gpleiss/efficient_densenet_pytorch

Uses PyTorch's checkpointing feature to discard intermediate feature maps during forward passes and recompute them during backpropagation, reducing memory consumption from quadratic to linear with network depth. Supports both efficient and standard modes with configurable depth and growth rates, compatible with CIFAR/SVHN (small inputs) and ImageNet (large inputs). Includes a demo script with single and multi-GPU training support via python-fire.

1,539 stars. No commits in the last 6 months.