CUDA error: out of memory about non-stationary_texture_syn HOT 3 CLOSED

Hi, I open the testing function during training. Normally, the testing function will load the image with its original size. Say you have an image over 1000x1000, then the output will be 2000x2000. Therefore, it might take up a lot of GPU space. I think this could be a cause leading to your problem. You can try to comment this line: test_func(opt, webpage, epoch=str(epoch)) in train.py to avoid the testing process during training or try a bit smaller image. You can also try to modify the options about loading image to a smaller size rather than the original size (in test_function.py).

from non-stationary_texture_syn.

Thanks for the response! Turned out I need to change the volatile flag to

with torch.no_grad():
self.real_A = Variable(self.fake_B.data)
self.fake_B = self.netG.forward(self.real_A)

because the volatile flag is deprecated in Torch 0.4.0. After this change, I was able to generate 2400x2400 image using a 1080 card.

Now I wonder how much extra juice can we squeeze out of it. Do you think it's possible to generate 4K image?

from non-stationary_texture_syn.

I'm not quite sure about space usage. My experiments are mostly conducted on one Titan Xp card with 12GB GPU memories. My most training images are around 800~.

from non-stationary_texture_syn.