The Arti Canon: Neural Text Generation

    "Why should I be angry with the childish pain of conditioned existence?"
                     - A Computer

Arti Canon is a neural text generation project designed to generate believable writing based on classic texts from Buddhism. It processes text from The Dhammapada, Zen Mind, Beginner's Mind, and The Gospel of Buddha, as well as two different translations of The Bodhicaryvatara (see references for links to the originals).

A brief overview of the process:

1. data/txt_to_npy.py parses raw .txt copies of the original texts into a numpy datset consisting of (x,y) pairs of (sequence of characters, next character). This array is saved to the data directory.
2. train.py builds a sequence to sequence language model in Keras, then trains it on the text data, saving the weights to the model_saves directory.
3. articanon.py contains wrappers and utility functions for the text generation process.
4. book.py handles conversion from the raw txt output of the generator to a pdf.
5. write.py is the main script for generating multiple chapters and assembling them into one 'book'/pdf.

(A full technical overview of the project can be found here)

Using the trained model to write your own book

You can generate your own version of the articanon with the following command:

 write.py --chapters *int* --verses *int* --k *int* --filter *True/False*

Where k is the width of the generator's beam search, chapters is the number of chapters to generate and verses is the number of pseudo-intellectual verses you'd like in each chapter. filter lets you cycle through the output and delete unwanted verses.

There are a number of settings to tweak at the top of the Articanon class, the most obvious of which are set up as properties of each Articanon object.

Note that the core capabilities of this library are located in the model training and generation scripts we've included. The write.py script in particular handles all the key interactions with most of the code base; that is the process that we prioritized debugging and getting to work. This is all just a fancy way of saying that, if you start to venture out from the default options-- like switching from beam search to temperature sampling or modifying the pdf output -- you will probably start running into cases we didn't consider pretty quickly. However, most methods have docstrings and the pydoc files have been included here for convenience.

That's cool, but what if I just want to run the entire project myself?

I don't know why you'd want to do this, but it took for me 4 seconds to write and will take your laptop 4 days to run-- that's a time efficiency opportunity that was too good to pass up:

chmod u+x articanon.sh
./articanon.sh

References

1. The Bodhicaryvatara v1
2. The Bodhicaryvatara v2
3. The Dhammapada
4. Zen Mind Beginner's Mind
5. The Gospel of Buddha
6. Deep Learning with Python
7. Neural Text Generation, a Practical Guide