- char_rnn_theano_orig.py is code for replicating the model they had with a few differences we can discuss.
- char_rnn_theano.py is the same, but with random parameter seeding to search over for the best network and training configuration.
- gen_patternltd_data.py is a quick and dirty generator of a toy dataset that I think stands in well for the problem of predicting target outcomes that require a good memory. In this case, the dataset is a string of digits. The network has to learn the rule that, if it sees ‘456’ within 20 steps of seeing ‘123’, then it has to switch it’s output from 0 to 1. I also added predictors of ‘123’ and ‘456’ (‘7’, and ‘8’, respectively), that have a relatively strong probability (0.5) of being followed immediately by the patterns. I haven’t started running this yet, however, so I don’t have the runner script ready for you.
- input.txt is the shakespear data
- jbpickle.py is a pickling helper
- submit* are wrappers for submitting to princeton gpu clusters
- merge char_rnn_theano_* and add flags for differences instead
- add model comparison script
- add clipping of gradients to keras opimizer
- modularize data prep out
- make runner for patternltd dataset
- create early stopping crit for bad training