feat: send pipeline to device with Pipeline.to(device)

[hbredin]

Fixes #1108 #1095 #1155 #1059 #1187

[codecov]

Codecov Report

Base: 29.91% // Head: 29.75% // Decreases project coverage by -0.16% ⚠️

Coverage data is based on head (79845fb) compared to base (7a60da9).
Patch coverage: 14.28% of modified lines in pull request are covered.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #1204      +/-   ##
===========================================
- Coverage    29.91%   29.75%   -0.17%     
===========================================
  Files           62       62              
  Lines         3817     3872      +55     
===========================================
+ Hits          1142     1152      +10     
- Misses        2675     2720      +45     

Impacted Files	Coverage Δ
pyannote/audio/pipelines/multilabel.py	0.00% <0.00%> (ø)
...ote/audio/pipelines/overlapped_speech_detection.py	0.00% <0.00%> (ø)
pyannote/audio/pipelines/resegmentation.py	0.00% <ø> (ø)
pyannote/audio/pipelines/segmentation.py	0.00% <ø> (ø)
pyannote/audio/pipelines/speaker_diarization.py	0.00% <ø> (ø)
pyannote/audio/pipelines/speaker_verification.py	0.00% <0.00%> (ø)
...annote/audio/pipelines/voice_activity_detection.py	0.00% <ø> (ø)
pyannote/audio/core/pipeline.py	21.98% <15.51%> (-6.25%)	⬇️
pyannote/audio/core/inference.py	61.42% <57.14%> (-0.34%)	⬇️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[prashanthellina]

@hbredin When I run the speaker diarization pipeline on latest develop branch, the GPU is not being used for embeddings computation (it used to work before the changes from this PR).

I suspect this piece of code in speaker_verification.py file.

pyannote-audio/pyannote/audio/pipelines/speaker_verification.py

Line 260 in 40783ce

run_opts={"device": self.device},

When I change self.device to device, it works as expected.

[hbredin]

Good catch. Would you make a PR?

[prashanthellina]

Sure, Will do.

[prashanthellina]

@hbredin #1217

[maxiek0071]

@hbredin

Would you mind telling how do I call pipeline.to(device) to decode on CPU?

This is my code:

from pyannote.audio import Pipeline
pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization",
                                    use_auth_token="<my_token>")
pipeline.to("cpu")

and am getting this error:

  File "transcription_module/my_custom_example.py", line 18, in <module>
    pipeline = pipeline.to("cpu")
  File "/home/ubuntu/.local/lib/python3.7/site-packages/pyannote/pipeline/pipeline.py", line 100, in __getattr__
    raise AttributeError(msg)
AttributeError: 'SpeakerDiarization' object has no attribute 'to'

[hbredin]

Pipeline.to API has not been released yet.

Install pyannote.audio from develop branch and it should work.

[maxiek0071]

Thanks, it works now!

[lasagna0]

Hey, for me its not working when using pip install https://github.com/pyannote/pyannote-audio/archive/refs/heads/develop.zip

[donzthefonz]

I'm running Mac M1 and think I've installed the develop branch by running pip install git+https://github.com/pyannote/pyannote-audio.git@develop#egg=pyannote-audio

Does this mean it should automatically use the gpu/mps or do I have to manually set with Pipeline.to("mps")?

The diarization step is still taking longer than I would have expected it to. Though the audio file is 1.5 hours long...

[hbredin]

It will run on CPU by default.

[Qualzz]

Took me a while to notice this post. Nothing would work until I manually add pipeline.to(device) to my code.

Is there any reason why it doesn't use the GPU by default (as written in the README) since CPU would be extremely slow nonetheless ?

Maybe i'm stupid, but I lost hours waiting the process to complete while it would took 300x less with my GPU. Tried multiple torch install, etc.

[lyehe]

Im having the same issue. The pipeline is running on the CPU by default.
BTW, I also tried pipeline.to(device), which broke the kernel.