This work is incredible! I'd love to have a GAN take something like a simple art style or type of doodle... and then create more of them... I have no idea where to start though... Your 1970s comic book GAN results are amazing!!
Very exciting! Especially the minecraft Gan! Now I've also been experimenting with the same thing that your minecraft GAN does. Take a lot of minecraft images, squash it down to 1024 images (I'm going to train on 1024 by 1024). Train it for a while. Now I've actually gotten way more different results, I've also rotated in the dimensional Z coordinate (up & down) to add a bit more randomness. I'm also training on about 6000 Images.
I love your videos, thank you for making these. I have a question - do you have a video on how to label images to create a conditional model (cond = True) with the training
Is there anyway to get some 1 on 1 help setting one of these up? Looking to train my own using m y images to create visuals for my DJ sets. Will pay for lessons
This is perfect! I finally know what I'll use all my GPUs for when I finally stop mining. I feel like I could have a lot of fun with this. Thank you for sharing how to actually use the source.
Awesome video ... Quick question though. If I train the model with my own data using the Pytorch implementation can I sell my product/services commercially ?
Is this really the same Jeff Heaton that gave us Encog and those videos about it with creased bed sheets in the background and everything? Don't get me wrong, those early videos taught me a lot but things have changed :)
It should be runnable like any other docker image. I have not tried by StyleGAN image in awhile, it is StyleGAN1, so there could be some software rot. I will consider it for a future video.
Amazing Video! Can we extract specific features for these custom data and change the output generated based on tweaking the features(Eg: Fish shape/Color etc)?
Hi @Jeff Heaton - thank you for the awesome tutorial. When I run the train command - i get the error: "nvcc fatal : Value 'sm_86' is not defined for option 'gpu-architecture'" - any advice on how to address this?
this might be a dumb question but is there a way to use the 70s scifi model with pytorch? i'm having trouble getting tensorflow to play nice with my system, but had no problem getting a pretrained pytorch BIGgan model to generate pictures and videos. thanks in advance!
Great video as always. Thanks. I wonna request you to also make a video of AttnGAN. Its a text-2-image GAN. I want to learn more about it and understand how to retrain it for something other than birds or flowers.
hello im a total noob on this subject, i feed a stylegan frames of a shortfilm i made. But the results after weeks of training on colabs with lots of accounts was just similar frames but all deformed. Could there be a way of recreating the short but in the style of the gan? Like ordering the frames and the movement to sync with the real short but all with that interpolation beetween frames? or maybe i should try another method
This worked well for me at first, but on my second try with a different dataset I'm getting an error saying "UserWarning: semaphore_tracker: There appear to be 34 leaked semaphores to clean up at shutdown len(cache))" on the initial training and then it just stops :/ my dataset is only about 350 images and I compressed them more and also tried on a different computer so I'm not sure how to fix this
I've really had no end of problems with WSL2, it is a bit of a disappointment to me. I think the fact that it has been in "beta" for years now shows the "seriousness" that Microsoft puts into GPU use under Docker in Windows. Linux has had this figured out for years.
hi Jeff ,thank you for the tutorial ! may I ask how do I correctly resume my training？ sometimes my training just crash,and I use --resume parameters path to my latest snap pkl file and the training will just restarts from 0 kimg and my training results became really weird(weird colors,weird shape)
I have 8 GB of dedicated GPU memory but it is saying that it is finding an error in allocating memory when I clearly have enough. Could you help me and provide a solution?
Maybe this is out of my league here, but how did NVIDIA make such an amazing user interface where one could drag a box over the selected area of images and it show live the new GAN results? And it was sort of in an animated flux between images: ruclips.net/video/9QuDh3W3lOY/видео.html
That is the number of kilo-images from the training set that are evaluated. The higher the number, generally, the better the result (lower FID). The size of the training set does not really impact the training time, just the quality, since the training is randomly sampling from the training data and also augmenting with random variations on the training data.
