Writing custom datasets

Follow this guide to create a new dataset (either in TFDS or in your own repository).

Check our list of datasets to see if the dataset you want is already present.


Datasets are distributed in all kinds of formats and in all kinds of places, and they're not always stored in a format that's ready to feed into a machine learning pipeline. Enter TFDS.

TFDS process those datasets into a standard format (external data -> serialized files), which can then be loaded as machine learning pipeline (serialized files -> tf.data.Dataset). The serialization is done only once. Subsequent access will read from those pre-processed files directly.

Most of the preprocessing is done automatically. Each dataset implements a subclass of tfds.core.DatasetBuilder, which specifies:

  • Where the data is coming from (i.e. its URLs);
  • What the dataset looks like (i.e. its features);
  • How the data should be split (e.g. TRAIN and TEST);
  • and the individual examples in the dataset.

Write your dataset

Default template: tfds new

Use TFDS CLI to generate the required template python files.

cd path/to/project/datasets/  # Or use `--dir=path/to/project/datasets/` bellow
tfds new my_dataset

This command will generate a new my_dataset/ folder with the following structure:

    my_dataset.py # Dataset definition
    my_dataset_test.py # Test
    dummy_data/ # Fake data (used for testing)
    checksum.tsv # URL checksums (see `checksums` section).

Search for TODO(my_dataset) here and modify accordingly.

If you're a googler, please use gtfds new instead.

Dataset example

All datasets are implemented using a subclasses of tfds.core.DatasetBuilder:

This tutorial uses tfds.core.GeneratorBasedBuilder, but implementing tfds.core.BeamBasedBuilder is very similar (see our beam dataset guide)

Here is a minimal example of dataset class:

class MyDataset(tfds.core.GeneratorBasedBuilder):
  """DatasetBuilder for my_dataset dataset."""

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

  def _info(self) -> tfds.core.DatasetInfo:
    """Dataset metadata (homepage, citation,...)."""
    return tfds.core.DatasetInfo(
            'image': tfds.features.Image(shape=(256, 256, 3)),
            'label': tfds.features.ClassLabel(names=['no', 'yes']),

  def _split_generators(self, dl_manager: tfds.download.DownloadManager):
    """Download the data and define splits."""
    extracted_path = dl_manager.download_and_extract('http://data.org/data.zip')
    train_path = os.path.join(extracted_path, 'train_images')
    test_path = os.path.join(extracted_path, 'test_images')
    # `**gen_kwargs` are forwarded to `_generate_examples`
    return [
        tfds.core.SplitGenerator('train', gen_kwargs=dict(path=train_path)),
        tfds.core.SplitGenerator('test', gen_kwargs=dict(path=test_path)),

  def _generate_examples(self, path) -> Iterator[Tuple[Key, Example]]:
    """Generator of examples for each split."""
    for filename in tf.io.gfile.listdir(path)
      # Yields (key, example)
      yield filename, {
          'image': os.path.join(path, filename),
          'label': 'yes' if filename.startswith('yes_') else 'no',

Let's see in detail the 3 abstract methods to overwrite.

_info: dataset metadata

_info returns the tfds.core.DatasetInfo containing the dataset metadata.

def _info(self):
  return tfds.core.DatasetInfo(
      # Description and homepage used for documentation
      Markdown description of the dataset. The text will be automatically
      stripped and dedent.
          'image_description': tfds.features.Text(),
          'image': tfds.features.Image(),
          # Here, 'label' can be 0-4.
          'label': tfds.features.ClassLabel(num_classes=5),
      # If there's a common `(input, target)` tuple from the features,
      # specify them here. They'll be used if as_supervised=True in
      # builder.as_dataset.
      supervised_keys=("image", "label"),
      # Bibtex citation for the dataset
               author = {Smith, John},"}

Most fields should be self-explainatory. Some precisions:

  • features: This specify the dataset structure, shape,... See the available features or the feature connector guide for more info.
  • citation: To find the BibText citation:
    • Search the dataset website for citation instruction (use that in BibTex format).
    • For arXiv papers: find the paper and click the BibText link on the right-hand side.
    • Find the paper on Google Scholar and click the double-quotation mark underneath the title and on the popup, click BibTeX.
    • If there is no associated paper (for example, there's just a website), you can use the BibTeX Online Editor to create a custom BibTeX entry (the drop-down menu has an Online entry type).

