Train Segmentation and Embedding Model From Scratch

[eggplant95]

Hi, Thank you for your awsome contribution.

I have learnt how to finetune your pretrained segmentation model in https://github.com/pyannote/pyannote-audio/blob/develop/tutorials/training_a_model.ipynb. However, it didn't mention how to train a segmentation from scratch.
I would like to know how could I reproduce your work with the listed data in your paper? Futhermore, how could I train a model from scratch with my own data? Is there any pipeline about the training details?

I would be much appreciated if you could tell me more about these details. Thank you very much.

[juanmc2005]

Hi @eggplant95,

I would like to know how could I reproduce your work with the listed data in your paper?

You need to download both audio and RTTM files for the datasets we used (we are not allowed to provide them ourselves) and then run diart.benchmark.
Notice that you probably won't get the exact same results because I've been working on some small improvements since the paper got published.

Futhermore, how could I train a model from scratch with my own data? Is there any pipeline about the training details?

To train a segmentation model from scratch you can follow the part of the tutorial with the title "Training a voice activity detection model from scratch", but replacing the VoiceActivityDetection task with Segmentation. The architecture used in the tutorial is the same as pyannote/segmentation.
The training configuration is available here, other hyper-parameters like learning rate and batch size will depend on your own data and hardware.

[m15kh]

hi @juanmc2005
i trained model in pyannote and i have a checkpoint so you didn't explain how can i useing my checkpoint in your code!
can you say example for that
thanks for your answering