TensorFlow Datasets (TFDS)

Google's open-source dataset library that downloads and loads a dataset into a trainable pipeline in one line of code.

TFDS is an open-source dataset library maintained by Google, cataloging thousands of ready-to-use datasets spanning images, text, audio, robotics, and more. Every dataset has a builder, responsible for downloading the raw data, converting it through a fixed pipeline, and storing it as sharded TFRecord or ArrayRecord files; calling tfds.load returns a tf.data.Dataset, which can also be converted to NumPy for use with JAX or PyTorch, along with a version number and metadata. It's a layer built on top of tf.data, not tf.data itself. In embodied AI, the RLDS format and the Open X-Embodiment dataset are both released following the TFDS convention, and the training code for Octo and OpenVLA reads data through it as well.

ExamplePointing tfds.builder_from_directory at Open X-Embodiment's fractal20220817_data directory on Google Cloud Storage lets you iterate through the RT-1 dataset episode by episode, getting images, language instructions, and actions.

Also called

tensorflow_datasets

Related

RLDS (Reinforcement Learning Datasets) · TFRecord · Open X-Embodiment · TensorFlow · LeRobotDataset

Sources

TensorFlow Datasets Overview
google-deepmind/open_x_embodiment (GitHub)
google-research/rlds (GitHub)