hover 0.7.0

Label data at scale. Fun and precision included.

Explore and label on a map of raw data.

Get enough to feed your model in no time.

hover speeds up data labeling through embedding + visualization + callbacks.

• You just need raw data and an embedding to start.

:sparkles: Features

It's fast because it labels in bulk.

:telescope: A 2D-embedded view of your dataset for labeling, equipped with

• Tooltip for each point and table view for groups of points.
• Search widgets for ad-hoc highlight of data matching search criteria.
• Toggle buttons that clearly distinguish data subsets ("raw"/"train"/"dev"/"test").

It's accurate because you can filter and extend.

:microscope: Supplementary views to provide further labeling precision, such as

• Advanced search view which can filter points by search criteria and provides stronger highlight.
• Active learning view which puts a model in the loop and can filter by confidence score.
• Function-based view which can leverage custom functions for labeling and filtering.

It's fun because the process never gets old.

• Explore the map to find out which "zones" are easy and which ones are tricky.
• Join the conquest of your data by coloring all of those zones through wisdom!

Check out @phurwicz/hover-binder for a list of demo apps.

:rocket: Quickstart

Code + Walkthrough -> Labeling App

• edit & run code right in your browser, with guides along the way.

Jump to Labeling App

• interactive plot for labeling data, pre-built and hosted on Binder.

:package: Install

Python: 3.7+

OS: Linux & Mac & Windows

PyPI (for all releases): pip install hover

Conda-forge (for 0.6.0 and above): conda install -c conda-forge hover

For Windows users, we recommend Windows Subsystem for Linux.

• On Windows itself you will need C++ build tools for dependencies.

:book: Resources

• Binder repo
• Changelog
• Documentation
• Tutorials

:flags: Project News

• Feb 25, 2022 version 0.7.0 is now available. Check out the changelog for details :partying_face:. Some tl-dr for the impatient:
  • audio and image support supply audio/image files through URLs to label with hover!
    • any type supported by HTML (and your browser) will be supported here.
  • high-dimensional support you can now use higher-than-2D embeddings.
    • hover still plots in 2D, but you can dynamically choose which two dimension to use.

:bell: Remarks

Shoutouts

• Thanks to Bokeh because hover would not exist without linked plots and callbacks, or be nearly as good without embeddable server apps.
• Thanks to Philip Vollet for sharing hover with the community even when it was really green.

Contributing

• All feedbacks are welcome, especially what you find lacking and want it fixed!
• ./requirements-dev.txt lists required packages for development.
• Pull requests are advised to use a superset of the pre-commit hooks listed in .pre-commit-config.yaml.

Citation

If you have found hover useful to your work, please let us know :hugs:

@misc{hover,
  title={{hover}: label data at scale},
  url={https://github.com/phurwicz/hover},
  note={Open software from https://github.com/phurwicz/hover},
  author={
    Pavel Hurwicz and
    Haochuan Wei},
  year={2021},
}