DenseNet

Why we included this project

This is the authors' own implementation of DenseNet, the architecture that won Best Paper at CVPR 2017. It defines the dense block design where each layer takes in the feature maps of all preceding layers, which lets you build deeper networks without exploding the parameter count. The repo contains the original Torch code plus a memory-efficient version documented in a separate technical report, and you can follow the exact training and evaluation setup the team used on CIFAR and ImageNet. Because DenseNet's connectivity pattern still shows up in modern vision backbones, having the reference source helps when you are reimplementing, comparing, or porting the architecture to another framework. If you want to reproduce the paper's numbers or understand how the dense connectivity works, this is more direct than reading a summary.