autoaugment.md

Augmentation

AutoAugment

Model	Size	Augmentation	Epoch	Top 1	Top 5	Time	Code	Log
ResNet_vd-50	160	AutoAugment	200	78.99	94.33	232.4	98	1 2
ResNet_vd-50	160	RandAugment	200	78.89	94.08	218.1	99	1 2
ResNet_vd-50	160	RA-AA	200	79.01	94.35	249.1	100	1 2
ResNet_vd-50	160	TrivalAugment	200	79.26	94.17	213.1	109	1 2 3

• Reference papers:
  • AutoAugment: Learning Augmentation Policies from Data
  • RandAugment: Practical automated data augmentation with a reduced search space
  • TrivialAugment: Tuning-free Yet State-of-the-Art Data Augmentation
• RA-AA randomly switches between AutoAugment and RandAugment for each sample.
• Note that TrivalAugment may not work well when using stronger models or combined with other data augmentation and regularization. See Assemble for more details.