Quasi-Recurrent Neural Networks were introduced at ICLR 2017 by Bradbury et al., see https://arxiv.org/abs/1611.01576. We provide basic Quasi-RNN layers that can be used together with the Lasagne framework. The implementation is not necessarily optimized for fast parallel execution, as explained in the paper, and is kept as basic as possible. A language modeling example is provided in the accompanying notebook.
Three pooling modes can be used, as explained in the original paper. They can be activated using the pooling argument. The modes are 'f', 'fo' and 'ifo'.