Some question about OSD in pyannote v2 and v1

[Mintcat10]

Thanks for share this project sincerely.
In paper, OSD in v2(develop) shows a significant improvement.
What are the changes to OSD in v2?
What are there besides number of LSTM layers, training data?
Such as data aug or other things?

About optimizing threshold , this code runs very slow. Do I use it in the wrong way?

    from pyannote.audio import Inference, Model

    model = Model.from_pretrained(checkpoint)
    model.eval()
    inference = Inference(model, device=torch.device("cuda:0"))
    validation_files = list(protocol.development())
    for file in validation_files:
        file['osd'] = inference(file)
    pipeline = OverlappedSpeechDetectionPipeline(segmentation=checkpoint)    
    optimizer = Optimizer(pipeline)