Omnizart

Official Open JOSS reviewed version.

Bugs
- Fix bug of name conflict while loading chord model.

Accumulated release. Various improvements and bug fix. See details below.

Feature
- Migrate checkpoints from private Google Drive to Github release.
See [here](https://github.com/Music-and-Culture-Technology-Lab/omnizart/releases/tag/checkpoints-20211001)
- Replace opencv

Dependency
- Upgrade Tensorflow version to 2.5.0 for Nvidia 30 series GPU compatibility.
- Upgrade Spleeter version to 2.3.0 for new TF version compatibility.
- Replace Opencv with PIL for drum feature resizing and remove opencv from the dependency.

Enhancement
- Simplify model loading mechanism by unifying the all checkpoint format to use TF format.
- Lazy import extraction functions to boost loading time.
- Change the order of Dockerfile commands for better utilizing cache.

Documentation
- Add notice about compatibility issue of running on certain CPU architecture.
- Add explaination about enabling auto completion.
- Rephrase sentences in paper according to JOSS review feedback.
- Add explaination about installing development dependencies.
- Use pepy as the alternative source for 'download' state badge.


Bugs
- Fix bug of unable to find vocal contour checkpoint.
- Fix bug of fail to custom layers of chord module.
- Fix various unit tests bugs.
- Fix minor linter errors.

Hotfix version according to issue [23](https://github.com/Music-and-Culture-Technology-Lab/omnizart/issues/23)

Feature
- Add a new piano transcription model and set it as the default model while using `music` module.

Bugs
- Fix bug while parsing weight files in the checkpoint folder.

---

Various improvements on music module and some critical bug fixes.

Enhancement
- Improve the peak finding and thresholding strategy for more stable and better performance.
- Modify the feeding strategy of feature slices with adjustable overlapping rate while making predictions.
- Apply learning rate scheduler for music module.
- Replace the usage of custom training loop of music module with the built-in TF `.fit()` function.

Bugs
- Fix a critical bug of inference of music module that would lead to missing onsets.
- Fix generation of pertubation of vocal module while training.

Documentation
- Merge the demo page into master from `build_doc` branch.

---

Hotifx version according to issue 19.

Bugs
- Fix bug of treating numpy array as list while appending elements.

---

Hotfix version according to issue 19.

Bugs
- Fix column inconsistency of `aggregate_f0_info` and `write_agg_f0_results`.
- Update version of dependencies according to the security alert.

---