Generating big datasets with Apache Beam

Some datasets are too big to be processed on a single machine. tfds supports generating data across many machines by using Apache Beam.

This doc has two sections:

  • For user who want to generate an existing Beam dataset
  • For developper who want to create a new Beam dataset

Table of content:

Generating a Beam dataset

Below are different examples of generating a Beam dataset, both on the cloud or locally.

On Google Cloud Dataflow

To run the pipeline using Google Cloud Dataflow and take advantage of distributed computation, first follow the Quickstart instructions.

Once your environment is set up, you can run the download_and_prepare script using a data directory on GCS and specifying the required options for the --beam_pipeline_options flag.

To make it easier to launch the script, it's helpful to define the following variables using the actual values for your GCP/GCS setup and the dataset you want to generate:


You will then need to create a file to tell Dataflow to install tfds on the workers:

echo "tensorflow_datasets[$DATASET_NAME]" > /tmp/beam_requirements.txt

If you're using tfds-nightly, make sure to to echo from tfds-nightly in case the dataset has been updated since the last release.

echo "tfds-nightly[$DATASET_NAME]" > /tmp/beam_requirements.txt

Finally, you can launch the job using the command below:

python -m tensorflow_datasets.scripts.download_and_prepare \
  --data_dir=$GCS_BUCKET/tensorflow_datasets \


To run your script locally using the default Apache Beam runner, the command is the same as for other datasets:

python -m tensorflow_datasets.scripts.download_and_prepare \

With a custom script

To generate the dataset on Beam, the API is the same as for other datasets, but you have to pass the Beam options or runner to the DownloadConfig.

# If you are running on Dataflow, Spark,..., you may have to set-up runtime
# flags. Otherwise, you can leave flags empty [].
flags = ['--runner=DataflowRunner', '--project=<project-name>', ...]

# To use Beam, you have to set at least one of `beam_options` or `beam_runner`
dl_config =

data_dir = 'gs://my-gcs-bucket/tensorflow_datasets'
builder = tfds.builder('wikipedia/20190301.en', data_dir=data_dir)

Implementing a Beam dataset


In order to write Apache Beam datasets, you should be familiar with the following concepts:


If you are familiar with the dataset creation guide, adding a Beam dataset only requires a few modifications:

  • Your DatasetBuilder will inherit from tfds.core.BeamBasedBuilder instead of tfds.core.GeneratorBasedBuilder.
  • Beam datasets should implement the abstract method _build_pcollection(self, **kwargs) instead of the method _generate_examples(self, **kwargs). _build_pcollection should return a beam.PCollection with the examples associated with the split.
  • Writing a unit test for your Beam dataset is the same as with other datasets.

Some additional considerations:

  • Use tfds.core.lazy_imports to import Apache Beam. By using a lazy dependency, users can still read the dataset after it has been generated without having to install Beam.
  • Be careful with Python closures. When running the pipeline, the beam.Map and beam.DoFn functions are serialized using pickle and sent to all workers. This can create bugs; for instance, if you are using a mutable object in your functions which has been declared outside of the function, you may encounter pickle errors or unexpected behavior. The fix is typically to avoid mutating closed-over objects.
  • Using methods on DatasetBuilder in the Beam pipeline is fine. However, the way the class is serialized during pickle, changes done to features during creation will be ignored at best.


Here is an example of a Beam dataset. For a more complicated real example, have a look at the Wikipedia dataset.

class DummyBeamDataset(tfds.core.BeamBasedBuilder):

  VERSION = tfds.core.Version('1.0.0')

  def _info(self):
    return tfds.core.DatasetInfo(
            'image': tfds.features.Image(shape=(16, 16, 1)),
            'label': tfds.features.ClassLabel(names=['dog', 'cat']),

  def _split_generators(self, dl_manager):
    return [

  def _build_pcollection(self, pipeline, file_dir):
    """Generate examples as dicts."""
    beam = tfds.core.lazy_imports.apache_beam

    def _process_example(filename):
      # Use filename as key
      return filename, {
          'image': os.path.join(file_dir, filename),
          'label': filename.split('.')[1],  # Extract label: ""

    return (
        | beam.Create(
        | beam.Map(_process_example)

Running your pipeline

To run the pipeline, have a look at the above section.

python -m tensorflow_datasets.scripts.download_and_prepare \
  --register_checksums \