Faceswap Forum

Every time I try to train I get instantly hit with
08/18/2020 18:05:20 CRITICAL Error caught! Exiting...
08/18/2020 18:05:20 ERROR Caught exception in thread: '_training_0'
08/18/2020 18:05:20 ERROR Alignments file does not exist: D:\Nueral Network programs\[i]My input B photo set[/i]\alignments.fsa

No crashlog or anything. I didnt think i was working with actual alignment files when I trained. This happens on every single type of model, trained models, new models, and many combinations of training settings. Reading some past comments on here, I attempted to see if I had a image that was different sized in my input a or b, but sorting by file size in both folders reveals 256x256 at top and bottom -so I dont think so.

Other info: While updating faceswap, it tried to update or install something related to cudnn and failed and told me to do it separately, and then faceswap would crash on command prompt before the gui. Not knowing what to do, I just uninstalled faceswap and mini cuda then reinstalled them, and while this solved the command prompt crash, I cant get anything to train. Extracting works.

The .fsa file is created in the folder with the file from which the faces are extracted. Rename this file to "alignments.fsa" and put it in the folder "D: \ Nueral Network programs \ My input B".
Sorry for the punctuation, English is not my first language.

The reason for this is that "Penalized Mask Loss" is enabled by default (where it wasn't in the old world), and this requires an alignments file. You can turn this off in Model Settings.

torzdf wrote: ↑Wed Aug 19, 2020 6:27 am

The reason for this is that "Penalized Mask Loss" is enabled by default (where it wasn't in the old world), and this requires an alignments file. You can turn this off in Model Settings.

This fixed my problem. Thank you very much. I'll admit that while I tinkered with lots of settings trying to get it to work, this is probably one of the few settings I didn't mess with.
Some follow up questions: is isn't penalized mask loss a good thing that I should want on if it didnt crash my training? and also second question: will the next update have this off by default like it used to be?

This will probably remain on by default, but your settings are saved.

Penalized mask loss basically means that the Neural Network gives less importance to parts of the image that aren't face (e.g. background etc.). This has 2 effects. It can lead to faster training, and it means that latent space within the model is not being wasted with details that are not important, which should result in a better final result.

I appreciate that it may trip some people up during this transition period, but I really want the default settings to be as close to getting the "best" result as possible.