/ Machine Learning

A simple regularization method for convolutional neural networks

A simple regularization method for convolutional neural networks

Cutout

2.56%, 15.20%, 1.30% on CIFAR10, CIFAR100, and SVHN

This repository contains the code for the paper Improved Regularization of Convolutional Neural Networks with Cutout.

Cutout is a simple regularization method for convolutional neural networks which consists of masking out random sections of input images during training. This technique simulates occluded examples and encourages the model to take more minor features into consideration when making decisions, rather than relying on the presence of a few major features.

Bibtex:

@article{devries2017cutout,  
  title={Improved Regularization of Convolutional Neural Networks with Cutout},  
  author={DeVries, Terrance and Taylor, Graham W},  
  journal={arXiv preprint arXiv:1708.04552},  
  year={2017}  
}

Results and Usage

Dependencies

PyTorch v0.4.0
tqdm

ResNet18

Test error (%, flip/translation augmentation, mean/std normalization, mean of 5 runs)

Network CIFAR-10 CIFAR-100
ResNet18 4.72 22.46
ResNet18 + cutout 3.99 21.96

To train ResNet18 on CIFAR10 with data augmentation and cutout:
python train.py --dataset cifar10 --model resnet18 --data_augmentation --cutout --length 16

To train ResNet18 on CIFAR100 with data augmentation and cutout:
python train.py --dataset cifar100 --model resnet18 --data_augmentation --cutout --length 8

WideResNet

WideResNet model implementation from https://github.com/xternalz/WideResNet-pytorch

Test error (%, flip/translation augmentation, mean/std normalization, mean of 5 runs)

Network CIFAR-10 CIFAR-100 SVHN
WideResNet 3.87 18.8 1.60
WideResNet + cutout 3.08 18.41 1.30

To train WideResNet 28-10 on CIFAR10 with data augmentation and cutout:
python train.py --dataset cifar10 --model wideresnet --data_augmentation --cutout --length 16

To train WideResNet 28-10 on CIFAR100 with data augmentation and cutout:
python train.py --dataset cifar100 --model wideresnet --data_augmentation --cutout --length 8

To train WideResNet 16-8 on SVHN with cutout:
python train.py --dataset svhn --model wideresnet --learning_rate 0.01 --epochs 160 --cutout --length 20

Shake-shake Regularization Network

Shake-shake regularization model implementation from https://github.com/xgastaldi/shake-shake

Test error (%, flip/translation augmentation, mean/std normalization, mean of 3 runs)

Network CIFAR-10 CIFAR-100
Shake-shake 2.86 15.58
Shake-shake + cutout 2.56 15.20

GitHub