Experiments when run using Python 3.8.5, CUDA Version 11.3 and Pytorch 1.10.0.
No changes are expected for different versions.
To install dependencies simply run
pip install -r requirements.txt.
To check the available flags run
python main.py --help
Most importantly the flag label_noise controls the percentage of noise in the labels of the dataset and the flag augmentation control the type of augmentations to use. Valid options include none for no augmentations full for the set of augmentations used (except mixup that has its own flag) and fullfixed_n that specifies a fixed number of augmentations per sample. The number n should be replaced by the specified number of augmentations per sample.
To reproduce our random label experiments with augmentations, simply run
CUDA_VISIBLE_DEVICES=0 python main.py --dataset CIFAR10 --model resnet18_small --epochs 10000 --batch-size 256 --comment random_labels --lr 3e-4 --image-size 32 --augmentation full --label-noise 1.0 --bottleneck-dim 65536
and wait for a few days.
If you like our paper please cite as:
@article{anagnostidis2022curious,
title={The Curious Case of Benign Memorization},
author={Anagnostidis, Sotiris and Bachmann, Gregor and Noci, Lorenzo and Hofmann, Thomas},
journal={arXiv preprint arXiv:2210.14019},
year={2022}
}