word-level-rnn

Word Level RNN for text generation using keras.

This is a project with a goal to make a Word Level Text generation using keras on Theano Framework. The code is based on the sample text generation character level code provided as a sample example at https://github.com/fchollet/keras/blob/master/examples/lstm_text_generation.py

Requirements:

Theano: https://github.com/Theano/Theano Keras: https://github.com/fchollet/keras

Usage:

To train a dataset, put the txt file inside the dataset directory, and run the script, additionally you can tweak the batchsize as per your machine's memory contraint

Future Update:

I'll add an option to have a choice between different RNN models

Sample Testing:

I tried training it on chapters of The Game of Thrones (individually, because of memory constraints), seems coherent, please try and give me feedback.

Read This:

Since the code is done at word level instead of character level, it needs more epochs than the character level text generation counter part, as well as much more memory because it has to be trained on thousands of words, (unlike character level code where it's only trained on 50-100 different characters (assuming ASCII).

Word level RNN text generation should in theory be more coherent then the Character level text generation at lesser amount of training. Suggestions and criticism are welcome to optimising and improve the code. thanks!