimator
    • keras - This directory contains a Keras based model.
      • constants.py — defines constants of the model
      • model.py, model_test.py — defines DNN model using Keras
  • local_runner.py, kubeflow_runner.py — define runners for each orchestration engine

You might notice that there are some files with _test.py in their name. These are unit tests of the pipeline and it is recommended to add more unit tests as you implement your own pipelines. You can run unit tests by supplying the module name of test files with -m flag. You can usually get a module name by deleting .py extension and replacing / with .. For example:

{sys.executable} -m models.features_test
{sys.executable} -m models.keras.model_test
Running tests under Python 3.7.5: /tmpfs/src/tf_docs_env/bin/python
[ RUN      ] FeaturesTest.testNumberOfBucketFeatureBucketCount
INFO:tensorflow:time(__main__.FeaturesTest.testNumberOfBucketFeatureBucketCount): 0.0s
I1204 11:33:54.064224 139808961349440 test_util.py:2076] time(__main__.FeaturesTest.testNumberOfBucketFeatureBucketCount): 0.0s
[       OK ] FeaturesTest.testNumberOfBucketFeatureBucketCount
[ RUN      ] FeaturesTest.testTransformedNames
INFO:tensorflow:time(__main__.FeaturesTest.testTransformedNames): 0.0s
I1204 11:33:54.064666 139808961349440 test_util.py:2076] time(__main__.FeaturesTest.testTransformedNames): 0.0s
[       OK ] FeaturesTest.testTransformedNames
[ RUN      ] FeaturesTest.test_session
[  SKIPPED ] FeaturesTest.test_session
----------------------------------------------------------------------
Ran 3 tests in 0.001s

