weiaicunzai/pytorch-cifar100

Implements 30+ CNN architectures with unified PyTorch training pipeline on CIFAR-100, featuring configurable warmup scheduling to stabilize early training and optional TensorBoard visualization. Standardizes hyperparameter settings (0.1 initial LR with step decay at epochs 60/120/160, batch size 128, Nesterov momentum 0.9) across all models for direct architectural comparison. Provides benchmark results showing top-1 accuracy ranging from 20.66% (SEResNet-152) to 34.02% (MobileNet) under identical training conditions, enabling empirical analysis of model efficiency versus performance trade-offs.

4,755 stars. No commits in the last 6 months.