A kitchen-sink benchmark of classic CNNs on CIFAR-10

What it does This repo gathers reference implementations of well-known convolutional networks—VGG, ResNet, DenseNet, DLA, and others—for training on CIFAR-10 with PyTorch. A single main.py script and a tidy accuracy table let you swap architectures and compare results without hunting down separate codebases.

The interesting bit The README doubles as an informal benchmark leaderboard: every model appears to be trained through the same entry point, so the numbers are at least apples-to-apples within this repo. The best reported score is DLA at 95.47%, though the README offers almost no detail on hyperparameters, hardware, or data augmentation.

Key highlights

• 14 architectures in one place, from VGG16 (92.64%) to DLA (95.47%).
• Single main.py interface with a --resume flag for restarting runs.
• Over 6,400 stars, suggesting it is a widely-cited baseline reference.
• Author’s stated goal is simply “playing with PyTorch”—no academic pretense.

Caveats

• The README is extremely terse: no training times, augmentation details, or reproducibility notes.
• A truncated arXiv link for SimpleDLA suggests the table has not been carefully proofed.

Verdict Useful if you need a quick CIFAR-10 baseline for classic CNNs; look elsewhere if you want documented training recipes or rigorous reproducibility notes.