This repository is our final project on the deep learning and neural network course.
Our team members are Yanfan\Zhaobin\Liyingxue\Hongyuxin.
-
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.
-
Mixup is a generic and straightforward data augmentation principle. In essence, mixup trains a neural network on convex combinations of pairs of examples and their labels. By doing so, mixup regularizes the neural network to favor simple linear behavior in-between training examples.
-
CutMix patches are cut and pasted among training images where the ground truth labels are also mixed proportionally to the area of the patches. By making efficient use of training pixels and retaining the regularization effect of regional dropout, CutMix consistently outperforms the state-of-the-art augmentation strategies on CIFAR classification tasks.
| CIFAR-10 classification | train_acc | test_acc |
|---|---|---|
| ResNet-18 | 1.0 | 0.8915 |
| ResNet-18 + cutout | 0.9998 | 0.9020 |
| ResNet-18 + cutmix | 0.8802 | 0.9073 |
| ResNet-18 + mixup | 0.7476 | 0.9604 |
To train ResNet18 on CIFAR10 with cutout:
python train.py --dataset cifar10 --model resnet18 --cutout --length 16
To train ResNet18 on CIFAR10 with cutmix:
python train_cutmix.py --dataset cifar10 --model resnet18 --beta 1 --cutmix_prob 1
To train ResNet18 on CIFAR10 with mixup:
python train_mixup.py --dataset cifar10 --model resnet18 --mixup
To train ResNet18 on CIFAR10(baseline):
python train_cutout.py --dataset cifar10 --model resnet18