Torch7 LSTM Experiment
My experiments hacking around with lstm for torch7.
My code was inspired by a great article published here: http://apaszke.github.io/lstm-explained.html
NOTES:
I moved on to the rnn library as I got very muddled trying to figure out how to train my own lstm.
This code still not managing to learn properly I don't think,
Current experiment design:
a. stream in csv timeseries data, don't read it all in at once. (future files maybe bigger than ram)
b. feed this data into the FastLSTM lib. Every batchsize steps, build a table of tensors and feed it to learning...
d. do some training - to make it simple, train on the first x rows of the timeseries, then guess the remainder
e. print the predictions, alongside the actual to get a minimal sense of whether it worked
to run:
th tc_rnn_lstm.lua
depends on an installation of torch7 and the csv library found here: https://github.com/geoffleyland/lua-csv