Page 1 of 1

What does it mean when training starts wonky?

Posted: Mon Jan 10, 2022 6:19 pm
by cagonzon

Hello! (I posted this erroneously in Training Discussion, first, apologies)

The first model I trained went pretty well..

The second model, doing the same procedure from scratch, not so much. From the start it's been showing this (image) for the A>A and A>B Image

the only difference in procedure this time compared to the first is that I have about 30k extracted images as a source for A, while for the first model I have about 3k.

This second model training is 10k iterations in and it hasn't changed. The masks are there in the extracted faces, but the images are these blanks. Also, face B is the same extracted sources I used for the first model. I wanted to do this second model from scratch.

Any thoughts? Corrupted from the start? This was from the beginning, so it's not like the model started warping and turning into a solid color (like in the FAQ) it was always wrong.

Thanks for any input.


Re: What does it mean when training starts wonky?

Posted: Mon Jan 10, 2022 10:29 pm
by cagonzon

I took it back to the start. New Face A and B, extracted, cleaned up, began training, and the same thing happens. Can't figure out what I'm doing wrong.


Re: What does it mean when training starts wonky?

Posted: Tue Jan 18, 2022 12:59 am
by abigflea

This is Phaze-A , correct?


Re: What does it mean when training starts wonky?

Posted: Wed Jan 19, 2022 3:32 am
by torzdf

Yeah, really depends on the model you're using. If it's failing to start training though, you should try lowering the learning rate.