An advanced singing voice synthesis system with high fidelity, expressiveness, controllability and flexibility based on DiffSinger: Singing Voice Synthesis via Shallow Diffusion Mechanism
Apache License 2.0
2.62k
stars
275
forks
source link
Feature proposal: freeze specific parameters when finetuning #116
This proposed feature allows freezing some parameters when finetuning from checkpoints. The frozen parameter does not need gradients and will not get updated during finetuning, thus making finetuning faster, more efficient and avoiding catastrophic forgetting.
Needs 1 new configuration key:
finetune_frozen_params: parameter key prefixes that need to be frozen during finetuning.
This proposed feature allows freezing some parameters when finetuning from checkpoints. The frozen parameter does not need gradients and will not get updated during finetuning, thus making finetuning faster, more efficient and avoiding catastrophic forgetting.
Needs 1 new configuration key:
finetune_frozen_params
: parameter key prefixes that need to be frozen during finetuning.