The relevant blog post is here: http://hjweide.github.io/char-rnn

This implementation is largely based on https://github.com/Lasagne/Recipes/blob/master/examples/lstm_text_generation.py. See http://karpathy.github.io/2015/05/21/rnn-effectiveness/ for a thorough explanation of how char-rnn works.

This implementation of char-rnn can be used to train on any text file. My goal, however, was to train it on the entire history of my Facebook conversations. If you have your own text file, you can skip to step 5 below.

1. Follow these instructions to get a copy of all your Facebook data. You may want to do this first, because it can take a while for them to send you the download link. When the download is complete, unzip the archive.

2. Clone and install this [Facebook chat parser](Facebook chat parser).

   git clone https://github.com/ownaginatious/fbchat-archive-parser 
   python setup.py develop

3. Run the parser on the messages.htm file from the extracted archive:

   fbcap html/messages.htm > messages.txt
   

4. Use this snippet of code to strip out all messages not written by you. Set the name appearing in your Facebook chats as the name variable, and run python parse_messages.py. You may need to write a more sophisticated parser if you want more control about which messages you want to extract, or if you had a name change, for example.

5. Set the text_fpath in train_char_rnn.py to the text file containing the training data. If you used the snippet mentioned above, this will already be appropriately set as parsed.txt.

6. Observe the sequences generated during training. Once you are happy that the model has reached reasonable convergence, end the training with ctrl-c.

7. Set the text_fpath in generate_samples.py, and run python generate_samples.py to continually supply phrases and sample from the model to amuse yourself.