Show HN: ResNet-50 hits 84.35% on CIFAR-100 with heavy augmentations

Achieved 84.35% test accuracy on CIFAR-100 using a standard ResNet-50. Most public implementations top out around 81%, so this result is unusually high for a classic architecture.

Key points:

Heavy augmentations: Mixup, CutMix, ColorJitter, RandomErasing, rotations, affine transforms, and Gaussian blur.

Progressive fine-tuning: ImageNet-pretrained ResNet-50 trained in stages with OneCycleLR and mixed precision.

Streamlit demo: Upload your own images and see real-time CIFAR-100 predictions with confidence scores.

Accessible hardware: Trained on a single GTX 1650 (~15 hours), no massive cluster needed.

Repo & demo: https://github.com/Amirali-SoltaniRad/cifar100-classificatio...

Question for the community: Has anyone pushed ResNet-50 beyond 85% on CIFAR-100? What tricks worked for you?