The StatisticsGen TFX Pipeline Component

The StatisticsGen TFX pipeline component generates features statistics over both training and serving data, which can be used by other pipeline components. StatisticsGen uses Beam to scale to large datasets.

  • Consumes: datasets created by an ExampleGen pipeline component.
  • Emits: Dataset statistics.

StatisticsGen and TensorFlow Data Validation

StatisticsGen makes extensive use of TensorFlow Data Validation for generating statistics from your dataset.

Using the StatsGen Component

A StatisticsGen pipeline component is typically very easy to deploy and requires little customization. Typical code looks like this:

compute_eval_stats = StatisticsGen(

Using the StatsGen Component With a Schema

For the first run of a pipeline, the output of StatisticsGen will be used to infer a schema. However, on subsequent runs you may have a manually curated schema that contains additional information about your data set. By providing this schema to StatisticsGen, TFDV can provide more useful statistics based on declared properties of your data set.

In this setting, you will invoke StatisticsGen with a curated schema that has been imported by an ImporterNode like this:

user_schema_importer = Importer(
    source_uri=user_schema_dir, # directory containing only schema text proto

compute_eval_stats = StatisticsGen(

Creating a Curated Schema

Schema in TFX is an instance of the TensorFlow Metadata Schema proto. This can be composed in text format from scratch. However, it is easier to use the inferred schema produced by SchemaGen as a starting point. Once the SchemaGen component has executed, the schema will be located under the pipeline root in the following path:


Where <artifact_id> represents a unique ID for this version of the schema in MLMD. This schema proto can then be modified to communicate information about the dataset which cannot be reliably inferred, which will make the output of StatisticsGen more useful and the validation performed in the ExampleValidator component more stringent.

More details are available in the StatisticsGen API reference.