Closely follows the reference Torch implementation, accompanying the work An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling by Shaojie Bai, J. Zico Kolter and Vladlen Koltun.
Currently only the adding problem from the original experiments is implemented
Furthermore there is a 'Sanity Check' experiment, which tries to predict the next value in a random sequence in [0,1) This experiment confirms that the network does not leak future information and converges to always predicting 0.5.
To install, run
pip install .
To run an experiment, run
python main.py
in the experiment's subdirectory.
(There are probably better ways to accomplish this)
Currently only tested with Keras 2.0.5; >2.1.0 is known to be incompatible due to an API change not yet incorporated into the weightnorm code by OpenAI we use.