_split_generators: downloads and splits data

Downloading and extracting source data

Most datasets need to download data from the web. This is done using the tfds.download.DownloadManager input argument of _split_generators. dl_manager has the following methods:

  • download: supports http(s)://, ftp(s)://
  • extract: currently supports .zip, .gz, and .tar files.
  • download_and_extract: Same as dl_manager.extract(dl_manager.download(urls))

Those methods supports arbitrary nested structure (list, dict), like:

    'foo': 'https://example.com/foo.zip',
    'bar': 'https://example.com/bar.zip',
})  # return {'foo': '/path/to/extracted_foo/', 'bar': '/path/extracted_bar/'}

Manual download and extraction

For data that cannot be automatically downloaded (e.g. require a login), the user will manually download the source data and place it in manual_dir/ (defaults to ~/tensorflow_datasets/manual/), which you can access with dl_manager.manual_dir .

Read archive directly

dl_manager.iter_archive reads an archives sequencially without extracting them. This can save storage space and improve performances on some file systems.

for filename, fobj in dl_manager.iter_archive('path/to/archive.zip'):

fobj has the same methods as with open('rb') as fobj: (e.g. fobj.read())

Specifying dataset splits

If the dataset comes with pre-defined splits (e.g. MNIST has train and test splits), keep those. Otherwise, only specify a single tfds.Split.TRAIN split. Users can dynamically create their own subsplits with the subsplit API (e.g. split='train[80%:]').

def _split_generators(self, dl_manager):
  # Download source data
  extracted_path = dl_manager.download_and_extract(...)

  # Specify the splits
  return [
              "images_dir_path": os.path.join(extracted_path, "train_imgs"),
              "labels": os.path.join(extracted_path, "train_labels.csv"),
              "images_dir_path": os.path.join(extracted_path, "test_imgs"),
              "labels": os.path.join(extracted_path, "test_labels.csv"),

SplitGenerator describes how a split should be generated. gen_kwargs will be passed as keyword arguments to _generate_examples, which we'll define next.

_generate_examples: Example generator

_generate_examples generates the examples for each split from the source data. For the TRAIN split with the gen_kwargs defined above, _generate_examples will be called as:

    images_dir_path=os.path.join(extracted_path, 'train_imgs'),
    labels=os.path.join(extracted_path, 'train_labels.csv'),

This method will typically read source dataset artifacts (e.g. a CSV file) and yield (key, feature_dict) tuples that correspond to the features specified in tfds.core.DatasetInfo.

  • key: unique identifier of the example. Used to deterministically suffle the examples using hash(key). If two examples use the same key, an exception will be raised.
  • feature_dict: A dict containing the example values. The structure should match the features kwarg defined in tfds.core.DatasetInfo. See the feature connector guide for more info.
def _generate_examples(self, images_dir_path, labels):
  # Read the input data out of the source files
  with tf.io.gfile.GFile(labels) as f:
    for row in csv.DictReader(f):
      image_id = row['image_id']
      # And yield (key, feature_dict)
      yield image_id, {
          'image_description': row['description'],
          'image': os.path.join(images_dir_path, f'{image_id}.jpeg'),
          "label": row['label'],

File access and tf.io.gfile

In order to support Cloud storage systems, use tf.io.gfile API instead of built-in for file operations. Ex:

Extra dependencies

Some datasets require additional Python dependencies only during generation. For example, the SVHN dataset uses scipy to load some data.

If you're adding dataset into the TFDS repository, please use tfds.core.lazy_imports to keep the tensorflow-datasets package small. Users will install additional dependencies only as needed.

To use lazy_imports:

  • Add an entry for your dataset into DATASET_EXTRAS in setup.py. This makes it so that users can do, for example, pip install 'tensorflow-datasets[svhn]' to install the extra dependencies.
  • Add an entry for your import to LazyImporter and to the LazyImportsTest.
  • Use tfds.core.lazy_imports to access the dependency (for example, tfds.core.lazy_imports.scipy) in your DatasetBuilder.

Corrupted data

Some datasets are not perfectly clean and contain some corrupt data (for example, the images are in JPEG files but some are invalid JPEG). These examples should be skipped, but leave a note in the dataset description how many examples were dropped and why.

