liuzhuang13/DenseNet

Each layer connects directly to every preceding layer within dense blocks, concatenating feature maps rather than summing them, enabling efficient feature reuse with significantly fewer parameters than ResNets. The architecture includes bottleneck variants (DenseNet-BC) with 1×1 convolutions and channel compression, plus memory-optimized implementations that reduce GPU footprint through gradient sharing and custom layer designs. Built on Torch with extensive cross-framework ports (PyTorch, TensorFlow, Caffe, MXNet, Keras), making it widely accessible for image classification and derivative tasks like semantic segmentation and object detection.

4,855 stars. No commits in the last 6 months.