PyTorch stop using cuda after reboot in Linux Mint 21.3

A good post about data loading bottlenecks can be found here.
However, you should also try to figure out what exactly might have changed between your previous run and now as it seems the data loading became the bottleneck now. E.g. did you move the data from a local SSD to a network storage, did you increase the size of the sample, or added more preprocessing?