This repo is the official implementation of the ECCV2022 paper: Efficient One Pass Self-distillation with Zipf's Label Smoothing.

Zipf's LS: Efficient One Pass Self-distillation with Zipf's Label Smoothing

Framework & Comparison

[2022.9] Pytorch Zipf's label smoothing is uploaded! CIFAR, TinyImageNet, ImageNet, INAT-21 are now supported by our training codes.

[2022.7] MegEngine Zipf's label smoothing is uploaded!

Main Results

Method	DenseNet121	DenseNet121	ResNet18	ResNet18
Arch	CIFAR100	TinyImageNet	CIFAR100	TinyImageNet
Pytorch Baseline	77.86±0.26	60.31±0.36	75.51±0.28	56.41±0.20
Pytorch Zipf's LS	79.03±0.32	62.64±0.30	77.38±0.32	59.25±0.20
Megengine Baseline	77.97±0.18	60.78±0.31	75.29±0.29	56.03±0.34
Megengine Zipf's LS	79.85±0.27	62.35±0.32	77.08±0.28	59.01±0.23

Training

train_baseline_cifar100_resnet18:

python3 train.py --ngpus 1 --dataset CIFAR100 --data_dir cifar100_data --arch CIFAR_ResNet18 --loss_lambda 0.0 --alpha 0.0 --dense

train_ZipfsLS_cifar100_resnet18:

python3 train.py --ngpus 1 --dataset CIFAR100 --data_dir cifar100_data --arch CIFAR_ResNet18 --loss_lambda 0.1 --alpha 0.1 --dense

See more examples in Makefile.

Liscense

Zipf's LS is released under the Apache 2.0 license. See LICENSE for details.