With the 0.7 pytorch and old course my gpu utilization was 100% at times. After upgrading to pytorch 1.0 and course v3, the gpu usage was only spikes.
@sgugger might also want to fix this for a future version of the notebooks by adding
fn_paths = [item.as_posix() for item in fn_paths]
into the https://github.com/fastai/course-v3/tree/master/nbs/dl1 notebooks that operate on file names. This would make Jeremyâs regexp work on Win and Posix without us poor Windows users having to figure out REs on themselves
The as_posix
is in the library now, it has been suggested as a bug fix and will be in v1.0.43 when itâs released.
set path = râ\([^\]+)_\d+.jpg$â will fix it.But in the meet a issueââPicklingErrorâ until now, i canât fix it.
I think that is because fastai0.7 has its own custom loader, while fastai1.0 uses Pytorchâs dataloader⌠As @mgorecki said, we should wait Pytorch fix for thisâŚ
The Pytorch issue youâve linked to is the largest issue I have using Fastai on Windows. What used to take me 20-30 seconds now takes around 2-5 minutes (minimum).