/Auto-Augment

An unofficial implementation of Google Brain's research in 2018

Primary LanguagePythonMIT LicenseMIT

UAlberta Multimedia Master Program - Reinforcement Learning for Auto Data Augmentation

An unofficial implementation of Google Brain's research in 2018 using tensorflow 1.15.0. Instead of using PPO, we use basic REINFORCE policy gradient algorithm with involving creative idea : depressed feedback.

Requirement

  • numpy
  • tensorflow 1.15.0
  • keras
  • PIL
  • matplotlib

Code files

  • child_net.py: containing python class representing child network. Can be replaced by any classifer as long as it is a keras model.
  • controller.py: containing python class representing the RNN controller. Implemented in tensorflow 1.x.
  • data_iterator.py: containing python class to load cifar10 dataset. If policy is given, it will automatically apply image operations.
  • run.py: code to run.
  • transformations.py: contains functions of image transformations. 16 in total.

How to run

python3 run.py