Training with the basic recipe
Hi, I am trying to replicate training of MetricGAN+ model exactly according to your recipe.
The recipe has downloaded vctk data and lexicon and then epochs started, about 50 of which passed now.
I observe that discriminator loss decreases rapidly and then oscilates around 0.05 or so, but generator loss oscilates much more without any observable trend to decrease. And other metrics (SI-SNR, pesq, stoi) oscilate accordingly.
Is this expected behaviour for first 50 epochs? Is generator loss starts to decrease more noticeable on later epochs?
Do you have your training logs to compare with?
Thanks,
Maxim
Yes, training logs are available at https://drive.google.com/drive/folders/1fcVP52gHgoMX9diNN1JxX_My5KaRNZWs?usp=sharing
Thank you! I have already found these logs and they are a bit different from those I obtain. But yes, generator loss started to decrease in my experiment too after more epochs passed