Anyone with Successful Training?

[Ahmer-444]

Hi! Anyone out there is able to replicate the results with this training stack. If Yes, then with which resolution, GPU configurations, epochs, and dataset?

I'm able to run the whole process for 224x224 resolution for about 1500 epochs on a dataset of around 10000 face images from StyleGAN2. I have trained it for around 4 days on a Tesla V100 but still can't see any results so far. Should I train it more, instead 500 epochs were mentioned in the original paper "Experiment" section?

[ztf512]

I have trained this net for about 2.5 days on a single GTX 1080Ti (840 epochs), and the training dataset I used is made by make_dataset.py from CelebA (800 images and latents are selected for training). I find the training result of each epoch in name/web/images folder and I think the training process is OK.

When I test my trained net by test_one_image.py, I meet some problems--the generated image is blurred with white light noises. By reading readme.md I find there should be some modification in test_one_image.py, and it's just like this:

from util.norm import SpecificNorm # add spNorm = SpecificNorm() # add img_fake = model(img_id, img_att, latend_id, latend_id, True) → img_fake = model(img_id, spNorm(img_att), latend_id, latend_id, True) # modify

Finally the generated image becomes normal as expected.

[Ahmer-444]

Hi @ztf512 ! Thanks for your reply. It gives me a way to move forward and ensure if it can be trained for a longer period, we can get something desirable.

I have used a 10K dataset from StyleGAN generated images with resolution 1024x1024. I have resized it to 224x224 and make_dataset downsize the number to around ~5K images.

I have trained it for around 1400 epochs on Tesla V100 for about 4 days. But, the results didn't look promising as attached.

===========================================================================================

Can you please share some of your results here?

[zhangyunming]

same with you

[tiansw1]

Hi @ztf512 ! Thanks for your reply. It gives me a way to move forward and ensure if it can be trained for a longer period, we can get something desirable.

I have used a 10K dataset from StyleGAN generated images with resolution 1024x1024. I have resized it to 224x224 and make_dataset downsize the number to around ~5K images.

I have trained it for around 1400 epochs on Tesla V100 for about 4 days. But, the results didn't look promising as attached.

===========================================================================================

Can you please share some of your results here?

It finally worked on my project in video swapping. The problem is in util.videoswap.py you need to set crop_size=512 in video_swap function, you can test a video swapping and trace that to find out how to swap a single image