I had the same problem, my PC was running out of memory and crashing, resulting in some incomplete leftover temporary files. I managed to figure out the issue by looking at the stack trace, though I must admit it isn’t too comprehensive. What I would suggest is to delete the temporary folder altogether, so that you give it a chance to start over without looking at the “corrupted” files.
You’d need to cd to your fastai root directory, and cd to courses/dl1. Assuming you have data directory in there (or symlink). You’d need to cd to data/dogscats/. If you run ls
you can see that you have a tmp folder. Run rm -r tmp
.
tldr;
cd ~/fastai/courses/dl1/data/dogscats
rm -r tmp
Also if you keep on having the crashing problem check out CUDA Out of Memory Error
Matthew’s advice to decrease batch size is good. I’m using 2 GTX1080 Ti GPUs and still managed to crash… might be because I’m running it on an nvidia-docker container