Continue training on provided pretrained models

Is it ok to continue training on the pretrained baseline models and submit it as entry?
If so, which leaderboard should it belongs? Did the provided baseline use external datasets?

1 Like

Yes, that’s fine - both baselines are currently trained on MUSDB18-HQ.

If you continue to train just with MUSDB18(-HQ) data but use, e.g., more data augmentation, then you still are eligible for leaderboard A.

If you continue to train with extra data from another dataset, then you need to set "external_dataset_used": false in aicrowd.json and your submission is for leaderboard B.


Sorry, it should have read "external_dataset_used": true in my reply.

Please also note that there is now another baseline called spleeter which was trained on external data. Hence, if you continue to train from this model, then your model needs to enter leaderboard B, independent on which dataset you use for your fine-tuning.