finetune

The MIT License (MIT)

Copyright (c) 2017-2020 CNRS

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

AUTHOR Hervé Bredin - http://herve.niderb.fr

Fine-tuning a pretrained model to your own data

This tutorial assumes that you have already followed the data preparation tutorial and teaches how to fine-tune a pretrained model to your own data.

The list of pretrained models available in pyannote.audio can be obtained with:

import torch
torch.hub.list('pyannote/pyannote-audio')

More precisely, we will fine-tune a speech activity detection model (which was originally trained on the DIHARD dataset) to the AMI dataset.

It is recommended to (at least) read this tutorial first, especially the part related to the configuration file.

Table of contents

• Citation
• Configuration
• Fine tuning
• More options

Citation

(↑up to table of contents)

If you use pyannote-audio, please cite the following paper:

@inproceedings{Bredin2020,
  Title = {{pyannote.audio: neural building blocks for speaker diarization}},
  Author = {{Bredin}, Herv{\'e} and {Yin}, Ruiqing and {Coria}, Juan Manuel and {Gelly}, Gregory and {Korshunov}, Pavel and {Lavechin}, Marvin and {Fustes}, Diego and {Titeux}, Hadrien and {Bouaziz}, Wassim and {Gill}, Marie-Philippe},
  Booktitle = {ICASSP 2020, IEEE International Conference on Acoustics, Speech, and Signal Processing},
  Address = {Barcelona, Spain},
  Month = {May},
  Year = {2020},
}

Configuration

(↑up to table of contents)

Because pretrained models come with their own configuration file, no configuration file is needed to fine-tune a model.

$ export EXP_DIR=tutorials/finetune
$ cat ${EXP_DIR}/config.yml
cat: config.yml: No such file or directory

Note that you can still choose to provide one (possibly partial) configuration file to override sections of the pretrained model configuration file.

# this (partial) configuration file overrides the batch size
task:
   name: SpeechActivityDetection
   params:
      batch_size: 16

Fine tuning

(↑up to table of contents)

The following command will fine-tune the sad_dihard pretrained model using the training subset of AMI database for 5 epochs:

$ pyannote-audio sad train --pretrained=sad_dihard --subset=train --to=5 --parallel=4 ${EXP_DIR} AMI.SpeakerDiarization.MixHeadset

Your model is now fine-tuned! Note that a configuration file has been created for you in EXP_DIR.

More options

(↑up to table of contents)

You may also use your own pretrained model. Simply provide the absolute path to the model checkpoint instead of sad_dihard:

$ pyannote-audio sad train \
  --pretrained=/path/to/train/YourDataset.SpeakerDiarization.YourProtocol.train/weights/0020.pt \
  --subset=train --to=5 --parallel=4 \
  ${EXP_DIR} AMI.SpeakerDiarization.MixHeadset

Go check this tutorial to learn how to validate and use your brand new model.

That's all folks!