I train base_medium on a custom dataset, but I get noisy outputs even after a lot of epochs (although the noise spectrogram is not the same as the uniform one of the beginning of training).
I wanted to thus ask:
Should the inputs be preprocessed in a certain way?
How many iterations (number of samples not number epochs) are expected to have visible improvement?
Did you encountered collapse scenarios while training and if yes how do they look like and how did you tackle them. For example in my case just decreasing the learning rate still yielded bad results.
Thanks for open-sourcing and supporting this work!
YS
Hi,
I train
base_mediumon a custom dataset, but I get noisy outputs even after a lot of epochs (although the noise spectrogram is not the same as the uniform one of the beginning of training).I wanted to thus ask:
Thanks for open-sourcing and supporting this work! YS