This version removes a few accidentally committed model files from the repo. Functionality should remain unchanged.

This version contains multiple bugfixes and performance improvements.

• The variant quality score predictor runs in parallel now, leading to a measurable speedup
• Nice progress bars for variant calling (you can disable with the --no-prog option)
• Improved management of batching and multiprocessing during calling, leading to slightly higher performance and better GPU utilization
• Fix all random seeds for repeatability
• Lots of other code polishing and cleanup

This version includes new models that offer improved accuracy, particularly for SNVs. The new transformer good44fix_epoch280.model and classifier g44e280_clf.model models deliver higher precision and sensitivity for SNVs, and similar indel performance (see charts in the README).
We also fixed an issue where an exception would be raised if the output variants were out of order. The exception caused calling to stop immediately, and could lead to a hang. It is still possible for output variants to be out of order in some cases, but now calling will finish correctly .

This release offers much improved performance for variant calling along with improved testing and lots of code refactoring and cleanup. There's also a new classifier model models/paraclf.model, which offers slightly higher accuracy than the previous classifier. Most people should use this release unless they're trying to closely replicate the original results.

Original code release for Jenever, as it was used to generate results for the manuscript.

Known issues:

• This code contains a bug in the buildclf.py training code used to create classifier models. The classifier models are incorrect and won't generate accurate scores (for the paper, models were generated in a separate notebook).