weiaicunzai/pytorch-cifar100

This repository is a collection of image classification models implemented in PyTorch and trained on a dataset called CIFAR-100. CIFAR-100 is a standard benchmark in machine learning research: it contains 60,000 small color images spread across 100 different categories, such as apples, bicycles, and various animals. Researchers use it to test how well different neural network designs learn to tell categories apart. The project implements more than 40 different network architectures in one place, covering designs that span roughly a decade of computer vision research. These include well-known families such as VGG, ResNet, DenseNet, GoogleNet, MobileNet, and many others, each representing a different approach to building layers in a neural network. Having all of them in one repository lets a researcher or student train and compare them side by side without hunting for separate implementations. Using the project is straightforward: you pick a network by name on the command line and run the training script. Training uses a GPU if one is available. The author trained each network with the same set of hyperparameters (learning rate schedule, batch size, number of training rounds) to keep comparisons fair, and the README includes a results table showing the error rates each model achieved. More complex models generally made fewer errors, with the best results coming from SENet and DenseNet variants. Optional TensorBoard integration lets you watch training progress in a browser as the model learns. After training, a separate test script evaluates a saved model on the test portion of the dataset. The repository is intended as a learning and practice resource rather than a production system. The author notes they deliberately avoided advanced training tricks to keep the code clear, and points to a separate repository if readers want to learn about those techniques. No license is stated in the README.