OK (skipped=1)
Running tests under Python 3.7.5: /tmpfs/src/tf_docs_env/bin/python
[ RUN      ] ModelTest.testBuildKerasModel
2021-12-04 11:33:57.507456: W tensorflow/stream_executor/platform/default/dso_loader.cc:60] Could not load dynamic library 'libcusolver.so.10'; dlerror: libcusolver.so.10: cannot open shared object file: No such file or directory
2021-12-04 11:33:57.508566: W tensorflow/core/common_runtime/gpu/gpu_device.cc:1757] Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at https://www.tensorflow.org/install/gpu for how to download and setup the required libraries for your platform.
Skipping registering GPU devices...
I1204 11:33:57.581331 139740839778112 layer_utils.py:191] Model: "model"
I1204 11:33:57.581501 139740839778112 layer_utils.py:192] __________________________________________________________________________________________________
I1204 11:33:57.581558 139740839778112 layer_utils.py:189] Layer (type)                    Output Shape         Param #     Connected to                     
I1204 11:33:57.581596 139740839778112 layer_utils.py:194] ==================================================================================================
I1204 11:33:57.581741 139740839778112 layer_utils.py:189] pickup_latitude_xf (InputLayer) [(None,)]            0                                            
I1204 11:33:57.581793 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.581883 139740839778112 layer_utils.py:189] trip_miles_xf (InputLayer)      [(None,)]            0                                            
I1204 11:33:57.581926 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.582010 139740839778112 layer_utils.py:189] trip_start_hour_xf (InputLayer) [(None,)]            0                                            
I1204 11:33:57.582052 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.582189 139740839778112 layer_utils.py:189] dense_features (DenseFeatures)  (None, 1)            0           pickup_latitude_xf[0][0]         
I1204 11:33:57.582241 139740839778112 layer_utils.py:189]                                                                  trip_miles_xf[0][0]              
I1204 11:33:57.582280 139740839778112 layer_utils.py:189]                                                                  trip_start_hour_xf[0][0]         
I1204 11:33:57.582315 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.582462 139740839778112 layer_utils.py:189] dense (Dense)                   (None, 1)            2           dense_features[0][0]             
I1204 11:33:57.582518 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.582629 139740839778112 layer_utils.py:189] dense_1 (Dense)                 (None, 1)            2           dense[0][0]                      
I1204 11:33:57.582674 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.582824 139740839778112 layer_utils.py:189] dense_features_1 (DenseFeatures (None, 34)           0           pickup_latitude_xf[0][0]         
I1204 11:33:57.582879 139740839778112 layer_utils.py:189]                                                                  trip_miles_xf[0][0]              
I1204 11:33:57.582921 139740839778112 layer_utils.py:189]                                                                  trip_start_hour_xf[0][0]         
I1204 11:33:57.582957 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.583053 139740839778112 layer_utils.py:189] concatenate (Concatenate)       (None, 35)           0           dense_1[0][0]                    
I1204 11:33:57.583099 139740839778112 layer_utils.py:189]                                                                  dense_features_1[0][0]           
I1204 11:33:57.583143 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.583260 139740839778112 layer_utils.py:189] dense_2 (Dense)                 (None, 1)            36          concatenate[0][0]                
I1204 11:33:57.583309 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.583389 139740839778112 layer_utils.py:189] tf.compat.v1.squeeze (TFOpLambd (None,)              0           dense_2[0][0]                    
I1204 11:33:57.583432 139740839778112 layer_utils.py:256] ==================================================================================================
I1204 11:33:57.583687 139740839778112 layer_utils.py:267] Total params: 40
I1204 11:33:57.583751 139740839778112 layer_utils.py:268] Trainable params: 40
I1204 11:33:57.583794 139740839778112 layer_utils.py:269] Non-trainable params: 0
I1204 11:33:57.583832 139740839778112 layer_utils.py:270] __________________________________________________________________________________________________
I1204 11:33:57.649701 139740839778112 layer_utils.py:191] Model: "model_1"
I1204 11:33:57.649825 139740839778112 layer_utils.py:192] __________________________________________________________________________________________________
I1204 11:33:57.649878 139740839778112 layer_utils.py:189] Layer (type)                    Output Shape         Param #     Connected to                     
I1204 11:33:57.649932 139740839778112 layer_utils.py:194] ==================================================================================================
I1204 11:33:57.650066 139740839778112 layer_utils.py:189] pickup_latitude_xf (InputLayer) [(None,)]            0                                            
I1204 11:33:57.650120 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.650207 139740839778112 layer_utils.py:189] trip_miles_xf (InputLayer)      [(None,)]            0                                            
I1204 11:33:57.650259 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.650356 139740839778112 layer_utils.py:189] trip_start_hour_xf (InputLayer) [(None,)]            0                                            
I1204 11:33:57.650398 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.650552 139740839778112 layer_utils.py:189] dense_features_2 (DenseFeatures (None, 1)            0           pickup_latitude_xf[0][0]         
I1204 11:33:57.650603 139740839778112 layer_utils.py:189]                                                                  trip_miles_xf[0][0]              
I1204 11:33:57.650644 139740839778112 layer_utils.py:189]                                                                  trip_start_hour_xf[0][0]         
I1204 11:33:57.650682 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.650812 139740839778112 layer_utils.py:189] dense_3 (Dense)                 (None, 1)            2           dense_features_2[0][0]           
I1204 11:33:57.650864 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.651007 139740839778112 layer_utils.py:189] dense_features_3 (DenseFeatures (None, 34)           0           pickup_latitude_xf[0][0]         
I1204 11:33:57.651061 139740839778112 layer_utils.py:189]                                                                  trip_miles_xf[0][0]              
I1204 11:33:57.651102 139740839778112 layer_utils.py:189]                                                                  trip_start_hour_xf[0][0]         
I1204 11:33:57.651146 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.651229 139740839778112 layer_utils.py:189] concatenate_1 (Concatenate)     (None, 35)           0           dense_3[0][0]                    
I1204 11:33:57.651274 139740839778112 layer_utils.py:189]                                                                  dense_features_3[0][0]           
I1204 11:33:57.651311 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.651462 139740839778112 layer_utils.py:189] dense_4 (Dense)                 (None, 1)            36          concatenate_1[0][0]              
I1204 11:33:57.651547 139740839778112 layer_utils.py:258] __________________________________________________________________________________________________
I1204 11:33:57.651632 139740839778112 layer_utils.py:189] tf.compat.v1.squeeze_1 (TFOpLam (None,)              0           dense_4[0][0]                    
I1204 11:33:57.651675 139740839778112 layer_utils.py:256] ==================================================================================================
I1204 11:33:57.651959 139740839778112 layer_utils.py:267] Total params: 38
I1204 11:33:57.652019 139740839778112 layer_utils.py:268] Trainable params: 38
I1204 11:33:57.652061 139740839778112 layer_utils.py:269] Non-trainable params: 0
I1204 11:33:57.652098 139740839778112 layer_utils.py:270] __________________________________________________________________________________________________
INFO:tensorflow:time(__main__.ModelTest.testBuildKerasModel): 0.84s
I1204 11:33:57.652639 139740839778112 test_util.py:2076] time(__main__.ModelTest.testBuildKerasModel): 0.84s
[       OK ] ModelTest.testBuildKerasModel
[ RUN      ] ModelTest.test_session
[  SKIPPED ] ModelTest.test_session
----------------------------------------------------------------------
Ran 2 tests in 0.836s

