Robust Vision Models

Out-of-Distribution (OOD) Robustness with AugMix

This project explored out-of-distribution (OOD) robustness in image classification using the AugMix data augmentation technique. AugMix generates diverse augmented views of input images and enforces consistency across them, which significantly improves robustness and uncertainty calibration without extensive tuning.

My study implemented ResNet18 and ConvNeXt-tiny architectures (both pretrained and non-pretrained) on CIFAR-10, CIFAR-10-C, and CIFAR-10-P datasets. I benchmarked models using different optimizers (AdamW, SGD) and learning rate schedulers (CosineAnnealingLR, LambdaLR), and tuned hyperparameters for ConvNeXt-tiny.

Key Highlights

  • Implemented ResNet18 & ConvNeXt-tiny in the AugMix framework.
  • Benchmarked across CIFAR-10, CIFAR-10-C, and CIFAR-10-P.
  • Compared optimizers & schedulers: AdamW + CosineAnnealingLR vs. SGD + LambdaLR.
  • Showed that AdamW + CosineAnnealingLR outperformed SGD for corruption robustness, while ConvNeXt-tiny achieved the best clean accuracy with larger batch sizes.
  • Demonstrated the trade-off between corruption robustness and flip probability.

Results

Performance comparison of ResNet18 and ConvNeXt-tiny models with different optimizers and schedulers on CIFAR-10 and CIFAR-10-C.

Resources