About the mel rvq training question

[ZZY2019-del]

Hello, may I ask how many steps are appropriate for training mel rvq, what is the final converged loss value, and should we mainly focus on commitment_loss or recon_loss?

[juhayna-zh]

Have you checked the released MelRVQ and SSL training code in https://github.com/tencent-ailab/MuQ/tree/main/src/recipes/pretrain#step2-prepare-the-mel-rvq? There are MelRVQ config and checkpoints for your reference.

[ZZY2019-del]

Yes, I observed that mel-rvq was trained for 4000 steps maybe in an hour, and I want to know if its loss has already converged. Are these steps of training sufficient for the after MUQ training?

Have you checked the released MelRVQ and SSL training code in https://github.com/tencent-ailab/MuQ/tree/main/src/recipes/pretrain#step2-prepare-the-mel-rvq? There are MelRVQ config and checkpoints for your reference.