search

shaoanlu / faceswap-GAN

A denoising autoencoder + adversarial losses and attention mechanisms for face swapping.
3.36k stars 844 forks source link

Just white squares after a day of TRAIN... #72

Open ZeroCool22 opened 6 years ago

ZeroCool22 commented 6 years ago

https://www.youtube.com/watch?v=9MaQYIpxnMY

Training done with: FaceSwap_GAN_v2_sz128_train.ipynb

Video: FaceSwap_GAN_v2_test_video.ipynb

Also tried FaceSwap_GAN_v2_test_video_MTCNN, same result.

petergerten commented 6 years ago

What did you see while training? Was the preview ok?

ZeroCool22 commented 6 years ago

Yes, preview windows was ok.

I'm going to bed now and i will let it training the same datasets but with FaceSwap_GAN_v2_train and i will report back later.

petergerten commented 6 years ago

I think you have to do the video generation with the routine inside the 128 notebook, as the others are for the 64px models.

ZeroCool22 commented 6 years ago

OK, FaceSwap_GAN_v2_test_video seems to "works".

https://www.youtube.com/watch?v=tFCDUoPwBJQ

There are still problems with the ALIGN of the faces in some frames.

@petergerten Can you explain me the "routine inside the 128 notebook" i don't see any test_video script with "128" in his title on it

petergerten commented 6 years ago

@ZeroCool22 the video generation for the 128 notebook is included in the same notebook, there is not separate one. See " Make video clips w/o face alignment" - no alignment available though

ZeroCool22 commented 6 years ago

@petergerten That's the video generator i used... w/o face alignment since the w/aligment said it's experimental.

petergerten commented 6 years ago

@ZeroCool22 I do have the alignment problem as well: see https://github.com/shaoanlu/faceswap-GAN/issues/67

ZeroCool22 commented 6 years ago

@petergerten Yeah it doesn't align well and also sometimes it's ZOOMED in/out, i was really hyped for this REPO, but looks like need some time to be a better/fast and easy alternative to the FakeApp.

petergerten commented 6 years ago

@ZeroCool22 same thoughts here :) I still don't quite understand if this zoomed out effect is simply an alignment problem or what exactly is happening

shaoanlu commented 6 years ago

In readme description it writes sz128 is NOT compatible with test_video notebook. In addition, if you use sz128 model in test_video notebooks, the will be an error message in “load model” step. Furthermore, there is a same issue you posted before, you can update info there... And if you are not satisfied with the mask, perceptual loss is always there for you to try.

Just, please, provide more detail about your problem like what is the output message of each step or more concrete example, what are the previws look like and what are they different from output videos, insread of throwing me a not-working video/image for every error you encountered. And please notice that I can not reproduce your 1.5 days training of 1080ti GPU (it will cost me 3 full days for tantamount training iterrations on K80).

shaoanlu commented 6 years ago

To be honest, this repo is nothing but a personal project. I did not make any cent from it and did not (and will never) ask for donation. I create this repo because I want to verify my understanding on deeplearning and perhaps have chance to interact with those who are also interested in DL (this is why I choose jupyter, which is suitable for fast prototyping). But untill now, I am feeling myself more like doing custom service for a “DL? Who cares?” product while learning very little through out the process. The “deepfakes” community cares much more about “faking” than making progress on going “deep”. These backfiring deviations really burn me out. I am afraid I can not meet the high expectation of the deepfakes community without others help.

ZeroCool22 commented 6 years ago

@shaoanlu I would have posted an error message if there was one to show, when i got that just white square there was no error, just the result you see on the video.

If there are errors to show, i ofc. will post these, like i did in my last post about the color correction issue.

shaoanlu commented 6 years ago

OK, I’ll look into the color correction issue.

ZeroCool22 commented 6 years ago

@shaoanlu Thx.

deepfakesclub commented 6 years ago

@shaoanlu You sound like a good guy, or girl but statistically unlikely, so I wanted to speak up in your support. I hope you don't mind.

If you are too nice, people will take advantage of you (not saying this thread starter, he's just excited about the tech, but in general). Don't let yourself burn out because of them. There are many authors of popular githubs that don't reply to messages anymore, probably for similar reasons. The WORST result is if it hurts your love of deep learning.

It's the internet. You spend hours working on something and share it for free. Then, people use it and complain, this sucks, why doesn't it work right, etc. Other people are trying to take your work or other work and profit from it while refusing to acknowledge or thank the original sources.

Frankly, none of these people deserve your time or your talents. Sometimes it is not bad people, but just young, socially insensitive nerds, or culturally different people. Still, no one has a right to ask anything of you.

The challenge of getting something new working is fun. Brainstorming ideas, being awed by the latest great paper, trying out crazy experiments - hell yeah. Polishing a free product for the ungrateful public... not so much.

Maybe just browse and reply here once a week or once a month. Stick with what you really love. If you stop contributing to this repo completely, you have already achieved far more for the deepfaking community than almost anyone else. THEY should be helping YOU, not the other way around!

As you also said, most are more concerned with "faking". Think about the type of people this work attracts. Some programmers I know have stated they are interested in the technology but don't want to be linked with it. That may be part of the reason why you are getting less technical discussions than you had hoped. Also, if people became seriously interested in deep learning because of deepfakes, it will still take them many months before they are ready to have a meaningful discussion.

There are definitely cool people around this project, like at the main repo, so I hope you stick around. But maybe keep to your own pace and own interests.

shaoanlu commented 6 years ago

@deepfakesclub You’re right. My wishful thinking mislead me not awaring of how sensitive a topic deepfakes is. And I have to change to a more casual mindset for opensource project. Big lessons learned. Thanks.