This repository provides the code for our paper, Improved Mixed-Example Data Augmentation.
Code has been tested with TensorFlow version 1.12.
First, make sure CIFAR-10 or CIFAR-100 has been downloaded and extracted to cifar10_data/cifar-10-batches-bin or cifar100_data/cifar-100-binary.
After that, basic usage is as follows:
python train.py --mixed_example_method=vh_mixup --model_dir=cifar10_models/vh_mixup --dataset=cifar10 --weight_decay=1e-4
On CIFAR-10, use a value of 1e-4 for weight decay for all mixed-example methods, and a value of 5e-4 for the baseline. On CIFAR-100, use a value of 5e-4 for all methods.
Due to the large maximum learning rate used by the default ResNet, training may be unstable for some methods. We recommend lowering the maximum learning rate (in cifar_model_fn) from 0.1 to 0.75 or, as done in the paper, running multiple copies for a few epochs, and only continuing the ones that didn't exhibit instability. Specifically, for the paper we ran 20 copies of each method for 3 epochs, continuing the 3 models with the lowest training loss.
If you use this code, please cite our paper:
Improved Mixed-Example Data Augmentation. Cecilia Summers and Michael J. Dinneen. IEEE Winter Conference on Applications of Computer Vision (WACV), 2019.
BibTeX:
@inproceedings{summers2019improved,
title={Improved Mixed-Example Data Augmentation},
author={Summers, Cecilia and Dinneen, Michael J},
booktitle={Applications of Computer Vision (WACV), 2019 IEEE Winter Conference on},
organization={IEEE}
year={2019}
}