Augmix is a new a data processing technique that mixes augmented images and enforces consistent embeddings of the augmented images, which results in increased robustness and improved uncertainty calibration. This technique achieves much better results as compared to other augmentation techniques. Not only it imporoves the accuracy of the models but also contributes in improving the robustness of the models.
The official code is in PyTorch. This is a just a port from PyTorch to Tensorflow 2.0 for the same work. I used ResNet20 as an example for the model but you can use whatever model you like.
Python>=3.6.x
numpy>= 1.17
Pillow>=6.2
tensorflow==2.0
I always recommend using anaconda for managing your environments but you can use virtualenv as well or you can directly install packages. It's your choice.
- Clone the repo.
pip install requirements.txtpython main.py --batch_size=128 --epochs=100
- Make everything modular
- Custom Early Stopping and History for custom training loops
- Distributed training
@article{hendrycks2020augmix,
title={{AugMix}: A Simple Data Processing Method to Improve Robustness and Uncertainty},
author={Hendrycks, Dan and Mu, Norman and Cubuk, Ekin D. and Zoph, Barret and Gilmer, Justin and Lakshminarayanan, Balaji},
journal={Proceedings of the International Conference on Learning Representations (ICLR)},
year={2020}
}