Awesome
Mixup Without Hesitation
This repo contains demo reimplementations of the CIFAR-100 training code in PyTorch based on the following paper:
Hao Yu, Huanyu Wang, Jianxin Wu. Mixup Without Hesitation.
How to use
git clone https://github.com/yuhao318/mwh.git
cd mwh
CUDA_VISIBLE_DEVICES=0 python easy_mwh.py --sess mwh_shufflenetv2_0.5 --alpha 0.5
Results
The following table shows accuracy with $\alpha = 0.5$ and 100 epochs in CIFAR-100:
| Model | PreAct ResNet-18 | DenseNet-161 | Wide ResNet28-10 |
|---|---|---|---|
| mixup | 75.27% | 78.71% | 78.86% |
| mwh | 77.00% | 79.94% | 79.94% |