The new model, previously introduced byNal Kalchbrenner, has been implemented in PyTorch, but due to time constraints for the original publication I implemented thek-max pooling andfolding operations in Python. To improve the speed of training it would be better if these were integrated with PyTorch at a lower-level.