During two separate training sessions I've received the following error during the convert process:
C:\Users\fs\faceswap_amd\lib\convert.py:197: RuntimeWarning: invalid value encountered in rint
patched_face = np.rint(patched_face,...
This results in black boxes being shown over the swapped face. The issue is only present during convert and not evident during training since everything appears normal.
I read in some previous posts that this was likely due to a NaN somewhere in the model and that the session is corrupted with little hope of recovery. I did run the model through Nan-Scan which did not uncover any issues.
Being curious as to why the training process showed no signs of problems, I modified the training configuration to disable Mixed Precision and then reran the training for as many iterations to ensure a [Saved model] step was run at least twice.
I then reran the conversion and the rint warning/error disappeared. I was able to convert successfully and continue training with no loss in progress or data.
Hope this helps some else out there that might run into this same issue.