Training on rotated Images

Want to understand the training process better? Got tips for which model to use and when? This is the place for you


Forum rules

Read the FAQs and search the forum before posting a new topic.

This forum is for discussing tips and understanding the process involved with Training a Faceswap model.

If you have found a bug are having issues with the Training process not working, then you should post in the Training Support forum.

Please mark any answers that fixed your problems so others can find the solutions.

Locked
User avatar
janipa
Posts: 3
Joined: Thu Dec 16, 2021 12:40 pm

Training on rotated Images

Post by janipa »

Hi , I would like to ask if it is possible for the current version of faceswap to train on rotated extracted images?
I used to be able to do that on a very old version, probably it was using tensorflow 1.14,
but now on the latest version when i rotate my extracted images it wouldn't detect the rotated image and train at all.

User avatar
bryanlyon
Site Admin
Posts: 793
Joined: Fri Jul 12, 2019 12:49 am
Answers: 44
Location: San Francisco
Has thanked: 4 times
Been thanked: 218 times
Contact:

Re: Training on rotated Images

Post by bryanlyon »

You shouldn't modify the image at all. We include metadata in the image that is required to train. Any modification to the image stands the chance of losing that data. And things such as rotating will invalidate the metadata making it incorrect and unusuable.

This sounds like an X-Y problem ( https://en.wikipedia.org/wiki/XY_problem ). Instead of asking us for help with fixing a symptom of an issue, you should ask us the problem you're actually having. I'm assuming that what's hapenning is that the image has been flipped upside down by the extract process. This sometimes happens when the alignments are incorrect. You can open that image in the manual tool (See the guides for information on how to do this) and adjust the alignments until it is upright.

If that's not your problem, please post your direct concerns so we can address them.

User avatar
janipa
Posts: 3
Joined: Thu Dec 16, 2021 12:40 pm

Re: Training on rotated Images

Post by janipa »

Hi,

Thanks for the reply.
Apologies for not stating my concern in detail.
What I actually wanted to is to do augmentation of my dataset , like example of the picture i posted below.

I realized that in setting there is already a rotation setting which can set up to 25 degree.

I used to be able train with self augmentated data, until i recently brought a rtx 3070 and realized that I have to upgrade to cuda 11 and after updating the faceswap I need to retrain my model again.

So I wanted to achieve retraining the same result as my previous model.

Image

User avatar
bryanlyon
Site Admin
Posts: 793
Joined: Fri Jul 12, 2019 12:49 am
Answers: 44
Location: San Francisco
Has thanked: 4 times
Been thanked: 218 times
Contact:

Re: Training on rotated Images

Post by bryanlyon »

Yeah, just use the augmentation setting. That's it's purpose. And while the slider only goes up to 25 degrees, it'll actually work with any number manually inserted.

Locked