dropblock

Why we included this project

Plain dropout often does little for convolutional layers, because nearby activations are spatially correlated and information keeps flowing around the dropped units. DropBlock attacks that directly: instead of zeroing individual activations, it removes whole contiguous blocks of a feature map, which is what actually forces the network to learn features that do not depend on their neighbors. This package is a faithful PyTorch implementation of the 2018 NeurIPS paper, with DropBlock2D and DropBlock3D modules plus a LinearScheduler that gradually raises the drop probability during training, a detail the paper credits for better accuracy and less sensitivity to hyperparameter choices. It installs from PyPI and fits into an existing training loop in a few lines, and the bundled ResNet-CIFAR10 example shows a realistic setup. If you are reproducing the paper's results or just want a stronger regularizer for your own vision models, this is a clean reference to build on.