h264 mmco: unref short failure

[SwarmTogether]

I've made two working DeepFakes LOVE THIS!! THANK You!!

So...

I've manually aligned a set and was re-extracting (Tools>Alignments>Extract) when console logged dozens of times in red is:
[h264 @ 000002112f48eb00] mmco: unref short failure

Even though the new output set looks complete I can't train, missing frames.

It appears that my original MP4 is flawed: when aligning I saw many "duplicate" frames, as if the player saw a missing frame and just repeated the last, but the training can't deal with it at speed. I've played with the Configs to limit batch size and Allow Growth but it still doesn't find those frames during training.

I'm guessing I need to improve/fix the original MP4 or update a decoder.

Any thoughts?
THANKS, you guys are amazing.

[torzdf]

Yeah, this is a problem with the mp4 file, so I would look to fix that in the first instance, or extract the frames to individual images prior to extracting.

[SwarmTogether]

Here's something I am trying:

• Scrubbed the MP4 by placing it in an Adobe Premiere project and exporting from there with the goal of creating fresh frames. Not actually sure this worked.

• used Tools > Alignment > remove-frames and remove-images to tweak the sets

• turned off second 4K monitor that runs on the same video card to reduce card load (I think this did help)

• Train using Original, without my masks. Masking always fails but without it the session runs, so I'm running a training session without masks for learning.

To add: the videos/alignment seem to always be "one frame late", where the alignment that is auto-generated is actually for the image one frame before.

[torzdf]

I've seen this latter issue a few times.... Going to look to hunt this issue down asap. I need to work out if it is a bug in the code or an issue with certain videos

[torzdf]

Ok, there was a bug (I think this has been fairly shortlived) which incremented the frame number 1 too many.

This should be fixed in latest extract....

I will throw together a script to fix previously extracted facesets

[torzdf]

I've added a job to alignments tool to fix the "out by 1 error"

[SwarmTogether]

Excellent!

I'm 12 hours into this training session and planning on pausing at 150k cycles to check the results.

I'll try this new feature on my alignments at that time.