KerasTextGeneration

This repository contains a tool to generate text using RNN (LSTM) using keras with Tensorflow backend. To run an instance using the example data given, you can run:

python main.py data/text.txt

The repository is structured as follows:

• main.py This code is the backbone of the project, calls the rest of the modules to train or generate.

• modelCreator.py This module creates the RNN and also handles the text generation

• parseText.py This module is responsible of extracting the vocabulary and the dictionaries of the input text. Also is in charge of generating the dataset for training.

Parameter options

usage: python main.py [-h] [--seq_length SEQ_LENGTH] [--epochs EPOCHS]
               [--batch_size BATCH_SIZE] [--verbose VERBOSE]
               [--gen_length GEN_LENGTH]
               path

Train a RNN to generate text.

positional arguments:
  path                  Path to text file with training text.

optional arguments:
  -h, --help            show this help message and exit
  --seq_length SEQ_LENGTH
                        Sequence length to train.
  --epochs EPOCHS       Number of epochs to train.
  --batch_size BATCH_SIZE
                        Batch size used in training.
  --verbose VERBOSE     Verbosity in training.
  --gen_length GEN_LENGTH
                        Length of generated string.

Data

• text.txt Example text from the "Fate Stay Night" series.
• hs.txt Example text from the "Fate Stay Night - Heaven's Feel" series.
• ubw.txt Example text from the "Fate Stay Night - Unlimited Blade Works" series.
• full.txt Example text from all the "Fate Stay Night" series.
• downloadData.py Code used to download text.txt, can be modified to download text from any static html page.

Dependencies

• keras
• Tensorflow backend
• Numpy

All the dependencies can be fullfilled by running: pip install --upgrade keras tensorflow numpy