search

jcjohnson / neural-style

Torch implementation of neural style algorithm
MIT License
18.3k stars 2.71k forks source link

GPU not fully utilized, Pytorch doesnt reserve all memory #488

Closed Pololinger closed 3 years ago

Pololinger commented 3 years ago

Hello there,

I'm new to Pytorch and noticed that I keep running out of memory. Now I have 8 GB total capacity but Pytorch is only reserving about 6 GB. According to torch.cuda.is_available() 8GB should also be available. But the problem persists even though I run torch.cuda.empty_cache()

Could this be solved by setting .cuda(non_blocking = True) ? Or changing the number of workers for Dataloader ? I was trying to do those changes but couldn't really figure out where. Help is very much appreciated.

htoyryla commented 3 years ago

This neural-style does not use pytorch at all, but (the original) torch using Lua language, not python. Pytorch did not exist back in 2015.

Anyway, you could check with nvidia-smi what other processes are using cuda and how much memory they use.