This is a self contained software accompanying the paper titled: Learning Longer Memory in Recurrent Neural Networks: http://arxiv.org/abs/1412.7753. The code allows you to reproduce our results on two language modeling datasets:
The code implements three recurrent models:
It also allows you to play around with various hyper-parameters.
Here are some of the examples of how to use the code.
you type
th -i main.lua -dset ptb -name srnn_sm -nhid 100 -batchsz 32 -eta 0.05
you type
th -i main.lua -dset text8 -name lstm_sm -nhid 100 -batchsz 32 -eta 0.05 -blen 20 -bfreq 5
you type
th -i main.lua -dset text8 -name scrnn_sm -nhid 100 -nslow 20 -batchsz 32 -eta 0.05 -blen 30 -bfreq 5
To list all the options available, you need to type
th main.lua --help
The software requires you to have the following two packages already installed on your systems:
Download the files in an appropriate directory and run the code from there. See below.
The top level file is called main.lua. In order to run the code you need to run the file using torch. For example:
th -i main.lua -<option1_name> option1_val -<option2_name> option2_val ...
In order to check what all options are available, type
th -i main.lua --help
Structurally Constrained Recurrent Neural Network is BSD-licensed. We also provide an additional patent grant.
See the CONTRIBUTING file for how to help out.