Hi, learned a lot from your videos. Is there a way to do hyperparameter tuning for this model? I was not able to find much resource for this. Can you give few tips from what you have explored while training? Thanks
Want to do this on CoLab? ruclips.net/video/L3JLzoe-dJU/видео.html Want to do this on Windows? ruclips.net/video/BCde68k6KXg/видео.html Want to see this as an article? towardsdatascience.com/generating-your-own-images-with-nvidia-stylegan2-ada-for-pytorch-on-ampere-a80fab52d6b5
Great Video! Can you go deeper in the Datasets Preparation? The PyTorch Version of StyleGAN allows to use a dataset.json for basic class labels. How I can generate this json file without typing line by line? What exactly are thees class labels? Its possible to bump the result quality with a labeled dataset? For example, I label faces with male/female or with glasses and without glasses. Thanks
Yeah, that would be very exciting to watch since I couldn't really find any real info about it over the internet (nor other people experimenting with that). I would be interested to have labeled data that might help with missing details on my target training data and see if that improves final model accuracy and prevents overfitting as well.
@@DigitalMarksman I have found a prelabled Flickr Dataset www.github.com/royorel/FFHQ-Aging-Dataset But I don't find the way to use the labels in StyleGAN2. I'm currently testing a basic pytorch script to figure out how to feed the data into StyleGAN2. If you have success feel free to share your experience, I will do it also.
how to train stylegan on grayscale images ? Sir, actually I have implemented without any chnages in channel input but it does not work. but the problem is it will works for only few number of images i.e face140.jpg and then program terminate. Error: tensorflow.python.framework.errors_impl.InvalidArgumentError: 2 root error(s) found. (0) Invalid argument: input depth must be evenly divisible by filter depth: 1 vs 3 [[{{node InceptionV3/_Run/InceptionV3/InceptionV3/import/conv/Conv2D}}]] [[InceptionV3/_Run/InceptionV3/Reshape/_5015]] (1) Invalid argument: input depth must be evenly divisible by filter depth: 1 vs 3 [[{{node InceptionV3/_Run/InceptionV3/InceptionV3/import/conv/Conv2D}}]] 0 successful operations. 0 derived errors ignored. please help me sir.
Yes, so long as the laptop has a CUDA 10/11 GPU. I have a whole series on using a ThinkPad P53 for this kind of thing. ruclips.net/video/EroJ-SzKXm4/видео.html
You could make some quick bucks if you unloaded some of that lenovo equipment. Then you just tell lenovo that somebody stole the computer. Let me know if you are interested.
![Noob To Max With DRAGON REWORK In Blox Fruits [FULL MOVIE]](http://i.ytimg.com/vi/LnBMOinoOvA/mqdefault.jpg)
![I.N "HALLUCINATION" | [Stray Kids : SKZ-PLAYER]](http://i.ytimg.com/vi/n5B5q1Hwt_U/mqdefault.jpg)
Very excited to experiment with this!
Thanks Ken!
This work is incredible! I'd love to have a GAN take something like a simple art style or type of doodle... and then create more of them... I have no idea where to start though... Your 1970s comic book GAN results are amazing!!
Thank you, yes the 70s GAN was fun! Glad it is helpful.
Very exciting! Especially the minecraft Gan! Now I've also been experimenting with the same thing that your minecraft GAN does. Take a lot of minecraft images, squash it down to 1024 images (I'm going to train on 1024 by 1024). Train it for a while. Now I've actually gotten way more different results, I've also rotated in the dimensional Z coordinate (up & down) to add a bit more randomness. I'm also training on about 6000 Images.
How did it go? I remember the ballpark estimate beink 10k pictures.
@@JohnSmith-ox3gy Very good.
If I use a pretrained model, does it require the same amount of GPU Ram for inference ?
I love your videos, thank you for making these. I have a question - do you have a video on how to label images to create a conditional model (cond = True) with the training
Thanks! I have not tried that feature of StyleGAN yet.
@@HeatonResearch I would also be interested in this, especially multi-conditional.
Is there anyway to get some 1 on 1 help setting one of these up? Looking to train my own using m y images to create visuals for my DJ sets. Will pay for lessons
This is perfect! I finally know what I'll use all my GPUs for when I finally stop mining. I feel like I could have a lot of fun with this. Thank you for sharing how to actually use the source.
Definitely be interested in a WSL2 video
Hey Jeff, thank you for making this. I can't seem to be able to find your video on what to do with the .pkl files.
Awesome video ... Quick question though. If I train the model with my own data using the Pytorch implementation can I sell my product/services commercially ?
Is this really the same Jeff Heaton that gave us Encog and those videos about it with creased bed sheets in the background and everything? Don't get me wrong, those early videos taught me a lot but things have changed :)
Hi Jeff thanks for the very detailed video. How would I go about running your image in WSL?
