One dataset, thirty architectures, zero tricks

What it does This repository gathers PyTorch implementations of dozens of classic image-classification networks—from VGG and ResNet to MobileNet, SENet, and NASNet—and trains them all on CIFAR-100 through a single unified script. Billed explicitly as a practice project, it offers a standard training loop, a reference custom dataset module, and a results table comparing parameter counts and top-1/top-5 error across the full zoo.

The interesting bit The author deliberately avoids training tricks, applying the same hyperparameter recipe—drawn from the Cutout paper—to every architecture regardless of temperament. That rigidity makes the repo an accidental controlled experiment: you can see which designs tolerate generic settings and which sulk, though the author admits some models likely have untapped headroom.

Key highlights

• Implements dozens of architectures including ResNet variants, DenseNet, the Inception family, ResNeXt, ShuffleNet, Xception, WideResNet, and Stochastic Depth networks.
• Fixed training regime: 200 epochs, initial lr 0.1 with step decay, Nesterov momentum, weight decay 5e-4, batch size 128.
• Results table lists parameter counts, top-1/top-5 error, and per-epoch learning rate boundaries for every model.
• Optional TensorBoard logging and a sample custom dataset implementation for reference.

Caveats

• Fixed hyperparameters likely underfit some architectures; the author notes better scores are possible with tuning.
• README typos and a pinned PyTorch 1.6 / Python 3.6 environment signal a personal snapshot rather than an actively maintained library.
• CIFAR-100 only; other datasets require writing your own loader.

Verdict Useful if you need clean, comparable baselines for legacy CNN architectures on a small-scale image dataset. Look elsewhere if you want a production training framework or state-of-the-art CIFAR-100 scores.