The fine-tune effect is not good. What is the optimal parameter setting?
#2
by a874417699 - opened
I'm using chinese-hubert-bert for audio content classification.
I used some regular configuration, splicing the classification header and only fine-tuning the classification header, and my audio max length was 30s, sampling rate was 16k, and learning rate was 5e-4.
But my fine tuning is not good, often divided into 1 categories, I don't know why.
Are there any training papers here, or any suggestions?