Tutorials: replace cell magic with shell command

[adamjstewart]

I'm hoping this is what's causing https://github.com/microsoft/torchgeo/runs/8172122269?check_suite_focus=true to fail. Also see treebeardtech/nbmake#80.

When I run the previous code in Google Colab, I see:

UsageError: Line magic function `%nvidia-smi` not found.

From my understanding, both % and ! will let you run shell commands, but % is a cell magic command for which there are only a finite number of mappings (pip and apt are allowed, but not nvidia-smi). The other difference is that ! runs in a subprocess, while % runs in the current process. This is required for something like pip install, but not for something like nvidia-smi.

[adamjstewart]

Second commit is something I missed in #714

[adamjstewart]

Bleh, this fixes Jupyter but breaks nbsphinx. See spatialaudio/nbsphinx#670.

[adamjstewart]

I knew this seemed familiar. I actually broke this (! -> %) back in #226, our first release. This PR attempts to revert that change.

[adamjstewart]

This fixes some of the issues. I'm still seeing unexplainable errors in https://github.com/microsoft/torchgeo/runs/8172778528?check_suite_focus=true. Basically, torch.cuda.is_available() is returning True, but CUDA doesn't work. Our transforms tutorial explicitly passes things to the GPU if it's available, and our trainers tutorial uses Lightning which does the same. Not sure why torch.cuda.is_available() is lying, and we can't easily reproduce this locally.

[adamjstewart]

This looks promising! github/roadmap#505

[calebrob6]

This fixes some of the issues. I'm still seeing unexplainable errors in https://github.com/microsoft/torchgeo/runs/8172778528?check_suite_focus=true. Basically, torch.cuda.is_available() is returning True, but CUDA doesn't work. Our transforms tutorial explicitly passes things to the GPU if it's available, and our trainers tutorial uses Lightning which does the same. Not sure why torch.cuda.is_available() is lying, and we can't easily reproduce this locally.

It seems that you have that line backwards device = "cpu" if torch.cuda.is_available() else "cuda", should be "cuda" if torch.cuda.is_available() ...