GPU training and pipe-line training from scratch

[vijaysumaravi]

Hi,

I have two questions related to 'pyannote-pipeline'

1. Fine-tuning pipelines on a GPU.
   I am trying to train the pretrained speaker diarization pipelines on my own dataset. I am running this command from the tutorial:
   pyannote-pipeline train --subset=development --iterations=50 ${EXP_DIR} my.personal.dataset

However, GPU usage is still 0%. The fine-tuning is happening on CPUs instead. I tried providing the CUDA_AVAILABLE_DEVICES but that doesn't help either.
The documentation says GPU is used by default when available but I don't see that happening. I see you've mentioned that the latest branch solves this issue but I have the latest dev branch and I am still facing this. Am I doing something wrong?

2. How do I train the whole pipeline from scratch? The documentation has training details for individual blocks. Is that the only way - train individual blocks separately, if I don't want to use pretrained models?

I was not sure of opening issues for these so I started a discussion thread instead. Thank you!

[hbredin]

However, GPU usage is still 0%. The fine-tuning is happening on CPUs instead. I tried providing the CUDA_AVAILABLE_DEVICES but that doesn't help either.

pyannote.pipeline does not run on GPU for the simple reason that it assumes that all neural network stuff has happened before (when precomputing VAD scores or embeddings.).

The documentation says GPU is used by default when available
Does it? Can you please provide a link?

1. How do I train the whole pipeline from scratch? The documentation has training details for individual blocks. Is that the only way - train individual blocks separately, if I don't want to use pretrained models?

Until a really end-to-end diarization is implemented, that is indeed the only way:

• first train individual models
• then train the diarization pipeline on top of those

I was not sure of opening issues for these so I started a discussion thread instead. Thank you!
I think you made the right choice :)