An implementation of mixup: Beyond Empirical Risk Minimization by Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz.
python main.py model.mixup={true,false} model.input_only={true,false}
mixup can (mathematically) avoid mixing labels by replacing beta(alpha, alpha) with beta(alpha+1, alpha). model.input_only=true is to confirm this. (*)
- PyTorch==1.6.0
- torchvision==0.7.0
- homura
pip install -U git+https://github.com/moskoumule/homura@v2020.08
| ERM | mixup | mixup (input_only=true) |
|---|---|---|
| 0.923 | 0.932 | 0.931 |
The results suggest that the alternative mixup strategy (*) is as effective as the original.