AugMix-pytorch

Pytorch Implementation of AugMix (ICLR2020)

Requirements

• Python 3.6 >
• PyTorch 1.4 >
• CIFAR-100, CIFAR-100-C

Details

• Baseline : 100 epochs, 30 mins
• AugMix : 100 epochs, 100 mins
• AugmentAndMix : 100 epochs, 50.1 mins
• AugMix with 1 augmented image : 100 epochs, 49.7 mins
• AugMix with 3 augmented image : 100 epochs, 142.8 mins

Results (Error Rate)

• Baseline : 25.72% (CIFAR-100 Test Set) / 55.02% (CIFAR-100-C Average)
• AugMix : 23.16% (CIFAR-100 Test Set) / 35.97% (CIFAR-100-C Average)
• AugmentAndMix (No JSD) : 24.62% (CIFAR-100 Test Set) / 39.65% (CIFAR-100-C Average)
• AugMix with 1 augmented image : 24.49% (CIFAR-100 Test Set) / 39.54% (CIFAR-100-C Average)
• AugMix with 3 augmented image : 23.23% (CIFAR-100 Test Set) / 36.04% (CIFAR-100-C Average)

Loss Curves

Sample Images (original / augmented 1 / augmented 2)

left : Original Image / middle and right : augmented samples

References

[1] Hendrycks, D., Mu, N., Cubuk, E. D., Zoph, B., Gilmer, J., & Lakshminarayanan, B. (2019). Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781.

[2] augmix, google-research, GitHub Repository, 2020, https://github.com/google-research/augmix