search

zhvng / open-musiclm

Implementation of MusicLM, a text to music model published by Google Research, with a few modifications.
https://arxiv.org/abs/2301.11325
MIT License
522 stars 59 forks source link

train loss of semantic stage has problem #20

Closed missaaoo closed 1 year ago

missaaoo commented 1 year ago

when train the semantic stage, the loss becomes very large at more than 2000 steps，I don't know what is causing this problem， do I need to adjust my training strategy?

zhvng commented 1 year ago

maybe try a large batch size and/or lower learning rate?

missaaoo commented 1 year ago

maybe try a large batch size and/or lower learning rate?

yeah, when use lower learning rate, the loss is normal .In addition, I would like to ask, the infer speed of coarse model is very slow, how to modify the decoding part?

zhvng commented 1 year ago

inference code is here: https://github.com/zhvng/open-musiclm/blob/main/open_musiclm/open_musiclm.py#L299-L319. I agree that inference is very slow, and it would be cool to explore ways to speed it up!