Expert Lip-Sync Discriminator is not converging on LRS2?

[Mayur28]

Hi,

I have experimenting with the code, but regardless of what I try with the expert lip-sync discriminator, the loss does not seem to go below 0.68. I am using the LRS2 (main) dataset. For some reason, even when I train the model exactly as it is, it is not showing signs of convergences. In previous issues, it was said that the model starts to converge after around 150K - 100K steps - I currently have the original model training now training for 350K steps but the loss has not gone below 0.68.

If anyone is experiencing this issue (or has not experienced this issue) when training with LRS2, please let me know, as this will also give an indication if there is an issue on my side.

Any assistance would be highly appreciated. Thanks