- Add support for custom encoder models. See https://github.com/flekschas/peax-avocado for an example
- Add support for custom data directory via the
--base-data-dir
argument ofstart.py
This is the version described in our paper (Lekschas et al., 2020).
- Add support for changing the classifier via the config file's
classifier
andclassifier_params
properties - Add ability to show more than 5 results per page via to HiGlass' new view scrolling
- Add legend to the UMAP projection view and highlight
- Add legend for the x-axis to the progress views
- Add explanatory help for the progress views
- Add ability to sort selected window by prediction probability
- Add ability to normalize tracks within the same window by setting
normalize_tracks: true
in the config file. This is useful for exploring differential peaks. - Add ability to show the window with the highest prediction probability in the query view instead of a fixed search window by setting
variable_target: true
in the config file. This is useful for exploring pre-loaded labels where there is no defined search query - Improve the visibility of the Re-Train and Compute Projection buttons
- Update HiGlass to
v1.7
- Fix several minor bugs
- Support multitrack search
- Update active learning sampling strategies
- Make windows selectable
- Visualize the progress of the actively-learned classifier
- Update and expand examples to use our newly trained autoencoders
- Tons of bug fixes and performance improvements
- Represent datasets and encoders as classes rather than loose collections of
dict
s - Make autoencoders optional, i.e., some tracks might be encodable (and thus searchable) but the encoding might not be decodable.
- Cache chunked, encoded, and autoencoded data as HDF5 files to avoid having to hold everything in memory.
- Replace internal scatterplot with
regl-scatterplot
- Update the UI to the latest version of HiGlassApp
- Remove Bootstrap v3, which is used by HiGlass's view header, as it's not needed and imposes a security risk
- Search across a single bigWig datasets over up to a few chromosomes