search

jishengpeng / WavTokenizer

SOTA discrete acoustic codec models with 40 tokens per second for audio language modeling
MIT License
801 stars 44 forks source link

Using EMA on the generator markedly improves the validation loss #33

Open erogol opened 2 months ago

erogol commented 2 months ago

Since there was to much fluctuation in the validation loss, I tried EMA with the generator parameters and it improved the validation loss and subjective quality a lot

image
erogol commented 2 months ago

there are bing jumps when commitment loss reduces suddenly due to probably reassignment of the codebooks

jishengpeng commented 2 months ago

Since there was to much fluctuation in the validation loss, I tried EMA with the generator parameters and it improved the validation loss and subjective quality a lot

image

Thanks for the reminder, we will try EMA in WavTokenizer2.