Pre-trained model

[danstowell]

Hi - Thanks for publishing this code, and all its documentation too. I was wondering if there is a pretrained model that you share anywhere? I don't see it, but I may have missed it. No worries either way. I guess it depends to some extent whether you intend others to use the pretrained weights, or to retrain for themselves.

[NickleDave]

Hi @danstowell !

Hi - Thanks for publishing this code, and all its documentation too. I was wondering if there is a pretrained model that you share anywhere? I don't see it, but I may have missed it. No worries either way.

Sorry it's a bit hidden: there's a separate README for the article code with a link to download pre-trained checkpoints here

https://github.com/yardencsGitHub/tweetynet/tree/master/article#4-download-datasets

I guess it depends to some extent whether you intend others to use the pretrained weights, or to retrain for themselves.

A functioning script to directly reproduce a final result is still a WIP, partly because we're waiting for Dryad review (as in #211)

As of right now there's no easy way to finetune models with those checkpoints (see https://github.com/NickleDave/vak/issues/5) but since that is something users have asked for, will be added in the near-distant future

Does that answer your question?

[NickleDave]

Also: happy to share additional results if those exact checkpoints are not what you're looking for. The total is ~300 GB 😬 You can also feel free to email @yardencsGitHub and me directly if you prefer: nicholdav at gmail + yardencsmail at gmail

[danstowell]

Thanks, looks good! I'll close this issue, but might come back if I hit a problem. I think I realise now that since each model is per-bird there won't be many use-cases for building on a pretrained model. It's useful anyway, to access the pre-trained models while reproducing the results. Thanks

[NickleDave]

Sure thing!

You might find this relevant if it's the kind of thing you have in mind: https://www.biorxiv.org/content/10.1101/2022.02.24.481827v2

Stay tuned for more along those lines.

Just let us know what else we can do to help.

[NickleDave]

Hi @danstowell it's been a while but just in case you're still interested, models pre-trained on 36 syllable classes from all 4 birds + a background class are available here:

• https://github.com/vocalpy/Nicholson-Cohen-SfN-2023-poster
• https://osf.io/3yv8s/

The exact dataset used is here: https://zenodo.org/records/10098250

This is part of a work in progress so look for more models along these lines. If you have any feedback on the results shared in the above repo, we'd love to hear them