Fine-Tuning on Schubert’s music

#4
by mikedimo - opened

Has anyone run the fine-tuning example on Schubert’s works as suggested in the README of the NotaGen model’s Git repository? I’m running with the large pre-trained model and the default config training parameters (batch size: 1, learning_rate: 1e-5, num_epochs: 64), but I don’t observe any improvement as the epochs progress. Specifically, the best epoch is the 3rd (out of 64) with a minimum eval loss of 0.131. I want to perform fine-tuning on a separate dataset and I’m trying to figure out whether it’s my mistake or if they’re simply providing training parameters that aren’t representative for the data they supply.

Sign up or log in to comment