This repository implements key experiments on quantifying lottery tickets under label noise presented in the following paper:

Quantifying lottery tickets under label noise: accuracy, calibration, and complexity. Viplove Arora, Daniele Irto, Sebastian Goldt, Guido Sanguinetti (UAI 2023).

MNIST	CIFAR-10

This framework is a fork of the OpenLTH project. We adopted most of the original implementations, and made additional modications.

• Python >= 3.6
• PyTorch >= 1.4
• TorchVision >= 0.5.0

• Install the requirements.
• Modify platforms/local.py so that it contains the paths where you want datasets and results to be stored. By default, they will be stored in ~/open_lth_data/ and ~/open_lth_datasets/.

python open_lth.py lottery --default_hparams=mnist_lenet_300 --training_steps=120ep --levels=17 --pruning_fraction=0.2 --random_labels_fraction 0.2

You can change the number of neurons in the hidden layer by changing the default_hyperparameter option and observe the double descent curve. The number of levels of pruning need to be changed accordingly to make sure that the entire sparse double descent curve is observed.

python open_lth.py lottery --default_hparams=cifar_resent_20_80 --training_steps=160ep --levels=30 --pruning_fraction=0.2 --random_labels_fraction 0.2 --rewinding_steps=10ep

We develop our project mainly on the OpenLTH, which is easy to use and extend. Thanks for the great work!

If you find this useful for your research, please cite the following paper.

@inproceedings{arora2023quantifying,
  title={Quantifying lottery tickets under label noise: accuracy, calibration, and complexity},
  author={Arora, Viplove and Irto, Daniele and Goldt, Sebastian and Sanguinetti, Guido},
  booktitle={The 39th Conference on Uncertainty in Artificial Intelligence},
  year={2023},
  url={https://openreview.net/forum?id=YVWw8OSEse}
}

If you have any question on these codes, please don't hesitate to contact me by [email protected]