OK (skipped=1)

Step 4. Run your first TFX pipeline

Components in the TFX pipeline will generate outputs for each run as ML Metadata Artifacts, and they need to be stored somewhere. You can use any storage which the KFP cluster can access, and for this example we will use Google Cloud Storage (GCS). A default GCS bucket should have been created automatically. Its name will be <your-project-id>-kubeflowpipelines-default.

Let's upload our sample data to GCS bucket so that we can use it in our pipeline later.

gsutil cp data/data.csv gs://{GOOGLE_CLOUD_PROJECT}-kubeflowpipelines-default/tfx-template/data/taxi/data.csv
BucketNotFoundException: 404 gs://tf-benchmark-dashboard-kubeflowpipelines-default bucket does not exist.

Let's create a TFX pipeline using the tfx pipeline create command.

!tfx pipeline create  --pipeline-path=kubeflow_runner.py --endpoint={ENDPOINT} \
--build-image
CLI
Usage: tfx pipeline create [OPTIONS]
Try 'tfx pipeline create --help' for help.

Error: no such option: --build-image

While creating a pipeline, Dockerfile will be generated to build a Docker image. Don't forget to add it to the source control system (for example, git) along with other source files.

Now start an execution run with the newly created pipeline using the tfx run create command.

tfx run create --pipeline-name={PIPELINE_NAME} --endpoint={ENDPOINT}
CLI
Creating a run for pipeline: my_pipeline
Detected Beam.
[WARNING] Default engine will be changed to "local" in the near future.
Use --engine flag if you intend to use a different orchestrator.
Pipeline "my_pipeline" does not exist.

Or, you can also run the pipeline in the KFP Dashboard. The new execution run will be listed under Experiments in the KFP Dashboard. Clicking into the experiment will allow you to monitor progress and visualize the artifacts created during the execution run.

However, we recommend visiting the KFP Dashboard. You can access the KFP Dashboard from the Cloud AI Platform Pipelines menu in Google Cloud Console. Once you visit the dashboard, you will be able to find the pipeline, and access a wealth of information about the pipeline. For example, you can find your runs under the Experiments menu, and when you open your execution run under Experiments you can find all your artifacts from the pipeline under Artifacts menu.

One of the major sources of failure is permission related problems. Please make sure your KFP cluster has permissions to access Google Cloud APIs. This can be configured when you create a KFP cluster in GCP, or see Troubleshooting document in GCP.

Step 5. Add components for data validation.

In this step, you will add components for data validation including StatisticsGen, SchemaGen, and ExampleValidator. If you are interested in data validation, please see Get started with Tensorflow Data Validation.

Double-click to change directory to pipeline and double-click again to open pipeline.py. Find and uncomment the 3 lines which add StatisticsGen, SchemaGen, and ExampleValidator to the pipeline. (Tip: search for comments containing TODO(step 5):). Make sure to save pipeline.py after you edit it.

You now need to update the existing pipeline with modified pipeline definition. Use the tfx pipeline update command to update your pipeline, followed by the tfx run create command to create a new execution run of your updated pipeline.

