parameterized_tfx_oss

Overview

Tensorflow Extended (TFX) is a Google-production-scale machine learning platform based on TensorFlow. It provides a configuration framework to express ML pipelines consisting of TFX components. Kubeflow Pipelines can be used as the orchestrator supporting the execution of a TFX pipeline.

This sample demonstrates how to author a ML pipeline in TFX and run it on a KFP deployment. Please refer to inline comments for the purpose of each step.

In order to successfully compile this sample, you'll need to have a TFX installation at HEAD. First, you can clone their repo and run python setup.py install from tfx/. The image used in the pipeline is specified as tfx_image in the KubeflowDagRunnerConfig. Please make sure you use the latest TFX build so that visualization support is included. List of officially released nightly build image available can be found here).

After that, running python3 chicago_taxi_pipeline_simple.py compiles the TFX pipeline into KFP pipeline package.

Permission

This pipeline requires Google Cloud Storage permission to run. If KFP was deployed through K8S marketplace, please follow instructions in the guideline to make sure the service account has storage.admin role.

Caveats

This sample uses pipeline parameters in a TFX pipeline, which is not yet fully supported. See here for more details. In this sample, however, the path to module file and path to data are parameterized. This is achieved by specifying those objects dsl.PipelineParam and appending them to the KubeflowDagRunner._params. Then, KubeflowDagRunner can correctly identify those pipeline parameters and interpret them as Argo placeholder correctly when compilation. However, this parameterization approach is a hack and we do not have plan for long-term support. Instead we're working with TFX team to support pipeline parameterization using their RuntimeParameter.

Known issues

• This approach only works for string-typed quantities. For example, you cannot parameterize num_steps of Trainer in this way.
• Name of parameters should be unique.
• By default pipeline root is always parameterized with the name pipeline-root.
• If the parameter is referenced at multiple places, the user should make sure that it is correctly converted to the string-formatted placeholder by calling str(your_param).
• The best practice is to specify TFX pipeline root to an empty dir. In this sample Argo automatically do that by plugging in the workflow unique ID (represented kfp.dsl.RUN_ID_PLACEHOLDER) to the pipeline root path.