Introducing pyannote.database.registry

[FrenchKrab]

PR summary

Previously, only a single configuration file could be loaded. All data was stored in global variables in the pyannote.database __init__.py file. The way things were loaded meant it was not possible to use multiple configuration files ("database.yml").

To solve that, a new class PyannoteDbConfig is responsible for loading yml files and "merging" the different configuration files . This simply means merging the defined databases and handling cases where the same protocols are defined multiple times.

Most of the logic remains similar, with global functions simply being moved to the new class, updating its state instead of global variables. Only the FileFinder class had some of its path resolving logic moved upstream to PyannoteDbConfig (because it relies on the database.yml path used, which FileFinder can't know).

What's left

• How to handle meta protocols defined multiple times
• Prepare datasets to replace entry points (for VoxCeleb)

[hbredin]

For some reason I cannot edit this PR. Can you make it editable by me?

[FrenchKrab]

I (should have) fixed the raised issues (down to 7 edited files, that's indeed better)

[FrenchKrab]

These commits should fix the raised issues

[hbredin]

I have found two packages where the change in FileFinder API would break things:

• pyannote.audio: https://github.com/pyannote/pyannote-audio/blob/6927ced52de68b6be978d3417acadee4d4ca5bf2/pyannote/audio/core/pipeline.py#L162
• pyannote.pipeline: https://github.com/pyannote/pyannote-pipeline/blob/3a8012883f1ed2e5f404439dcc039088dd342b02/pyannote/pipeline/experiment.py#L194

Not sure what to do about this yet but I wanted to keep track of this potential problem.

[hbredin]

@FrenchKrab I would not mind a final review on this (but that can wait for next week)