PyTorch Implementation of Mixup

Usage

$ python main.py --block_type basic --depth 110 --use_mixup --mixup_alpha 1 --outdir results

Results on CIFAR-10

Model	Test Error (1 run)
ResNet-preact-56 w/ Mixup alpha = 0.5 (160 epochs)	5.55
ResNet-preact-56 w/ Mixup alpha = 1 (160 epochs)	5.62 (median of 3 runs)
ResNet-preact-56 w/ Mixup alpha = 2 (160 epochs)	6.14
ResNet-preact-56 w/ Mixup alpha = 1 (300 epochs)	5.11 (median of 5 runs)
ResNet-preact-110 w/ Mixup alpha = 1 (300 epochs)	4.26

Model	Test Error (median of 5 runs)	Training Time
ResNet-preact-56 w/o Mixup (160 epochs)	5.85	98 min
ResNet-preact-56 w/ Mixup (300 epochs)	5.11	191 min

w/o Mixup

$ python -u main.py --depth 56 --block_type basic --base_lr 0.2 --epochs 160 --milestones '[80, 120]' --seed 7 --outdir results/wo_mixup/00

w/ Mixup

$ python -u main.py --depth 56 --block_type basic --base_lr 0.2 --use_mixup --mixup_alpha 1 --epochs 300 --milestones '[150, 225]' --seed 7 --outdir results/w_mixup/00

References

• Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, David Lopez-Paz. "mixup: Beyond Empirical Risk Minimization." arXiv preprint arXiv:1710.09412. arXiv:1710.09412