Pretrained embeddings

[popcornell]

This PR supersedes https://github.com/DCASE-REPO/DESED_task/pull/35 Adding AST and PANNs Audioset-based pretrained models. Right now supports e2e training. Also the script to extract the embedding to use them offline is included.

[popcornell]

@RonFrancesca @turpaultn ready for review. Right now the code supports both fine-tuning of the model and using pre-extracted embeddings

[popcornell]

Once the embeddings are in dropbox and we agree on how participants can download them. I ll add instructions in the readme

[turpaultn]

I have uploaded the embeddings to zenodo:

• https://zenodo.org/record/6541454#.YnzHq2YzbDI (unalabeled ast)
• https://zenodo.org/record/6539466#.YnvtWmYzbAM (ast synth train, ast synth val,ast weak val)
• https://zenodo.org/record/6518380#.YnvWZGYzbAM (panns, ast weak train, ast devtest)

[popcornell]

I added the instructions on the README.md on how to run the baseline with pre-trained models. Plus the results for the best model.

[popcornell]

@RonFrancesca can you review ?

[RonFrancesca]

@popcornell overwall the code is great! Thanks for the work!

In particular:

• I added some minor comments on the text of the README file with suggestions
• I would remove the two empty init.py files
• I would not delete the line 99 on the datasets.py file, otherwise the KeyAttribute error would be thrown.

[popcornell]

I can't see your comments on README

[RonFrancesca]

I can't see your comments on README

can you see them here?

[popcornell]

Nope I am sorry. Did you submit your review ?

[RonFrancesca]

Nope I am sorry. Did you submit your review ?

Now you should be able to see them, I guess

[popcornell]

I should ve addressed your comments

[popcornell]

Can we merge this @turpaultn @RonFrancesca