New faker general questions

Discussions about research, Faceswapping and things that don't fit in the other categories here.


Post Reply
User avatar
JTcgs
Posts: 4
Joined: Sat Feb 03, 2024 4:05 pm
Has thanked: 1 time

New faker general questions

Post by JTcgs »

So, just started for the first time a few days ago, read a ton of guides and posts, did some searches and answered just about all my early questions and have hit a snag on a topic.

Saving/backups

I am wanting to do several things. Make a full backup on a storage drive that I can quickly grab if I decide to work on my current models later and I also want to test out training without warp for a few hours just to see how what I currently have comes out knowing I will almost certainly will need to go back to training with warp and will need to restore a backup pre no-warp. I am also wondering about coming back to the current model after I started doing another at a later time.

So, my questions I could not find answers to.

  1. I know the model training folder is the training data set which is a must to backup and use to restore, what about the snapshots?
  2. If I decide to come back to this model after doing another model, does the face photos used to train have to be placed into the exact folder tree structure?

Example. If the current model A is located in C:faceswap/work/model A, and then subfolders, does everything have to be placed exactly into that tree or does that not matter as long as I point the program to the new folders?

User avatar
torzdf
Posts: 2621
Joined: Fri Jul 12, 2019 12:53 am
Answers: 155
Has thanked: 128 times
Been thanked: 610 times

Re: New faker general questions

Post by torzdf »

JTcgs wrote: Sun Feb 04, 2024 2:37 am

So, just started for the first time a few days ago, read a ton of guides and posts, did some searches and answered just about all my early questions and have hit a snag on a topic.

Welcome to Faceswap and the forum :)

I know the model training folder is the training data set which is a must to backup and use to restore, what about the snapshots?

The snapshot folder is a copy of the training folder at each 25,000 iterations (or at whatever value you set for the number of iterations between snapshots). It is created purely so that you can rollback to that state if needs be. If you are not going to ever need to rollback to a certain snapshot, it can safely be deleted as it will be unused.

It is just a copy of the model folder at that point, so to rollback to a certain snapshot, it is just a question of deleting your model folder and renaming the snapshot folder to your model folder's name (that is, remove the snapshot_<iteratations> information from the end of the folder name.

It exists as auto-backups become less and less frequent as training progresses, so is an easy mechanism for rolling back the training to a specific point in time. I tend to delete early snapshots the longer I train (or at least, start removing interim snapshots... for example I may delete every 25k, 50k, 75k snapshots and just keep the 100k, 200k, 300k snapshots). If I am finished with a model I will delete all the snapshots.

If I decide to come back to this model after doing another model, does the face photos used to train have to be placed into the exact folder tree structure?

No. You can put the faces anywhere, and you could even use entirely new faces when you go back to your model. As long as you point the Input A and Input B to where your faces are stored, these can be placed wherever you like.

The model folder can also be moved, but you should not change anything (that faceswap has created) within the model folder.

My word is final

User avatar
JTcgs
Posts: 4
Joined: Sat Feb 03, 2024 4:05 pm
Has thanked: 1 time

Re: New faker general questions

Post by JTcgs »

torzdf wrote: Sun Feb 04, 2024 1:20 pm

The snapshot folder is a copy of the training folder at each 25,000 iterations It is just a copy of the model folder at that point, so to rollback to a certain snapshot, it is just a question of deleting your model folder and renaming the snapshot folder to your model folder's name (that is, remove the snapshot_<iteratations> information from the end of the folder name.

It exists as auto-backups become less and less frequent as training progresses, so is an easy mechanism for rolling back the training to a specific point in time. I tend to delete early snapshots the longer I train (or at least, start removing interim snapshots... for example I may delete every 25k, 50k, 75k snapshots and just keep the 100k, 200k, 300k snapshots). If I am finished with a model I will delete all the snapshots.

No. You can put the faces anywhere, and you could even use entirely new faces when you go back to your model. As long as you point the Input A and Input B to where your faces are stored, these can be placed wherever you like.

The model folder can also be moved, but you should not change anything (that faceswap has created) within the model folder.

Thanks for the reply.

If I am reading this correct, the snapshot is the training model and if I am restoring a model I only need the snapshot folder and no longer need the source photos? Meaning, if I do enough iterations that my models are matching in a quality I like and I finish making the videos I want, I can back up the snapshot folder and come back to it after doing other models via the snapshot folder only?

User avatar
torzdf
Posts: 2621
Joined: Fri Jul 12, 2019 12:53 am
Answers: 155
Has thanked: 128 times
Been thanked: 610 times

Re: New faker general questions

Post by torzdf »

You can do that, yes. However, if you are happy with the final model, then there is no need to keep the snapshots (unless you wish to specifically rollback to a particular point in the model's history)

re: images, that is correct. You can feed the model any faces you wish. No need to keep the original faces around if you don't think you'll need them again.

Personally, I never keep facesets anyway, as they take a lot of space. I always just keep the original video + clean alignments file. This way I can just keep 2 files, and can always re-extract the faces from source video (using these 2 files) if I need to in the future

My word is final

User avatar
JTcgs
Posts: 4
Joined: Sat Feb 03, 2024 4:05 pm
Has thanked: 1 time

Re: New faker general questions

Post by JTcgs »

Thanks again and I actually just figured out the thing I was missing...

The Realface.h5 file is the trained facesets file.

I wasnt sure if the program required the original face files it keep drawing from or not. This great info, thanks for helping point me in the right direction!

User avatar
JTcgs
Posts: 4
Joined: Sat Feb 03, 2024 4:05 pm
Has thanked: 1 time

Re: New faker general questions

Post by JTcgs »

torzdf wrote: Sun Feb 04, 2024 2:00 pm

Y

If I may ask an unrelated question so I dont have to create another thread.

I have a second Desktop, its older but the proc is good (its an i7 3.7ghz, 16GB DDR5), I basically only use it to upconvert videos and some 3D graphic design. Its running Windows 7 and I was thinking I may upgrade the GPU to a 4060 to use to run this when I need my main PC for other things.

I tried to find the apps system requirements and cant find it anywhere. Will this run on Windows 7?

User avatar
torzdf
Posts: 2621
Joined: Fri Jul 12, 2019 12:53 am
Answers: 155
Has thanked: 128 times
Been thanked: 610 times

Re: New faker general questions

Post by torzdf »

Honestly, I have no idea. Windows 7 is well out of support now, so it may or may not work. Not been tested.

My word is final

Post Reply