This repository provides a torch implementation of the UORO algorithm described in https://arxiv.org/abs/1702.05043.

Torch must be installed for the implementation to work. Installation instructions are provided here: http://torch.ch/docs/getting-started.html.

The following command line

th train.lua

trains a GRU model on the anbn dataset presented in the paper using the UORO algorithm. The command line

th train.lua -h

provides a detailled list of the additional command line arguments that you can provide.

The UORO implementation provided in the file uoro.lua provides a general implementation of UORO that you freely use in all cases covered by the framework detailled in the paper.