It should be runnable like any other docker image. I have not tried by StyleGAN image in awhile, it is StyleGAN1, so there could be some software rot. I will consider it for a future video.
@@HeatonResearch thanks Jeff, will give it a shot when I get a chance
Amazing Video! Can we extract specific features for these custom data and change the output generated based on tweaking the features(Eg: Fish shape/Color etc)?
Would also like to know this
You have WSL2 working with an Nvidia GPU? Didn't think that was possible except through an insider build...
Hi @Jeff Heaton - thank you for the awesome tutorial.
When I run the train command - i get the error: "nvcc fatal : Value 'sm_86' is not defined for option 'gpu-architecture'" - any advice on how to address this?
this might be a dumb question but is there a way to use the 70s scifi model with pytorch? i'm having trouble getting tensorflow to play nice with my system, but had no problem getting a pretrained pytorch BIGgan model to generate pictures and videos. thanks in advance!
Great video as always. Thanks. I wonna request you to also make a video of AttnGAN. Its a text-2-image GAN. I want to learn more about it and understand how to retrain it for something other than birds or flowers.
Interesting, thanks!
What cloud service do you recommend for running a GAN if I don't want to invest in expensive GPU's?
How did you compile 30k comic images and then feed them into the gan?
Hi , what would you recommend if I want to generate image from binary images using only a laptop..
hello im a total noob on this subject, i feed a stylegan frames of a shortfilm i made. But the results after weeks of training on colabs with lots of accounts was just similar frames but all deformed. Could there be a way of recreating the short but in the style of the gan? Like ordering the frames and the movement to sync with the real short but all with that interpolation beetween frames? or maybe i should try another method
This worked well for me at first, but on my second try with a different dataset I'm getting an error saying "UserWarning: semaphore_tracker: There appear to be 34 leaked semaphores to clean up at shutdown
len(cache))" on the initial training and then it just stops :/ my dataset is only about 350 images and I compressed them more and also tried on a different computer so I'm not sure how to fix this
Thanks a lot for the video. Can you explain 3000 kimgs mean 3 million images there are in your dataset?
Hi, great tutorial, thank you a lot! Could you make a tutorial about conditional StyleGAN (multiclass)?
Hi Jeff, I have emailed you.. Trying to get the CoLab working again. Getting a few errors.. Any help? :)
Would love to see your comments on WSL2
I've really had no end of problems with WSL2, it is a bit of a disappointment to me. I think the fact that it has been in "beta" for years now shows the "seriousness" that Microsoft puts into GPU use under Docker in Windows. Linux has had this figured out for years.
Have you produced a video like this using colab pro?
Not yet, but working on that now, actually.
@@HeatonResearch awesome! Been learning a ton from your channel.
Hi there, Is it possible to train and create GANs using only my Mac laptop? I am not running linux. thanks!
GANs in general, yes, but not sure what software to use. StyleGAN2, no, because it requires a CUDA GPU.
@@HeatonResearch Thanks for your reply! Hmm this is what I was afraid of. I will look into who is working on Macs. Thanks again
hi Jeff ,thank you for the tutorial !
may I ask how do I correctly resume my training？
sometimes my training just crash,and I use --resume parameters path to my latest snap pkl file
and the training will just restarts from 0 kimg
and my training results became really weird(weird colors,weird shape)
4:12 Lex? :)
Great Job ,
i am still waiting find digits in Random image Not just writed hand ones .
Hey, great tutorial! In theory, would style gan work with images that are vector style graphics and not photos?
I think if you can render the vectors as a flattened PNG with no transparency, you should be able to fit the parameters required to run the GAN
jeff knows what the ppl want!
Thanks! I try my best. And am always glad to hear what the ppl want too!
How could I change the last command to allocate more memory to my GPU?
I have 8 GB of dedicated GPU memory but it is saying that it is finding an error in allocating memory when I clearly have enough. Could you help me and provide a solution?
😍😍
Best one sir❤
So do the image dimension have to be powers of 2? i have a dataset of images that are all 100x100, but i'm worried it won't work.
Yes, powers of two... generally 256x256, 512x512, 1024x1024, 2048x2048... higher seems to crash the current stylegan.
@@HeatonResearch Could that be a problem with RAM?
