This repository provides a Keras implementation of CapsNet, a neural network architecture proposed by Sara Sabour, Nicholas Frosst, and Geoffrey Hinton. The model uses dynamic routing between capsules to preserve hierarchical pose relationships, which the authors argue better captures spatial hierarchies than traditional CNNs. The implementation supports TensorFlow 1.2 and 2.x, trains on MNIST dataset, and achieves competitive test error rates around 0.30-0.34%.