Aug-for-LNL-Extras

Extra bits of unsanitized code for plotting, training, etc. related to our CVPR 2021 paper "Augmentation Strategies for Learning with Noisy Labels".

Why does this exist?

I found many uncommitted scripts and notebooks while cleaning up my drive. I wanted to sanitize and publish the extra code, but unfortunately, I'm lazy... so here we are!

On top of these scripts, the repo includes an extension of the Augmentation-for-LNL repo as a submodule. The submodule introduces loss saving, which is useful for generating animated loss plots like the one seen above.

Cloning

This repository uses submodules. Make sure you clone with the --recursive flag.

git clone [email protected]:KentoNishi/Aug-for-LNL-Extras.git --recursive

Scripts and notebooks are only published for demonstration purposes only. There is no guarantee that they will run on your local machine.

Some of these scripts even contain absolute paths specific to our own machines!

Some useful commands

(for my own use)

Save per-sample loss at each epoch as a tensor (for plotting)

cd Augmentation-for-LNL
python train_cifar.py --preset c10.50sym.AugDesc-WS --machine extras