How can I resume when I stopped the training
Maybe this is out of my league here, but how did NVIDIA make such an amazing user interface where one could drag a box over the selected area of images and it show live the new GAN results? And it was sort of in an animated flux between images:
ruclips.net/video/9QuDh3W3lOY/видео.html
Could you explain what kimgs are and why they impact training time?
That is the number of kilo-images from the training set that are evaluated. The higher the number, generally, the better the result (lower FID). The size of the training set does not really impact the training time, just the quality, since the training is randomly sampling from the training data and also augmenting with random variations on the training data.
@@HeatonResearch Thanks!
Hi, learned a lot from your videos. Is there a way to do hyperparameter tuning for this model? I was not able to find much resource for this. Can you give few tips from what you have explored while training?
Thanks
Want to do this on CoLab? ruclips.net/video/L3JLzoe-dJU/видео.html
Want to do this on Windows? ruclips.net/video/BCde68k6KXg/видео.html
Want to see this as an article? towardsdatascience.com/generating-your-own-images-with-nvidia-stylegan2-ada-for-pytorch-on-ampere-a80fab52d6b5
This cannot be done on Nvidia Jetson Nano I suppose?
I believe so, but not something I've tried.
Great Video!
Can you go deeper in the Datasets Preparation?
The PyTorch Version of StyleGAN allows to use a dataset.json for basic class labels.
How I can generate this json file without typing line by line?
What exactly are thees class labels? Its possible to bump the result quality with a labeled dataset?
For example, I label faces with male/female or with glasses and without glasses.
Thanks
Yeah, that would be very exciting to watch since I couldn't really find any real info about it over the internet (nor other people experimenting with that).
I would be interested to have labeled data that might help with missing details on my target training data and see if that improves final model accuracy and prevents overfitting as well.
@@DigitalMarksman I have found a prelabled Flickr Dataset www.github.com/royorel/FFHQ-Aging-Dataset
But I don't find the way to use the labels in StyleGAN2. I'm currently testing a basic pytorch script to figure out how to feed the data into StyleGAN2. If you have success feel free to share your experience, I will do it also.
Can you train a StyleGan on Mac M1?
I have not been successful with that. It has CUDA specific code.
Anyone else having problems trying to set up nvidia-docker on wsl2?
why has noone built a GUI for Stylegan yet?
Well, it costs money, but I believe this system has a GUI on GANs. runwayml.com/. But agreed, someone should create a free GUI on top. hmmmmmm :-)
how to train stylegan on grayscale images ?
Sir, actually I have implemented without any chnages in channel input but it does not work. but the problem is it will works for only few number of images i.e face140.jpg and then program terminate.
Error:
tensorflow.python.framework.errors_impl.InvalidArgumentError: 2 root error(s) found.
(0) Invalid argument: input depth must be evenly divisible by filter depth: 1 vs 3
[[{{node InceptionV3/_Run/InceptionV3/InceptionV3/import/conv/Conv2D}}]]
[[InceptionV3/_Run/InceptionV3/Reshape/_5015]]
(1) Invalid argument: input depth must be evenly divisible by filter depth: 1 vs 3
[[{{node InceptionV3/_Run/InceptionV3/InceptionV3/import/conv/Conv2D}}]]
0 successful operations.
0 derived errors ignored.
please help me sir.
Have not tried that. I THINK, so long as they are ALL grayscale it would work. Failing that I would convert them to color.
Only love - :o)
how to train stylegan on rectangular images ?
That is not trivial, it would require modification of StyleGAN. But it is something I am thinking of trying.
Can you do a video on U-GAT-IT please?
I will have to have a look at that, thanks!
The images also need to be at least 64x64 pixels
interested in a WSL2 video
Can we do it in a laptop
Yes, so long as the laptop has a CUDA 10/11 GPU. I have a whole series on using a ThinkPad P53 for this kind of thing. ruclips.net/video/EroJ-SzKXm4/видео.html
I'm interested in Windows WLS2!
So am I! There will be something coming from me soon on WSL2 and Win11.
IT DOESNT WORK, $ bash: nvidia-docker: command not found
, DOES IT WORK ON WINDOWS PLATFORM
Did I hear Intel Xenon processors 0:59 🤭🤭🤫🤔
You could make some quick bucks if you unloaded some of that lenovo equipment. Then you just tell lenovo that somebody stole the computer. Let me know if you are interested.