A TFX component to sample examples.

The sampling component wraps an Apache Beam pipeline to process data in an TFX pipeline. This component loads in tf.Record files from an earlier example artifact, processes the 'train' split by default, samples the split by a given label's classes, and stores the new set of sampled examples into its own example artifact in tf.Record format.

The sampling is probabilistic estimation. Note that in small datasets this may result in worse datasets or such. This module is meant to approximate sampling using probability.

By default, the component will ignore all examples with a null value (more precisely, a value that evaluates to False) for the given label, although more values can be added in as necessary. Additionally, it will copy all non-'train' splits, through this behavior can be changed as well. The component will save the examples in a user-specified number of files, and it can be given a name as well.

import tfx_addons as tfxa

under = tfxa.sampling.Sampler(
  examples=example_gen.outputs['examples'],
  sampling_strategy=tfxa.sampling.SamplingStrategy.UNDERSAMPLE
)

Component outputs contains:

• sampled_examples: Channel of type standard_artifacts.Examples for materialized sampled examples, based on the input splits, which includes copied splits unless otherwise specified by copy_others.

Initial Proposal