This code implements a configurable LSTM to generate text similiar to a set you train it on.
You need any version of python with numpy and keras installed. Note that for keras to work you need to either install tensorflow or theano or CTNK.
Configure the config.py file to suit your needs and provide a dataset under the name input.txt (changable in the configs).
If there is a model file, the neural network will continue working on that, otherwise it will start from scratch. In order to change the structure of the machine you need to clear out the old model file.
Run textgen.py to train the model and run run.py in order to print out samples based on the trained model.
As of this version, stateful machines must run with batch size 1. This makes each iteration very slow. Sorry for that!
- Save the iteration for the model, so that when resumed, it would continue counting and not reset to 1.
- Save the model name based on the input file
- Write a better Readme
- Write better normalizer
- Provide some sample datasets
- Write more TODOs