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.
Normandy
Posts: 3 Joined: Mon Jan 27, 2020 11:31 am
Post
by Normandy » Mon Mar 23, 2020 12:10 pm
Hello, I have trained on a dataset extracted from 2 videos.
The faceA folder contains 1690 images and the faceB folder contains 2590 images.
I have trained for 40,000 iterations but the results are very bad.
Could it be because I have trained for too few iterations?
Could it be for the image number of my dataset?
Thank you a lot.
torzdf
Posts: 2649 Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 128 times
Been thanked: 623 times
Post
by torzdf » Mon Mar 23, 2020 12:11 pm
Which model?
Either way, 40,000 is not that high.
Normandy
Posts: 3 Joined: Mon Jan 27, 2020 11:31 am
Post
by Normandy » Mon Mar 23, 2020 9:50 pm
Is the model based on the input size using the CLI?
because I can't use the GUI and I don't select a model type.
however the images are all 256x256.
The network can train for 1000000 iterations but i don't know when it is enough.
Thank you a lot.
torzdf
Posts: 2649 Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 128 times
Been thanked: 623 times
Post
by torzdf » Tue Mar 24, 2020 10:28 am
If you didn't specify a model then you will be using the Original model.
yueshitian
Posts: 3 Joined: Tue Apr 07, 2020 3:56 am
Post
by yueshitian » Mon Apr 13, 2020 3:25 am
Did you solve your problem and get a good result ? I get same problem, I run 15000 iterations and my training loss is still 0.03 , It is hard for me to decrease the loss , and convert face is blurry.
torzdf
Posts: 2649 Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 128 times
Been thanked: 623 times
Post
by torzdf » Mon Apr 13, 2020 11:32 am
Loss is diminishing returns. The closer you get to convergence, the slower the loss level will drop.
See the section on monitoring training for more information:
viewtopic.php?f=6&t=146#monitor