rl_autoaug

This is the repository for our ACM MM 2021 paper "Learning Sample-Specific Policies for Sequential ImageAugmentation".

Dependency

• Ubuntu ≥ 14.04
• Python ≥ 3.6.8
• tensorflow == 1.15.0

Clone this repository

git clone https://github.com/Paul-LiPu/rl_autoaug.git

Data

The dataset can be downloaded from Google Drive. We includes the CIRAR-10, CIFAR-100, Trash Dataset used in our experiments. Those datasets can be readily used in our code after extracting the *.zip files. The extracted folders also contains *.pkl file which is the index file we split the training and validation dataset.

Download those files and extract them in 'CIFAR100_wrn28-10/data'

cd CIFAR100_wrn28-10/data
unzip cifar100.zip

Trained Model

The trained models for our experiments in the paper could be downloaded in Google Drive. The policy models are in folder "policy". And classifier is in folders "classifier".

Model training

The iterative training of classifier and policy network can be done by

cd CIFAR100_wrn28-10
python workflow.py