This is the code used for the paper "Inferring algorithmic patterns with a stack augmented recurrent network", by Armand Joulin and Tomas Mikolov.
Stack RNN is a project gathering the code from the paper Inferring Algorithmic Patterns with Stack-Augmented Recurrent Nets by Armand Joulin and Tomas Mikolov (pdf). In this research project, we focus on extending Recurrent Neural Networks (RNN) with a stack to allow them to learn sequences which require some form of persistent memory.
Examples are given in the script
script_tasks.sh. The code is still under construction. We are working on releasing the code for the list RNN. If you have any suggestion, please let us know (contacts below).
To run the code on a task: ```
make toy ./traintoy -ntask 1 -nchar 2 -nhid 10 -nstack 1 -lr .1 -nmax 10 -depth 2 -bptt 50 -mod 1To run the code on binary addition:make add ./trainadd ```
Stack RNN works on: * Mac OS X * Linux
It was not tested on Windows. To compile the code a relatively recent version of g++ is required.
maketo compile everything.
For more help about the options: ```
make toy ./traintoy --help ``Note thattrainaddcan take the same options astrain_toy`.
See the CONTRIBUTING file for how to help out.
Stack RNN is BSD-licensed. We also provide an additional patent grant