The training process saves a backup copy of the model each time the loss goes down. The loss wouldn't have caught up from a short train while reversed so wouldn't overwrite the previous best score. There are also snapshots saved separately every (by default) 25000 iterations.
Search found 790 matches
- Thu Feb 13, 2020 9:23 pm
- Forum: Training Discussion
- Topic: Training halfway then swap? Start over again?
- Replies: 4
- Views: 3769
- Thu Feb 13, 2020 7:50 pm
- Forum: Training Discussion
- Topic: Training halfway then swap? Start over again?
- Replies: 4
- Views: 3769
Re: Training halfway then swap? Start over again?
- Wed Feb 12, 2020 4:01 pm
- Forum: Installation Support
- Topic: Training doesn't start
- Replies: 3
- Views: 3927
Re: Training doesn't start
- Tue Feb 11, 2020 5:38 pm
- Forum: Artificial Intelligence
- Topic: tattoo removal? GANs?
- Replies: 6
- Views: 20255
Re: tattoo removal? GANs?
Not saying that you can't do it or that there is no value, just that it's out of scope for faceswap. Feel free to download StyleGan if you want to try face generation.
- Tue Feb 11, 2020 4:24 pm
- Forum: Artificial Intelligence
- Topic: tattoo removal? GANs?
- Replies: 6
- Views: 20255
Re: tattoo removal? GANs?
For faces, yes. https://www.thispersondoesnotexist.com/ shows a good example of it. They've also got information on how and what they're doing.
For generating new faces, GANs are great. But it's outside the domain of faceswap, which focuses on identity swapping.
- Sun Feb 09, 2020 11:35 pm
- Forum: Artificial Intelligence
- Topic: tattoo removal? GANs?
- Replies: 6
- Views: 20255
Re: tattoo removal? GANs?
Best way that I can think of to do tattoo removal in AI would be a Tattoo detection model followed by a patch redrawing (infill) network. But like Torzdf said, I'm unaware of any complete packages that do it.
- Sun Feb 09, 2020 11:34 pm
- Forum: Artificial Intelligence
- Topic: CUDA book recommendation?
- Replies: 4
- Views: 16931
Re: CUDA book recommendation?
- Sun Feb 09, 2020 10:34 pm
- Forum: Training Discussion
- Topic: [Guide] Training best practices.
- Replies: 11
- Views: 107491
Re: [Tip] Training best practices.
Fit training is replace all your data for A with the set you're actually going to swap. Changing B is unecessary but can also help avoid overtraining problems..
- Sun Feb 09, 2020 5:01 am
- Forum: Artificial Intelligence
- Topic: CUDA book recommendation?
- Replies: 4
- Views: 16931
Re: CUDA book recommendation?
You're not going to get a whole lot more out of faceswap by reading books. That said, if you're interested in AI/ML books in general, I highly recommend http://manning.com/ especially https://www.manning.com/books/deep-learning-with-python written by Chollet (The primary author of Keras).
- Thu Feb 06, 2020 4:49 pm
- Forum: Installation Support
- Topic: can't get working on new workstation
- Replies: 17
- Views: 11680
Re: can't get working on new workstation
It's best to disable the Nouveau driver and install using the .run file.
- Thu Feb 06, 2020 4:18 pm
- Forum: Installation Support
- Topic: can't get working on new workstation
- Replies: 17
- Views: 11680
Re: can't get working on new workstation
- Wed Feb 05, 2020 4:19 pm
- Forum: Installation Support
- Topic: using your own desktop
- Replies: 2
- Views: 3334
Re: using your own desktop
If you have a built in IGPU on your CPU, I highly recommend using that for your desktop environment. It'd definitely make training while using the system smoother.
- Wed Feb 05, 2020 4:15 pm
- Forum: Hardware
- Topic: nvidia vs AMD
- Replies: 6
- Views: 9988
Re: nvidia vs AMD
Nvidia is an order of magnitude faster. There really is no competition when it comes to speed. Extra ram would be nice in letting you run a larger model, but it wont speed things up.
- Wed Feb 05, 2020 4:14 pm
- Forum: Artificial Intelligence
- Topic: Where do I learn about how the nurael net is working instead of how to use the program?
- Replies: 4
- Views: 22206
Re: Where do I learn about how the nurael net is working instead of how to use the program?
Also, you could watch our Devoxx talk where we went into some surface explanation of that.
- Wed Feb 05, 2020 12:40 am
- Forum: Training Support
- Topic: lost several days of work to a bug
- Replies: 8
- Views: 3995
Re: lost several days of work to a bug
550w with a 400w GPU is pretty tight. I'd suggest underclocking the card slightly to ensure that you're not going over. Also, make sure that your GPU isn't "factory overclocked" as even that is likely to be unstable when you're doing compute tasks.
- Mon Feb 03, 2020 9:45 pm
- Forum: Training Discussion
- Topic: what type of face photo sources should i extract if i want to train vgg obstructed?
- Replies: 1
- Views: 2109
Re: what type of face photo sources should i extract if i want to train vgg obstructed?
We don't actually support directly training the mask models. Instead those just help you by providing a mask for your data.
- Fri Jan 31, 2020 8:49 pm
- Forum: Training Discussion
- Topic: Can you keep using 1 single model?
- Replies: 1
- Views: 1813
Re: Can you keep using 1 single model?
It's not recommended. You get identity bleed. That said, you should make sure that "B" is the face that's going on to the "A" original as some models actually care about that.
- Fri Jan 31, 2020 5:45 pm
- Forum: Hardware
- Topic: Has deepfaking for long periods caused anyone long-term hardware issues?
- Replies: 9
- Views: 7723
Re: Has deepfaking for long periods caused anyone long-term hardware issues?
- Wed Jan 29, 2020 8:02 pm
- Forum: Extract Support
- Topic: CRASH on extraction
- Replies: 1
- Views: 1915
Re: CRASH on extraction
Code: Select all
ImportError: `load_model` requires h5py.
Your install failed. You can either try again with the installer or if you're comfortable doing it, you can run update_deps.py in your faceswap environment.
- Wed Jan 29, 2020 2:56 am
- Forum: Extract Discussion
- Topic: viewing/modifying alignments.fsa?
- Replies: 7
- Views: 11388
Re: viewing/modifying alignments.fsa?
The GUI doesn't require a GPU and can run on any OS. So you can still use the manual tool locally. However, you can still see the alignments by using
Code: Select all
python tools.py alignments -j draw -a [alignments] -fr [frames_folder_or_video]
which will output the files with the alignments drawn on.