Learning-Deep-Learning

AutoAugment: Learning Augmentation Policies from Data

May 2019

@YuShen1116

tl;dr: Use reinforcement learning to train a agent to search data augmentation policy.

Overall impression

Instead of hard code the data augmentation(image flip, resize, rotation), they trained a small network(reinforcement learning) to decide which data augmentation should be applied on the training data. They also tried auto augmentation transferring(trained on CIFAR-10 and applied on difference dataset for instance.) and the results show that transferring could also work.

Key ideas

Technical details

Notes