Installing requirements on windows/linux takes a huge time due to nvidia stuff

[julesvanrie]

Since we included transformers, the python requirements install takes even longer than before.

Reasons (double check): transformers needs pytorch, which installs plenty of nvidia stuff too.

Couple of options:

• Remove transformers stuff: we only use it one day, students could run the notebooks on Colab (like in the previous days of DL).
• For pytorch to install a cpu-only version. Now we install nvidia cuda stuff, but most students don't have any benefit from it. (Pretty rare to have a (powerful) nvidia gpu in laptops.)

[julesvanrie]

On windows / linux 9.4 Gb of which:

• nvidia 2.8G
• tangled_up_in_unicode 1.8G
• torch 1.4G
• tensorflow 1.4G

On macos 4.3 Gb of which:

• tangled_up_in_unicode 1.7G
• torch 0.3G
• tensorflow 0.9G

Maybe worth trying a cpu only of tensorflow too?

[julesvanrie]

tangled-up-in-unicode :

Dependency of pandas-profiling and visions, which in turn is a dependency of pandas-profiling.

So we can drop this if we drop pandas-profiling.

➜ pip install "pipdeptree<2.17.0"
➜ pipdeptree -r -p tangled-up-in-unicode
tangled-up-in-unicode==0.2.0
├── pandas-profiling==3.3.0 [requires: tangled-up-in-unicode==0.2.0]
└── visions==0.7.5 [requires: tangled-up-in-unicode>=0.0.4]
    └── pandas-profiling==3.3.0 [requires: visions==0.7.5]
➜  pipdeptree -r -p visions
visions==0.7.5
└── pandas-profiling==3.3.0 [requires: visions==0.7.5]

[julesvanrie]

ydata-profiling does not need tangled-up-in-unicode directly. visions doesn't need it anymore either in the latest version. But that requires pandas>=2.0.0 and we're still on 1.4.4).