Tutorial: Multi-layer Recurrent Neural Networks (LSTM) for text models in Python using Keras.

Before going through this tutorial, I suggest to read the very very good blog note from Andrej Karpathy: http://karpathy.github.io/2015/05/21/rnn-effectiveness/

• Tensorflow 1.1.0rc0
• Python 3.6
• Keras 2.0.5

The project has train.py script.

The script generate and save locally a model and text features.

Just run it using python train.py some items can be modified by editing the script:

• data_dir = 'data/Artistes_et_Phalanges-David_Campion'# data directory containing input.txt
• save_dir = 'save' # directory to store models
• rnn_size = 128 # size of RNN
• batch_size = 30 # minibatch size
• seq_length = 15 # sequence length
• num_epochs = 8 # number of epochs
• learning_rate = 0.001 #learning rate
• sequences_step = 1 #step to create sequences

note: 8 epochs is cleraly not enough. Should be above 25.

The project has a generate,py script.

Just run it using python generate.py some items can be modified by editing the script:

• save_dir = 'save' # directory where model is stored
• batch_size = 30 # minibatch size
• seq_length = 25 # sequence length
• words_number = 400 #number of words to generate
• seed_sentences = "Il y a" #sentence for seed generation

The project comes with two types of input:

• data/tinyshakespeare/input.txt:
  • a small condensate of Shakespeare books
• data/Artistes_et_Phalanges-David_campion/input.txt:
  • The complete text of a french fantasy book "Artistes et Phalanges", by David Campion
  • This file book is under the following licence: Licence Creative Commons CC BY-NC-ND