How to control multi-processing / threading in pipelines

[treya-lin]

Hi I am trying the speaker-diarization pipelines to detect speaker numbers, and I noticed that it can be very cpu hungry and each time I run it it creates a lot of processes and I can see many many processes in htop. My saw that there is the num_workers param that can control the processes used, but only when we call the classes like Segmentation, VAD etc...

here is how I use the model.

checkpoint_path = "/data/resources/models/speaker-diarization-3.1/config.yaml"
pipeline = Pipeline.from_pretrained(
            checkpoint_path
    )
pipeline.to(torch.device(device))

diarization = pipeline(audio)

        speakers = set()
        for turn, _, speaker in diarization.itertracks(yield_label=True):
            start = int(round(turn.start,3)*1000)
            end = int(round(turn.end,3)*1000)
            speaker = int(speaker.split("_")[-1]) # SPEAKER_00, SPEAKER_01
            speakers.add(speaker)

I checked the code and the wrappers seems very deep and I didn't see where I can specify the num_workers param. Anyone has any idea about it? Any advice is appreciated!

[Purfview]

Try this:

 torch.set_num_threads(2)
 pipeline.to(torch.device(device))

[treya-lin]

Thanks. It works!