I’m trying to run the course notebooks on a GPU cluster I have access to.
However, nothing works. So far, I’ve only run into import errors. I cannot run anything.
Initially I was able to install fastai and all its dependencies (pytorch etc.). When I try to run the first notebook of the course, I get a “GLIBC_2.14 not found” error. Google gave me nothing. After some prodding, I figured this is a torch issue, so I downgraded to PyTorch 0.3.0. Now I’m able to run “import torch” without errors.
So I run the “from fastai import *” command and I get some “cannot import name ‘as_tensor’” error. I guess this is because fastai uses a function from a latter version of pytorch. Back to square one. Now I could go in and add this function to my installed pytorch folder but this obviously won’t be the only ‘new’ function fastai uses. It would be impossible to keep adding these till I stop getting errors.
Now I could downgrade fastai to version 0.7 but the notebook mentions it uses 1.0. So same problem as above, this time with fastai library instead of pytorch.
Obviously updating GLIBC is not an option. Apparently I do have GLIBC_2.14 but from what I can see pytorch requires a higher version. Anyway, I can’t do anything about it as I don’t have sudo access.
tl;dr: Stuck between independently working versions of course notebooks, fastai, and pytorch but no way to get them all to work together.
What are my options? I could use paperspace but it breaks every time I try to upload a large-ish file (~5GB). That’s not an option.