Dataset configuration/variants (tfds.core.BuilderConfig)

Some datasets may have multiple variants, or options for how the data is preprocessed and written to disk. For example, cycle_gan has one config per object pairs (cycle_gan/horse2zebra, cycle_gan/monet2photo,...).

This is done through tfds.core.BuilderConfigs:

  1. Define your own configuration object as a subclass of tfds.core.BuilderConfig. For example, MyDatasetConfig.
  2. Define the BUILDER_CONFIGS = [] class member in MyDataset that lists MyDatasetConfigs that the dataset exposes.
  3. Use self.builder_config in MyDataset to configure data generation. This may include setting different values in _info() or changing download data access.


  • Each config has a unique name. The fully qualified name of a config is dataset_name/config_name (e.g. coco/2017).
  • If not specified, the first config in BUILDER_CONFIGS will be used (e.g. tfds.load('c4') default to c4/en)

See anli for an example of a dataset that uses BuilderConfigs.


Version can refer to two different meaning:

  • The "external" original data version: e.g. COCO v2019, v2017,...
  • The "internal" TFDS code version: e.g. rename a feature in tfds.features.FeatureDict, fix a bug in _generate_examples

To update a dataset:

  • For "external" data update: Multiple users may want to access a specific year/version simlutaneously. This is done by using one tfds.core.BuilderConfig per version (e.g. coco/2017, coco/2019) or one class per version (e.g. Voc2007, Voc2012).
  • For "internal" code update: Users only download the most recent version. Any code update should increase the VERSION class attribute (e.g. from 1.0.0 to VERSION = tfds.core.Version('2.0.0')) following semantic versioning.

Add an import for registration

Don't forget to import the dataset module to your project __init__ to be automatically registered in tfds.load, tfds.builder.

import my_project.datasets.my_dataset  # Register MyDataset

ds = tfds.load('my_dataset')  # MyDataset available

For example, if you're contributing to tensorflow/datasets, add the module import to its subdirectory's __init__.py (e.g. image/__init__.py.

Test your dataset

Run the generation code

From the my_dataset/ directory, run download_and_prepare to ensure that data generation works:

cd path/to/my_dataset/
python -m tensorflow_datasets.scripts.download_and_prepare \
    --register_checksums \
    --module_import=my_dataset \
  • --datasets: The dataset to build (e.g. my_dataset/config)
  • --module_import: Required if the dataset is defined outside of TFDS, so your class can be detected (e.g. my_project.submodules.datasets)
  • --register_checksums: Record the checksums of downloaded urls. Should only be used while in development.


It is recommanded to record the checksums of your datasets to guarantee determinism, help with documentation,... This is done by generating the dataset with the --register_checksums (see previous section).

If you are releasing your datasets through PyPI, don't forget to export the checksums.tsv files (e.g. in the package_data of your setup.py).

Unit-test your dataset

tfds.testing.DatasetBuilderTestCase is a base TestCase to fully exercise a dataset. It uses "dummy data" as test data that mimic the structure of the source dataset.

  • The test data should be put in my_dataset/dummy_data/ directory and should mimic the source dataset artifacts as downloaded and extracted. It can be created manually or automatically with a script (example script).
  • Make sure to use different data in your test data splits, as the test will fail if your dataset splits overlap.
  • The test data should not contain any copyrighted material. If in doubt, do not create the data using material from the original dataset.
import tensorflow_datasets as tfds
from . import my_dataset

class MyDatasetTest(tfds.testing.DatasetBuilderTestCase):
  """Tests for my_dataset dataset."""
  DATASET_CLASS = my_dataset.MyDataset
  SPLITS = {
      'train': 3,  # Number of fake train example
      'test': 1,  # Number of fake test example

  # If you are calling `download/download_and_extract` with a dict, like:
  #   dl_manager.download({'some_key': 'http://a.org/out.txt', ...})
  # then the tests needs to provide the fake output paths relative to the
  # fake data directory
      "name1": "path/to/file1",  # Relative to dummy_data/my_dataset dir.
      "name2": "file2",

if __name__ == '__main__':

Run the following command to test the dataset.

python my_dataset_test.py

Send us feedback

We are continously trying to improve the dataset creation workflow, but can only do so if we are aware of the issues. Which issues, errors did you encountered while creating the dataset ? Was there a part which was confusing, boilerplate or wasn't working the first time ? Please share your feedback on github.