capsules

December 23, 2019 ยท View on GitHub

Implementation of capsule network with dynamic routing that actually works.

New features:

  1. modification of the loss for anomaly detection based on https://arxiv.org/pdf/1909.02755.pdf
  2. anomaly score based on https://arxiv.org/pdf/1909.02755.pdf
  3. normality scores based on https://arxiv.org/pdf/1907.06312.pdf
  4. HitOrMiss capsules and Centripetal Loss from https://arxiv.org/pdf/1806.06519.pdf
  5. Kernelized Capsule Networks (https://arxiv.org/abs/1906.03164) example with GPytorch - partially done, reconstruction doesn't work yet

TODO:

  1. add CIFAR10 example
  2. train CatsVsDogs to above 80% validation accuracy
  3. add Autoencoder example
  4. add unsupervised capsule layer