ArcFace embedding task is broken

[olvb]

Describe the bug

The SpeakerEmbedding / SupervisedRepresentationLearningWithArcFace cannot be used anymore for training since commit 0fb42ad08cb92539a37702cb7afcc20dad8871d7. It is missing an implementation of the collate_fn method.

Is this known/expected? I see that this task is not tested in pyannote-audio/tests/test_train.py

To Reproduce

import pytorch_lightning as pl

from pyannote.database import get_protocol, FileFinder
from pyannote.audio.models.embedding.debug import SimpleEmbeddingModel
from pyannote.audio.tasks import SpeakerEmbedding

protocol = get_protocol(
    "Debug.SpeakerDiarization.Debug",
    preprocessors={"audio": FileFinder()}
)

emb = SpeakerEmbedding(protocol)
model = SimpleEmbeddingModel(task=emb)
trainer = pl.Trainer(max_epochs=1)
_ = trainer.fit(model)

[hbredin]

Thanks for reporting this.

This is indeed expected but should definitely be fixed -- but unfortunately not at the top of my priority list (as I have been using speechbrain speaker embedding instead).

Would you like to contribute a fix? Or at least a test?

[olvb]

Sure, I can give it a try! I suppose restoring the previous behavior relying on default_collate should be enough?

Also, do you fine tune the speechbrain ecapa embeddings or do you directly use the ones pretrained on VoxCeleb?

[hbredin]

Sure, I can give it a try! I suppose restoring the previous behavior relying on default_collate should be enough?

"Enough", I am not 100% sure. But it definitely is "needed".

Also, do you fine tune the speechbrain ecapa embeddings or do you directly use the ones pretrained on VoxCeleb?

I use them directly for the time being...

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[hbredin]

@olvb should be fixed in develop branch.

[olvb]

Thanks! Will give it a try