akanimax / BMSG-GAN

[MSG-GAN] Any body can GAN! Highly stable and robust architecture. Requires little to no hyperparameter tuning. Pytorch Implementation
MIT License
629 stars 105 forks source link

How much GPU memory to use for 1024*1024? #17

Closed wodsoe closed 2 years ago

wodsoe commented 5 years ago

How much GPU memory to use for 1024*1024?

robbiebarrat commented 5 years ago

bumping this- i have the same question

robbiebarrat commented 5 years ago

@wodsoe i just tried it out; not sure exactly but more than 12gb it seems. it's stated that they used v100s (16gb gpus) - so the full 1024x1024 model is somewhere between 12 and 16 gb

wodsoe commented 5 years ago

@robbiebarrat Thank you!So it doesn't work on 1080ti or 2080ti.

akanimax commented 5 years ago

@wodsoe, @robbiebarrat. That is a little unlikely. Do you mean that on your GPU, even a batch size of 1 doesn't fit?

powerspowers commented 5 years ago

I think that this code would have to be updated to use mixed precision floating point for a 1080 or 2080 to be able to handle 1024x1024. I started to modify their newer BBMSGAN to try to add nVidia's code for fp16. My coding skills are definitely getting stretched. Going mixed precision would effectively double what GPU memory could handle.