# Update the pipeline
!tfx pipeline update \
--pipeline-path=kubeflow_runner.py \
--endpoint={ENDPOINT}
# You can run the pipeline the same way.
!tfx run create --pipeline-name {PIPELINE_NAME} --endpoint={ENDPOINT}
CLI
Updating pipeline
Detected Beam.
[WARNING] Default engine will be changed to "local" in the near future.
Use --engine flag if you intend to use a different orchestrator.
beam runner not found in dsl.
CLI
Creating a run for pipeline: my_pipeline
Detected Beam.
[WARNING] Default engine will be changed to "local" in the near future.
Use --engine flag if you intend to use a different orchestrator.
Pipeline "my_pipeline" does not exist.

Check pipeline outputs

Visit the KFP dashboard to find pipeline outputs in the page for your pipeline run. Click the Experiments tab on the left, and All runs in the Experiments page. You should be able to find the latest run under the name of your pipeline.

Step 6. Add components for training.

In this step, you will add components for training and model validation including Transform, Trainer, Resolver, Evaluator, and Pusher.

Double-click to open pipeline.py. Find and uncomment the 5 lines which add Transform, Trainer, Resolver, Evaluator and Pusher to the pipeline. (Tip: search for TODO(step 6):)

As you did before, you now need to update the existing pipeline with the modified pipeline definition. The instructions are the same as Step 5. Update the pipeline using tfx pipeline update, and create an execution run using tfx run create.

!tfx pipeline update \
--pipeline-path=kubeflow_runner.py \
--endpoint={ENDPOINT}
!tfx run create --pipeline-name {PIPELINE_NAME} --endpoint={ENDPOINT}
CLI
Updating pipeline
Detected Beam.
[WARNING] Default engine will be changed to "local" in the near future.
Use --engine flag if you intend to use a different orchestrator.
beam runner not found in dsl.
CLI
Creating a run for pipeline: my_pipeline
Detected Beam.
[WARNING] Default engine will be changed to "local" in the near future.
Use --engine flag if you intend to use a different orchestrator.
Pipeline "my_pipeline" does not exist.

When this execution run finishes successfully, you have now created and run your first TFX pipeline in AI Platform Pipelines!

Step 7. (Optional) Try BigQueryExampleGen

BigQuery is a serverless, highly scalable, and cost-effective cloud data warehouse. BigQuery can be used as a source for training examples in TFX. In this step, we will add BigQueryExampleGen to the pipeline.

Double-click to open pipeline.py. Comment out CsvExampleGen and uncomment the line which creates an instance of BigQueryExampleGen. You also need to uncomment the query argument of the create_pipeline function.

We need to specify which GCP project to use for BigQuery, and this is done by setting --project in beam_pipeline_args when creating a pipeline.

Double-click to open configs.py. Uncomment the definition of GOOGLE_CLOUD_REGION, BIG_QUERY_WITH_DIRECT_RUNNER_BEAM_PIPELINE_ARGS and BIG_QUERY_QUERY. You should replace the region value in this file with the correct values for your GCP project.

Change directory one level up. Click the name of the directory above the file list. The name of the directory is the name of the pipeline which is my_pipeline if you didn't change.

Double-click to open kubeflow_runner.py. Uncomment two arguments, query and beam_pipeline_args, for the create_pipeline function.

Now the pipeline is ready to use BigQuery as an example source. Update the pipeline as before and create a new execution run as we did in step 5 and 6.

!tfx pipeline update \
--pipeline-path=kubeflow_runner.py \
--endpoint={ENDPOINT}
!tfx run create --pipeline-name {PIPELINE_NAME} --endpoint={ENDPOINT}
CLI
Updating pipeline
Detected Beam.
[WARNING] Default engine will be changed to "local" in the near future.
Use --engine flag if you intend to use a different orchestrator.
beam runner not found in dsl.
CLI
Creating a run for pipeline: my_pipeline
Detected Beam.
[WARNING] Default engine will be changed to "local" in the near future.
Use --engine flag if you intend to use a different orchestrator.
Pipeline "my_pipeline" does not exist.

Step 8. (Optional) Try Dataflow with KFP

Several TFX Components uses Apache Beam to implement data-parallel pipelines, and it means that you can distribute data processing workloads using Google Cloud Dataflow. In this step, we will set the Kubeflow orchestrator to use dataflow as the data processing back-end for Apache Beam.

