Faceswap Forum

The scenario is somewhat like this:
I start training the model. Some time after on the preview I see an error in the mask I want to fix. While fixing this error, I see some more errors in the mask and landmarks. Nothing too damaging, I presume (the resulting model looks ok, think), but still.

The question is - if I fix those mistakes in masks/landmarks and resume training, will the resulting model be better?

I presumed "yes".
But now I'm thinking - the model learns with the mistakes and probably learns to make mistakes to compensate those mistakes... and it will not forget those mistakes, so what if fixing those mistakes in the data set would cause the model to be... I don't know, less consistent?

And a follow up question.
Do you think it's an effective way to improve the model after fixing those mistakes - to start a new model with loading weights from the previous model that incorporated the mistakes?
I didn't try loading weights with the same faces, but it feels like there's a lot of data that is discarded.
And my model is already "finished" in the sense that I would already use it if I had the data set for another model ready. And since I'm still working on the next data set, I keep training this model. Wouldn't hurt. I don't see the improvements, but the numbers change...

P.S. Obviously I re-export and overwrite the images after making the changes.

Fed wrote: ↑Thu Apr 20, 2023 10:34 am

The question is - if I fix those mistakes in masks/landmarks and resume training, will the resulting model be better?

The answer is a 'theoretical yes' but in reality 'no'. Ultimately machine learning is an averaging game. The better the data you feed into it, the better the final output. However, a few outliers will have a negligible effect (if any) on the final result, as they will end up getting averaged out of the final result.

Fed wrote: ↑Thu Apr 20, 2023 10:34 am

Do you think it's an effective way to improve the model after fixing those mistakes - to start a new model with loading weights from the previous model that incorporated the mistakes?

See above. Put it this way, I spend a LOT more time fixing data for conversion than I do for training, as these will have a far larger impact on the final result.