In this repository, I investigate the impact of various data augmentation techniques — MixUp, FMix, and AGMix (introduced in the MiAMix paper: https://arxiv.org/abs/2308.02804) — on improving the generalization capability of deep neural networks. All experiments were conducted on the CIFAR-10 dataset.
The model used for evaluation is ResNet-18, trained with 5-fold cross-validation for 20 epochs per fold. The training pipeline includes the 1cycle policy and AdamW optimizer. A custom implementation of LRFinder identified a peak learning rate of 1.5e-3, which was used during training.
-
Vanilla: 91.17% (91.35%, 91.37%, 90.75%, 91.07%, 91.30%)
-
MixUp: 91.89% (91.96%, 91.95%, 91.92%, 91.96%, 91.67%)
-
FMix: 92.18% (92.45%, 92.21%, 91.92%, 92.27%, 92.07%)
-
AGMix: 91.88% (91.72%, 91.82%, 91.86%, 92.10%, 91.90%)
The results clearly demonstrate that data augmentation techniques enhance the generalization of deep neural networks on CIFAR-10 compared to standard training. While all evaluated methods provided improvements, FMix achieved the highest test accuracy, showing the most consistent gains across all folds. This suggests that feature-level mixing strategies, such as FMix, can be particularly effective for improving robustness and generalization in image classification tasks.
Created by Denys Bondarchuk. Feel free to reach out or contribute to the project!