Double-click pipeline to change directory, and double-click to open configs.py. Uncomment the definition of GOOGLE_CLOUD_REGION, and DATAFLOW_BEAM_PIPELINE_ARGS.

Change directory one level up. Click the name of the directory above the file list. The name of the directory is the name of the pipeline which is my_pipeline if you didn't change.

Double-click to open kubeflow_runner.py. Uncomment beam_pipeline_args. (Also make sure to comment out current beam_pipeline_args that you added in Step 7.)

Now the pipeline is ready to use Dataflow. Update the pipeline and create an execution run as we did in step 5 and 6.

!tfx pipeline update \
--pipeline-path=kubeflow_runner.py \
--endpoint={ENDPOINT}
!tfx run create --pipeline-name {PIPELINE_NAME} --endpoint={ENDPOINT}
CLI
Updating pipeline
Detected Beam.
[WARNING] Default engine will be changed to "local" in the near future.
Use --engine flag if you intend to use a different orchestrator.
beam runner not found in dsl.
CLI
Creating a run for pipeline: my_pipeline
Detected Beam.
[WARNING] Default engine will be changed to "local" in the near future.
Use --engine flag if you intend to use a different orchestrator.
Pipeline "my_pipeline" does not exist.

You can find your Dataflow jobs in Dataflow in Cloud Console.

Step 9. (Optional) Try Cloud AI Platform Training and Prediction with KFP

TFX interoperates with several managed GCP services, such as Cloud AI Platform for Training and Prediction. You can set your Trainer component to use Cloud AI Platform Training, a managed service for training ML models. Moreover, when your model is built and ready to be served, you can push your model to Cloud AI Platform Prediction for serving. In this step, we will set our Trainer and Pusher component to use Cloud AI Platform services.

Before editing files, you might first have to enable AI Platform Training & Prediction API.

Double-click pipeline to change directory, and double-click to open configs.py. Uncomment the definition of GOOGLE_CLOUD_REGION, GCP_AI_PLATFORM_TRAINING_ARGS and GCP_AI_PLATFORM_SERVING_ARGS. We will use our custom built container image to train a model in Cloud AI Platform Training, so we should set masterConfig.imageUri in GCP_AI_PLATFORM_TRAINING_ARGS to the same value as CUSTOM_TFX_IMAGE above.

Change directory one level up, and double-click to open kubeflow_runner.py. Uncomment ai_platform_training_args and ai_platform_serving_args.

Update the pipeline and create an execution run as we did in step 5 and 6.

!tfx pipeline update \
--pipeline-path=kubeflow_runner.py \
--endpoint={ENDPOINT}
!tfx run create --pipeline-name {PIPELINE_NAME} --endpoint={ENDPOINT}
CLI
Updating pipeline
Detected Beam.
[WARNING] Default engine will be changed to "local" in the near future.
Use --engine flag if you intend to use a different orchestrator.
beam runner not found in dsl.
CLI
Creating a run for pipeline: my_pipeline
Detected Beam.
[WARNING] Default engine will be changed to "local" in the near future.
Use --engine flag if you intend to use a different orchestrator.
Pipeline "my_pipeline" does not exist.

You can find your training jobs in Cloud AI Platform Jobs. If your pipeline completed successfully, you can find your model in Cloud AI Platform Models.

Step 10. Ingest YOUR data to the pipeline

We made a pipeline for a model using the Chicago Taxi dataset. Now it's time to put your data into the pipeline.

Your data can be stored anywhere your pipeline can access, including GCS, or BigQuery. You will need to modify the pipeline definition to access your data.

  1. If your data is stored in files, modify the DATA_PATH in kubeflow_runner.py or local_runner.py and set it to the location of your files. If your data is stored in BigQuery, modify BIG_QUERY_QUERY in pipeline/configs.py to correctly query for your data.
  2. Add features in models/features.py.
  3. Modify models/preprocessing.py to transform input data for training.
  4. Modify models/keras/model.py and models/keras/constants.py to describe your ML model.
    • You can use an estimator based model, too. Change RUN_FN constant to models.estimator.model.run_fn in pipeline/configs.py.

Please see Trainer component guide for more introduction.

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.

Alternatively, you can clean up individual resources by visiting each consoles: