Product feedback for the TensorFlow Datasets team

Overview

TensorFlow Datasets (TFDS) is a collection of datasets ready to use, with TensorFlow or other Python ML frameworks, such as JAX. Consuming datasets with TFDS enables your team to build easy-to-use and highly performant input pipelines.

The TensorFlow team is excited to announce that we would like to work with external teams who are using TensorFlow Datasets in their projects.

Did you know you can use TFDS for both public and private datasets?

Qualifications:

We’re looking to work with companies who have a use case that falls into either of the following category:

  • Public use: Anyone interested in sharing your dataset to the publicly available repository hosted on TFDS. This involves giving authorization to TFDS to host and redistribute the data ourselves.
  • Private use: Any enterprise, company, or institution that would like to utilize TFDS for internal datasets, while benefiting from its features such as extraction, formatting, and distribution but would like to keep access to the data restricted.

What does working together with TFDS mean?

We’d like to work together and provide technical support to help you integrate TFDS. This includes GitHub question support, and a consultation with the TFDS team.

In return, we're looking to work with companies that will give us feedback. This includes companies across all industry segments that include use cases from startups, large enterprise companies or academic institutions. Working together will help us ensure we build the right set of features that supports the TFDS community.

Why TFDS

TFDS already contains hundreds of public datasets, with an easy and flexible API to access datasets of arbitrary size.

For partners who have a private use scenario, integrating your own internal dataset enables easy transition between running/testing on public and internal data.

In addition, adding your dataset to TFDS makes accessing it easier for additional departments in your organization.

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.

TFDS provides a way to transform all those datasets into a standard format, do the preprocessing necessary to make them ready for a machine learning pipeline, and provides a standard input pipeline.

Datasets provided through TFDS offers many features:

  • Expose datasets metadata through info objects (label names, number of examples, dataset size)
  • Use the default datasets for training and test splits or dynamically create your own custom sub-splits
  • Versioning

Examples

Google Research utilizes TFDS heavily across projects.

Visualization

Google's AI Research team, released a research paper on T5, a shared text-to-text framework, that produces state-of-the-art results on many NLP benchmarks while being flexible enough to be fine-tuned to a variety of important downstream tasks. Some applications of T5 include closed book questions and fill in the blank text generation. The pre-training objective used by T5 aligns more closely with a fill-in-the-blank task where the model predicts missing words within a corrupted piece of text. The team utilized TFDS for handling the dataset they created and used to train their models.

Visualization

In addition, Google’s research team has utilized pre-trained representations in TFDS to make advances in the field of Computer Vision. In VTAB, they evaluated visual models on a challenging set of downstream vision tasks, coming from diverse domains: natural images, artificial environments (structured) and images captured with non-standard cameras (specialized). The availability of the datasets helped enable their research to improve extracting features from images.

Visualization

Lastly, TFDS has been used in another NLP example. Google researcher's have released EXTREME, a set of language tasks and benchmarks to encourage more research on multilingual learning.

You can check out more repos that are using TFDS here.

Contact

Please contact us at tfds-feedback@tensorflow.org and the Tensorflow Datasets team will reach out to you.